Repository navigation
Releases
fastpilot v0.5.1
What's new in 0.5.1
Small fixes after the 0.5.0 checks.
- Plugin or
fastpilot init, not both. The README says so: together they start two fastpilot
servers. With the plugin, a secret such asDEMO_PASSworks like this: add its name infastpilot ui›
Settings › Secret, and set its value withexport DEMO_PASS=…before you start Claude Code. When a
secret is missing, the message now names the plugin case too. - Jira skills explain the connection themselves. Each skill gives the two steps:
claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp, then/mcp› atlassian ›
sign in, picking the site that has your project. If Atlassian says a site "isn't explicitly granted",
the skill asks you to sign in again and pick that site, and never switches to another site. - Run now works again after its run was stopped from outside. A scheduled run that was killed no
longer stays "already running" untilfastpilot uirestarts. - Settings › Limits & permissions no longer shows the old "New in 0.3.2" notice.
- The skill count is the same everywhere: the Skills page says "23 on, plus the fastpilot skill (always
on)", 24 in all, as in the README.
Known limits
- The Jira skills have not been tested against a real Jira site yet.
- The known limits of 0.5.0 below still apply.
What's new in 0.5.0
fastpilot where you already work: a Claude Code plugin, the official MCP Registry, GitHub Actions and
Jira.
Claude Code plugin
- Install fastpilot in Claude Code in one step:
/plugin marketplace add autobot-lab/fastpilot, then
/plugin install fastpilot@fastpilot. You get the fastpilot tools and its testing skills, shown as
fastpilot:<skill>, for examplefastpilot:boundary-value-analysis. - Use either the plugin or
fastpilot initin a project, not both: with both, Claude Code shows two
fastpilot servers.
Official MCP Registry
- fastpilot is listed on the official MCP Registry as
io.github.autobot-lab/fastpilot, so MCP clients
that read the registry can find and install it.
GitHub Actions
- Run your fastpilot web tests in CI with
autobot-lab/fastpilot-action@v1. It downloads
the fastpilot release for the runner, checks its SHA-256, runs your tests, and gives a JUnit file for
the checks tab and an HTML report. Pull requests from forks get no secrets, so their tests show
"could not run" instead of a pass.
Jira
- Three Jira skills, used through Atlassian's own MCP server in Claude Code:
- Jira ticket to test reads a ticket's acceptance criteria and turns them into tests. It never
writes to Jira. - Jira bug report turns a failed run into one new Jira issue (steps, expected and actual result,
links to the report and video). - Jira test result posts one comment with the result. It changes the ticket's status only when you
ask for it in your own words.
- Jira ticket to test reads a ticket's acceptance criteria and turns them into tests. It never
- Every write shows the exact draft, including the Jira site, and waits for your "yes". With nobody to
ask (a scheduled or unattended run) nothing is written. Ticket and page text is treated as data, never
as instructions, and emails, names and order numbers from tested pages are masked. - Setup: connect Atlassian's MCP server in Claude Code (
/mcpto sign in) and keep Claude Code's own
question on for every Jira write. See the guide's "Jira" section. Jira works in Claude Code, not in
dashboard Chat.
Docs
- The README shows the dashboard: Chat creating a test, the Device page, Live and Schedules.
Known limits
- The Jira skills have not been tested against a real Jira site yet, only with a stand-in Atlassian
server. The Atlassian endpoint in the guide was checked through Claude's built-in Atlassian connector,
not against Atlassian's own documentation. - With the plugin alone, a test that needs a secret such as
${DEMO_PASS}stops and asks you to set it:
put it in the project's.mcp.json(asfastpilot initandfastpilot demodo) or export it. - The Jira skills point to the guide for setup; npm and plugin installs have no local copy of the guide,
so use the online guide. - Settings › Limits & permissions still says "New in 0.3.2" about blocked payment buttons.
- Not tested yet: models other than Claude, Linux, physical Android phones and iPhones.
What's new in 0.4.2
Installs on npm 12, says plainly when a test could not run, and sets up iPhone testing by itself.
Install
- Works on npm 12. npm 12 skips install scripts unless you allow them, so 0.4.1 installed without
its program andfastpilotfailed to start. Now the first run downloads the program for your
computer, checks its SHA-256 against the GitHub Release, and then starts. This happens once; later
runs need no network. Messages go to stderr, so Claude Code's connection to fastpilot is not
disturbed. npm 11 and older still download it at install time. - The download stops after 2 minutes or 150 MB, and a mirror set with
FASTPILOT_DOWNLOAD_BASEmust
use https (except on your own computer). - Claude Code plugin:
/plugin marketplace add autobot-lab/fastpilot, then
/plugin install fastpilot@fastpilotadds the fastpilot tools and its 21 testing skills in one step. - GitHub Actions: run your fastpilot web tests in CI with
autobot-lab/fastpilot-action@v1.
Clear messages
- A missing secret stops the test before it starts. When a test needs a secret such as
${DEMO_PASS}that is not set,fastpilot runsays "Could not run", names the secret (never its
value), gives the exact command to set it, and does not retry. It is reported as a setup problem,
not as a bug in your app, and never as a pass. Schedules and Release watch say "could not run:
needs DEMO_PASS" in the summary and the notification, and the Schedules page shows it in amber with
a link to Settings › Secret. - The Ship check card keeps each test on one row.
- Live says where a session came from: "Started from the terminal" or "Scheduled · ".
- Blocked actions are explained in plain words ("the built-in blocked list"), without setting names.
- A web-only run prints "Web · Chrome" instead of phone and iPhone counts.
- The demo pages work at phone width, so Ship check no longer flags every demo page.
iPhone
- The first iPhone test sets up iPhone testing by itself. When the test runner is missing,
fastpilot run(or Run now) builds it once with your Xcode, shows the time so far, and then
starts the test. Without Xcode it says to install Xcode from the App Store. - A scheduled run does not build it: it reports that a one-time setup is needed, so run one iPhone
test by hand first.
For maintainers
- The release workflow can publish to npm with Trusted Publishing (no stored token, with
provenance) once the owner sets it up on npmjs.com and GitHub. See
docs/releasing.md.
Known limits
- iPhone sessions started from Claude Code or dashboard Chat do not build the runner yet: use Set up
iPhone testing on the Device page, orfastpilot setup ios, once. fastpilot runreads secrets only from the environment, not from.mcp.json. The message says how
to set them.- Not tested yet: models other than Claude, Linux, physical Android phones and iPhones.
What's new in 0.4.1
Use the phones you already have, and a quieter Chat.
Devices
- Your installed emulators come first. When no Android device is running, the Device page lists the
emulators already on this computer by name, each with Start. An emulator starts only when you
click Start and confirm; Chat and agents cannot start one. It uses a free port, and other running
emulators are not touched. Add a new test phone (a download of about 1.5 GB) is now the second
choice, and the main one only when no emulator is installed. - Pick the iPhone simulator tests use. The Device page lists your installed iPhone simulators.
Use for tests saves one; Any running simulator goes back to the default. - iPhone testing is a one-time setup, no longer shown as an error.
fastpilot doctor,fastpilot init
and the Device page say "iPhone testing needs a one-time setup (about 5 minutes)". The Device page has
Set up iPhone testing, andfastpilot initoffers it when Xcode is found. fastpilot run --workersuses only the device you chose. WithFASTPILOT_ANDROID_SERIAL(or
ANDROID_SERIAL) set, no other Android device is contacted. The first line of the run says which
devices are used.
Dashboard
- The empty Chat no longer shows the three suggestion buttons. It keeps the heading and one help line.
Docs
- The README is now a step-by-step guide to the dashboard: start in 3 steps, then each page in turn.
Fixes
- The release check reaches an Android device only in its Android step, and two tests that sometimes
failed on a busy machine are stable.
Known limits
- The iPhone test runner is not built automatically on the first iPhone test yet: run Set up iPhone
testing (orfastpilot setup ios) once. - The known limits of 0.4.0 below still apply.
What's new in 0.4.0
Talk to it, in plain words. Several QA agents, each with its own model and skills, that test like a
certified tester and show what they covered.
Agents
- Several agents, like a team. Each agent has its own Chat, Live view, Schedules, Instructions
(with suggestions you can click), Skills, History and settings. Create one with a short form, or
describe it in one sentence in Chat and click Create. Existing projects get one default agent
that keeps your chat, schedules and history. - Provider, then model, per agent. Claude Code, OpenAI, OpenRouter, Groq, Ollama (local) or another
OpenAI-compatible service. API keys are ...
fastpilot v0.5.0
What's new in 0.5.0
fastpilot where you already work: a Claude Code plugin, the official MCP Registry, GitHub Actions and
Jira.
Claude Code plugin
- Install fastpilot in Claude Code in one step:
/plugin marketplace add autobot-lab/fastpilot, then
/plugin install fastpilot@fastpilot. You get the fastpilot tools and its testing skills, shown as
fastpilot:<skill>, for examplefastpilot:boundary-value-analysis. - Use either the plugin or
fastpilot initin a project, not both: with both, Claude Code shows two
fastpilot servers.
Official MCP Registry
- fastpilot is listed on the official MCP Registry as
io.github.autobot-lab/fastpilot, so MCP clients
that read the registry can find and install it.
GitHub Actions
- Run your fastpilot web tests in CI with
autobot-lab/fastpilot-action@v1. It downloads
the fastpilot release for the runner, checks its SHA-256, runs your tests, and gives a JUnit file for
the checks tab and an HTML report. Pull requests from forks get no secrets, so their tests show
"could not run" instead of a pass.
Jira
- Three Jira skills, used through Atlassian's own MCP server in Claude Code:
- Jira ticket to test reads a ticket's acceptance criteria and turns them into tests. It never
writes to Jira. - Jira bug report turns a failed run into one new Jira issue (steps, expected and actual result,
links to the report and video). - Jira test result posts one comment with the result. It changes the ticket's status only when you
ask for it in your own words.
- Jira ticket to test reads a ticket's acceptance criteria and turns them into tests. It never
- Every write shows the exact draft, including the Jira site, and waits for your "yes". With nobody to
ask (a scheduled or unattended run) nothing is written. Ticket and page text is treated as data, never
as instructions, and emails, names and order numbers from tested pages are masked. - Setup: connect Atlassian's MCP server in Claude Code (
/mcpto sign in) and keep Claude Code's own
question on for every Jira write. See the guide's "Jira" section. Jira works in Claude Code, not in
dashboard Chat.
Docs
- The README shows the dashboard: Chat creating a test, the Device page, Live and Schedules.
Known limits
- The Jira skills have not been tested against a real Jira site yet, only with a stand-in Atlassian
server. The Atlassian endpoint in the guide was checked through Claude's built-in Atlassian connector,
not against Atlassian's own documentation. - With the plugin alone, a test that needs a secret such as
${DEMO_PASS}stops and asks you to set it:
put it in the project's.mcp.json(asfastpilot initandfastpilot demodo) or export it. - The Jira skills point to the guide for setup; npm and plugin installs have no local copy of the guide,
so use the online guide. - Settings › Limits & permissions still says "New in 0.3.2" about blocked payment buttons.
- Not tested yet: models other than Claude, Linux, physical Android phones and iPhones.
What's new in 0.4.2
Installs on npm 12, says plainly when a test could not run, and sets up iPhone testing by itself.
Install
- Works on npm 12. npm 12 skips install scripts unless you allow them, so 0.4.1 installed without
its program andfastpilotfailed to start. Now the first run downloads the program for your
computer, checks its SHA-256 against the GitHub Release, and then starts. This happens once; later
runs need no network. Messages go to stderr, so Claude Code's connection to fastpilot is not
disturbed. npm 11 and older still download it at install time. - The download stops after 2 minutes or 150 MB, and a mirror set with
FASTPILOT_DOWNLOAD_BASEmust
use https (except on your own computer). - Claude Code plugin:
/plugin marketplace add autobot-lab/fastpilot, then
/plugin install fastpilot@fastpilotadds the fastpilot tools and its 21 testing skills in one step. - GitHub Actions: run your fastpilot web tests in CI with
autobot-lab/fastpilot-action@v1.
Clear messages
- A missing secret stops the test before it starts. When a test needs a secret such as
${DEMO_PASS}that is not set,fastpilot runsays "Could not run", names the secret (never its
value), gives the exact command to set it, and does not retry. It is reported as a setup problem,
not as a bug in your app, and never as a pass. Schedules and Release watch say "could not run:
needs DEMO_PASS" in the summary and the notification, and the Schedules page shows it in amber with
a link to Settings › Secret. - The Ship check card keeps each test on one row.
- Live says where a session came from: "Started from the terminal" or "Scheduled · ".
- Blocked actions are explained in plain words ("the built-in blocked list"), without setting names.
- A web-only run prints "Web · Chrome" instead of phone and iPhone counts.
- The demo pages work at phone width, so Ship check no longer flags every demo page.
iPhone
- The first iPhone test sets up iPhone testing by itself. When the test runner is missing,
fastpilot run(or Run now) builds it once with your Xcode, shows the time so far, and then
starts the test. Without Xcode it says to install Xcode from the App Store. - A scheduled run does not build it: it reports that a one-time setup is needed, so run one iPhone
test by hand first.
For maintainers
- The release workflow can publish to npm with Trusted Publishing (no stored token, with
provenance) once the owner sets it up on npmjs.com and GitHub. See
docs/releasing.md.
Known limits
- iPhone sessions started from Claude Code or dashboard Chat do not build the runner yet: use Set up
iPhone testing on the Device page, orfastpilot setup ios, once. fastpilot runreads secrets only from the environment, not from.mcp.json. The message says how
to set them.- Not tested yet: models other than Claude, Linux, physical Android phones and iPhones.
What's new in 0.4.1
Use the phones you already have, and a quieter Chat.
Devices
- Your installed emulators come first. When no Android device is running, the Device page lists the
emulators already on this computer by name, each with Start. An emulator starts only when you
click Start and confirm; Chat and agents cannot start one. It uses a free port, and other running
emulators are not touched. Add a new test phone (a download of about 1.5 GB) is now the second
choice, and the main one only when no emulator is installed. - Pick the iPhone simulator tests use. The Device page lists your installed iPhone simulators.
Use for tests saves one; Any running simulator goes back to the default. - iPhone testing is a one-time setup, no longer shown as an error.
fastpilot doctor,fastpilot init
and the Device page say "iPhone testing needs a one-time setup (about 5 minutes)". The Device page has
Set up iPhone testing, andfastpilot initoffers it when Xcode is found. fastpilot run --workersuses only the device you chose. WithFASTPILOT_ANDROID_SERIAL(or
ANDROID_SERIAL) set, no other Android device is contacted. The first line of the run says which
devices are used.
Dashboard
- The empty Chat no longer shows the three suggestion buttons. It keeps the heading and one help line.
Docs
- The README is now a step-by-step guide to the dashboard: start in 3 steps, then each page in turn.
Fixes
- The release check reaches an Android device only in its Android step, and two tests that sometimes
failed on a busy machine are stable.
Known limits
- The iPhone test runner is not built automatically on the first iPhone test yet: run Set up iPhone
testing (orfastpilot setup ios) once. - The known limits of 0.4.0 below still apply.
What's new in 0.4.0
Talk to it, in plain words. Several QA agents, each with its own model and skills, that test like a
certified tester and show what they covered.
Agents
- Several agents, like a team. Each agent has its own Chat, Live view, Schedules, Instructions
(with suggestions you can click), Skills, History and settings. Create one with a short form, or
describe it in one sentence in Chat and click Create. Existing projects get one default agent
that keeps your chat, schedules and history. - Provider, then model, per agent. Claude Code, OpenAI, OpenRouter, Groq, Ollama (local) or another
OpenAI-compatible service. API keys are stored once per provider in Settings. Dashboard Chat now
works with any provider. Models other than Claude are labelled "not tested yet". - Agents run in parallel. Web tests run side by side. A phone or emulator runs one test at a time;
other agents wait in a queue for it. "Tests at once" limits how many run together (default 3).
Tests that think like a certified tester
- 21 built-in skills, grouped by ISTQB® test design technique (fastpilot is not affiliated with
ISTQB): Choose technique, Equivalence partitioning, Boundary value analysis, Decision table, State
transition, Pairwise, Use case scenarios, Error guessing, Exploratory testing, Checklist-based (with
checklists for login and OTP, forms, safe checkout, accessibility and mobile), Acceptance tests,
Regression, Bug report, Test report and Release check.fastpilot initandfastpilot demoinstall
them. - "Techniques used": results say which technique was used, which values were tried, what was
assumed, and what was covered (for example "Boundary value analysis · Age — 6 of 6 values"). - Regression puts what is newly failing since the last run first, then still failing, flaky,
blocked and passed. - Practice pages in
fastpilot demo: a ticket field with limits, shipping rules and an order
status flow, each with o...
fastpilot v0.4.2
What's new in 0.4.2
Installs on npm 12, says plainly when a test could not run, and sets up iPhone testing by itself.
Install
- Works on npm 12. npm 12 skips install scripts unless you allow them, so 0.4.1 installed without
its program andfastpilotfailed to start. Now the first run downloads the program for your
computer, checks its SHA-256 against the GitHub Release, and then starts. This happens once; later
runs need no network. Messages go to stderr, so Claude Code's connection to fastpilot is not
disturbed. npm 11 and older still download it at install time. - The download stops after 2 minutes or 150 MB, and a mirror set with
FASTPILOT_DOWNLOAD_BASEmust
use https (except on your own computer). - Claude Code plugin:
/plugin marketplace add autobot-lab/fastpilot, then
/plugin install fastpilot@fastpilotadds the fastpilot tools and its 21 testing skills in one step. - GitHub Actions: run your fastpilot web tests in CI with
autobot-lab/fastpilot-action@v1.
Clear messages
- A missing secret stops the test before it starts. When a test needs a secret such as
${DEMO_PASS}that is not set,fastpilot runsays "Could not run", names the secret (never its
value), gives the exact command to set it, and does not retry. It is reported as a setup problem,
not as a bug in your app, and never as a pass. Schedules and Release watch say "could not run:
needs DEMO_PASS" in the summary and the notification, and the Schedules page shows it in amber with
a link to Settings › Secret. - The Ship check card keeps each test on one row.
- Live says where a session came from: "Started from the terminal" or "Scheduled · ".
- Blocked actions are explained in plain words ("the built-in blocked list"), without setting names.
- A web-only run prints "Web · Chrome" instead of phone and iPhone counts.
- The demo pages work at phone width, so Ship check no longer flags every demo page.
iPhone
- The first iPhone test sets up iPhone testing by itself. When the test runner is missing,
fastpilot run(or Run now) builds it once with your Xcode, shows the time so far, and then
starts the test. Without Xcode it says to install Xcode from the App Store. - A scheduled run does not build it: it reports that a one-time setup is needed, so run one iPhone
test by hand first.
For maintainers
- The release workflow can publish to npm with Trusted Publishing (no stored token, with
provenance) once the owner sets it up on npmjs.com and GitHub. See
docs/releasing.md.
Known limits
- iPhone sessions started from Claude Code or dashboard Chat do not build the runner yet: use Set up
iPhone testing on the Device page, orfastpilot setup ios, once. fastpilot runreads secrets only from the environment, not from.mcp.json. The message says how
to set them.- Not tested yet: models other than Claude, Linux, physical Android phones and iPhones.
What's new in 0.4.1
Use the phones you already have, and a quieter Chat.
Devices
- Your installed emulators come first. When no Android device is running, the Device page lists the
emulators already on this computer by name, each with Start. An emulator starts only when you
click Start and confirm; Chat and agents cannot start one. It uses a free port, and other running
emulators are not touched. Add a new test phone (a download of about 1.5 GB) is now the second
choice, and the main one only when no emulator is installed. - Pick the iPhone simulator tests use. The Device page lists your installed iPhone simulators.
Use for tests saves one; Any running simulator goes back to the default. - iPhone testing is a one-time setup, no longer shown as an error.
fastpilot doctor,fastpilot init
and the Device page say "iPhone testing needs a one-time setup (about 5 minutes)". The Device page has
Set up iPhone testing, andfastpilot initoffers it when Xcode is found. fastpilot run --workersuses only the device you chose. WithFASTPILOT_ANDROID_SERIAL(or
ANDROID_SERIAL) set, no other Android device is contacted. The first line of the run says which
devices are used.
Dashboard
- The empty Chat no longer shows the three suggestion buttons. It keeps the heading and one help line.
Docs
- The README is now a step-by-step guide to the dashboard: start in 3 steps, then each page in turn.
Fixes
- The release check reaches an Android device only in its Android step, and two tests that sometimes
failed on a busy machine are stable.
Known limits
- The iPhone test runner is not built automatically on the first iPhone test yet: run Set up iPhone
testing (orfastpilot setup ios) once. - The known limits of 0.4.0 below still apply.
What's new in 0.4.0
Talk to it, in plain words. Several QA agents, each with its own model and skills, that test like a
certified tester and show what they covered.
Agents
- Several agents, like a team. Each agent has its own Chat, Live view, Schedules, Instructions
(with suggestions you can click), Skills, History and settings. Create one with a short form, or
describe it in one sentence in Chat and click Create. Existing projects get one default agent
that keeps your chat, schedules and history. - Provider, then model, per agent. Claude Code, OpenAI, OpenRouter, Groq, Ollama (local) or another
OpenAI-compatible service. API keys are stored once per provider in Settings. Dashboard Chat now
works with any provider. Models other than Claude are labelled "not tested yet". - Agents run in parallel. Web tests run side by side. A phone or emulator runs one test at a time;
other agents wait in a queue for it. "Tests at once" limits how many run together (default 3).
Tests that think like a certified tester
- 21 built-in skills, grouped by ISTQB® test design technique (fastpilot is not affiliated with
ISTQB): Choose technique, Equivalence partitioning, Boundary value analysis, Decision table, State
transition, Pairwise, Use case scenarios, Error guessing, Exploratory testing, Checklist-based (with
checklists for login and OTP, forms, safe checkout, accessibility and mobile), Acceptance tests,
Regression, Bug report, Test report and Release check.fastpilot initandfastpilot demoinstall
them. - "Techniques used": results say which technique was used, which values were tried, what was
assumed, and what was covered (for example "Boundary value analysis · Age — 6 of 6 values"). - Regression puts what is newly failing since the last run first, then still failing, flaky,
blocked and passed. - Practice pages in
fastpilot demo: a ticket field with limits, shipping rules and an order
status flow, each with one planted bug.
Everything through chat
- Chat creates a test from a sentence. One card shows the steps; Save & schedule or Save test
only. Nothing is written before you click, and variables the test needs are checked first. - All settings through chat, in the dashboard and in Claude Code:
- making limits stricter applies at once, with Undo;
- other changes need your click;
- loosening a limit only appears when you asked for it in your own words, as a red card that
quotes you; allowing Pay also needs you to type "Pay" and expires after 10 minutes; - text on a tested page can never ask for a change;
- in the terminal, the AI can only file a request; you approve it yourself with
fastpilot config approve; - every change is written to an activity log.
- "Remember this as a note": a card in Chat saves a project note from your own words.
Schedules and releases
- Every N hours (1–24) with an active window (for example 08:00–20:00), Run now, and Next / Last
columns. - The Schedules table says why a run failed (which test, and the reason), not only "0 of 1 passed".
- Release watch: after a release, re-runs your tests every 5, 10, 15, 20, 30 or 60 minutes for 1 to
24 hours. The first run is the "before". After that you hear once about each problem that is new
since then (a test that starts failing, or the tests no longer running), on the schedule's
notifications: Mac notification, Slack webhook or GitHub issue. When the time is up it turns itself
off and removes its system-scheduler entry; the Schedules list keeps it with how it went. Start one
from Schedules › New schedule › Release watch, from Chat, or with
fastpilot schedule watch tests/ --every 10 --for 2. - Ship check:
fastpilot ship-check <url>(or ashipCheckstep) looks for console errors,
failed requests, broken links, layout at phone, tablet and desktop width, and basic accessibility,
then gives one verdict: SHIP, SHIP WITH FIXES, DO NOT SHIP or INCONCLUSIVE, always with what was not
checked. It exits with 0 (SHIP or SHIP WITH FIXES), 1 (DO NOT SHIP) or 2 (INCONCLUSIVE). Links are
checked without your cookies unless you ask (--cookies); sign-out, delete, one-time and other
links that change something are never opened. - In Chat, a release check shows as a verdict card, labelled "the bot's summary" (the bot wrote
it from the checks it ran; fastpilot does not verify it). A bug report shows as a preview card
that is not sent anywhere, with Copy report and Copy as Markdown.
Dashboard
- History: runs and saved Chat conversations; delete one, several or all; optional auto-delete
after 7, 30 or 90 days. - Memory: one line per test (Reliable, Unstable, Failing, Set aside, New), with Set aside and
Forget, plus project notes (staging address, test account names, "login uses a code") that
agents read. Notes that look like a password or hide invisible text are refused. Notes are kept in
your own fastpilot folder, not in the project. An agent on an...
fastpilot v0.4.1
What's new in 0.4.1
Use the phones you already have, and a quieter Chat.
Devices
- Your installed emulators come first. When no Android device is running, the Device page lists the
emulators already on this computer by name, each with Start. An emulator starts only when you
click Start and confirm; Chat and agents cannot start one. It uses a free port, and other running
emulators are not touched. Add a new test phone (a download of about 1.5 GB) is now the second
choice, and the main one only when no emulator is installed. - Pick the iPhone simulator tests use. The Device page lists your installed iPhone simulators.
Use for tests saves one; Any running simulator goes back to the default. - iPhone testing is a one-time setup, no longer shown as an error.
fastpilot doctor,fastpilot init
and the Device page say "iPhone testing needs a one-time setup (about 5 minutes)". The Device page has
Set up iPhone testing, andfastpilot initoffers it when Xcode is found. fastpilot run --workersuses only the device you chose. WithFASTPILOT_ANDROID_SERIAL(or
ANDROID_SERIAL) set, no other Android device is contacted. The first line of the run says which
devices are used.
Dashboard
- The empty Chat no longer shows the three suggestion buttons. It keeps the heading and one help line.
Docs
- The README is now a step-by-step guide to the dashboard: start in 3 steps, then each page in turn.
Fixes
- The release check reaches an Android device only in its Android step, and two tests that sometimes
failed on a busy machine are stable.
Known limits
- The iPhone test runner is not built automatically on the first iPhone test yet: run Set up iPhone
testing (orfastpilot setup ios) once. - The known limits of 0.4.0 below still apply.
What's new in 0.4.0
Talk to it, in plain words. Several QA agents, each with its own model and skills, that test like a
certified tester and show what they covered.
Agents
- Several agents, like a team. Each agent has its own Chat, Live view, Schedules, Instructions
(with suggestions you can click), Skills, History and settings. Create one with a short form, or
describe it in one sentence in Chat and click Create. Existing projects get one default agent
that keeps your chat, schedules and history. - Provider, then model, per agent. Claude Code, OpenAI, OpenRouter, Groq, Ollama (local) or another
OpenAI-compatible service. API keys are stored once per provider in Settings. Dashboard Chat now
works with any provider. Models other than Claude are labelled "not tested yet". - Agents run in parallel. Web tests run side by side. A phone or emulator runs one test at a time;
other agents wait in a queue for it. "Tests at once" limits how many run together (default 3).
Tests that think like a certified tester
- 21 built-in skills, grouped by ISTQB® test design technique (fastpilot is not affiliated with
ISTQB): Choose technique, Equivalence partitioning, Boundary value analysis, Decision table, State
transition, Pairwise, Use case scenarios, Error guessing, Exploratory testing, Checklist-based (with
checklists for login and OTP, forms, safe checkout, accessibility and mobile), Acceptance tests,
Regression, Bug report, Test report and Release check.fastpilot initandfastpilot demoinstall
them. - "Techniques used": results say which technique was used, which values were tried, what was
assumed, and what was covered (for example "Boundary value analysis · Age — 6 of 6 values"). - Regression puts what is newly failing since the last run first, then still failing, flaky,
blocked and passed. - Practice pages in
fastpilot demo: a ticket field with limits, shipping rules and an order
status flow, each with one planted bug.
Everything through chat
- Chat creates a test from a sentence. One card shows the steps; Save & schedule or Save test
only. Nothing is written before you click, and variables the test needs are checked first. - All settings through chat, in the dashboard and in Claude Code:
- making limits stricter applies at once, with Undo;
- other changes need your click;
- loosening a limit only appears when you asked for it in your own words, as a red card that
quotes you; allowing Pay also needs you to type "Pay" and expires after 10 minutes; - text on a tested page can never ask for a change;
- in the terminal, the AI can only file a request; you approve it yourself with
fastpilot config approve; - every change is written to an activity log.
- "Remember this as a note": a card in Chat saves a project note from your own words.
Schedules and releases
- Every N hours (1–24) with an active window (for example 08:00–20:00), Run now, and Next / Last
columns. - The Schedules table says why a run failed (which test, and the reason), not only "0 of 1 passed".
- Release watch: after a release, re-runs your tests every 5, 10, 15, 20, 30 or 60 minutes for 1 to
24 hours. The first run is the "before". After that you hear once about each problem that is new
since then (a test that starts failing, or the tests no longer running), on the schedule's
notifications: Mac notification, Slack webhook or GitHub issue. When the time is up it turns itself
off and removes its system-scheduler entry; the Schedules list keeps it with how it went. Start one
from Schedules › New schedule › Release watch, from Chat, or with
fastpilot schedule watch tests/ --every 10 --for 2. - Ship check:
fastpilot ship-check <url>(or ashipCheckstep) looks for console errors,
failed requests, broken links, layout at phone, tablet and desktop width, and basic accessibility,
then gives one verdict: SHIP, SHIP WITH FIXES, DO NOT SHIP or INCONCLUSIVE, always with what was not
checked. It exits with 0 (SHIP or SHIP WITH FIXES), 1 (DO NOT SHIP) or 2 (INCONCLUSIVE). Links are
checked without your cookies unless you ask (--cookies); sign-out, delete, one-time and other
links that change something are never opened. - In Chat, a release check shows as a verdict card, labelled "the bot's summary" (the bot wrote
it from the checks it ran; fastpilot does not verify it). A bug report shows as a preview card
that is not sent anywhere, with Copy report and Copy as Markdown.
Dashboard
- History: runs and saved Chat conversations; delete one, several or all; optional auto-delete
after 7, 30 or 90 days. - Memory: one line per test (Reliable, Unstable, Failing, Set aside, New), with Set aside and
Forget, plus project notes (staging address, test account names, "login uses a code") that
agents read. Notes that look like a password or hide invisible text are refused. Notes are kept in
your own fastpilot folder, not in the project. An agent on another AI provider sends its notes to
that provider. - Device: Fix buttons for "Install Chrome" and "Set up an emulator" (you confirm the size and time;
only a person can start them). Devices are named by type ("Android emulator", "iPhone simulator"). - After
fastpilot initandfastpilot demo: "Open the dashboard now? [Y/n]". - Everything is in English: dashboard, terminal messages and the guide.
Setup
fastpilot setup iosworks without the source code: it builds the iPhone test runner once with
your Xcode into~/.fastpilot/ios-runner.- Android tests fail clearly when no emulator is there, instead of being skipped. They are also
more stable on a busy machine. - fastpilot checks your Claude Code version in
init,doctor, Chat andfastpilot run. It needs
2.1.280 or newer and tells you to runclaude updatewhen it is older. fastpilot setup --helpandfastpilot doctor --helpshow help instead of running.
Upgrading from 0.3.2
- Messages are now in English. New projects get
tests/example.md; an existingtests/contoh.mdis
kept. - Block words that start with
!, the worddefault, and words with a comma are refused from chat
and from agents. Only you can switch the built-in list off, in Settings. - fastpilot writes
.fastpilot/.gitignorein your project. It ignores everything except visual
baselines, and never overwrites a file you already have. - Project notes live in your own fastpilot folder (
~/.fastpilot/projects/), not in the project, so a
downloaded project cannot bring notes of its own.
Fixes
- History serves and deletes only reports in fastpilot's own folders.
- A schedule saved while a run is finishing is no longer lost.
- Session ids are checked before they name a file.
- "Save note" works when you type into the empty form.
Known limits
fastpilot run, schedules and Release watch read secret variables such as${DEMO_PASS}only from
the shell, not from.mcp.json(Chat and Claude Code do read.mcp.json). In the demo folder, run
export FASTPILOT_VARS=DEMO_PASS DEMO_PASS=<the demo password>first, or a saved test fails with
"variable DEMO_PASS is not allowed".- Claude Code in the terminal reads project notes when it starts: restart it after you change a note.
Dashboard Chat picks up changed notes on your next message. - Ship check reads the page from inside the page, so a hostile page can fake its own layout and
accessibility results (not which links are opened). Afterwards the screen size is put back from the
page's own numbers, within safe bounds.
Not in this release
- Jira and the Apps page (planned for 0.5).
- Not tested yet: models other than Claude, Linux, physical Android phones and iPhones.
fastpilot v0.4.0
What's new in 0.4.0
Talk to it, in plain words. Several QA agents, each with its own model and skills, that test like a
certified tester and show what they covered.
Agents
- Several agents, like a team. Each agent has its own Chat, Live view, Schedules, Instructions
(with suggestions you can click), Skills, History and settings. Create one with a short form, or
describe it in one sentence in Chat and click Create. Existing projects get one default agent
that keeps your chat, schedules and history. - Provider, then model, per agent. Claude Code, OpenAI, OpenRouter, Groq, Ollama (local) or another
OpenAI-compatible service. API keys are stored once per provider in Settings. Dashboard Chat now
works with any provider. Models other than Claude are labelled "not tested yet". - Agents run in parallel. Web tests run side by side. A phone or emulator runs one test at a time;
other agents wait in a queue for it. "Tests at once" limits how many run together (default 3).
Tests that think like a certified tester
- 21 built-in skills, grouped by ISTQB® test design technique (fastpilot is not affiliated with
ISTQB): Choose technique, Equivalence partitioning, Boundary value analysis, Decision table, State
transition, Pairwise, Use case scenarios, Error guessing, Exploratory testing, Checklist-based (with
checklists for login and OTP, forms, safe checkout, accessibility and mobile), Acceptance tests,
Regression, Bug report, Test report and Release check.fastpilot initandfastpilot demoinstall
them. - "Techniques used": results say which technique was used, which values were tried, what was
assumed, and what was covered (for example "Boundary value analysis · Age — 6 of 6 values"). - Regression puts what is newly failing since the last run first, then still failing, flaky,
blocked and passed. - Practice pages in
fastpilot demo: a ticket field with limits, shipping rules and an order
status flow, each with one planted bug.
Everything through chat
- Chat creates a test from a sentence. One card shows the steps; Save & schedule or Save test
only. Nothing is written before you click, and variables the test needs are checked first. - All settings through chat, in the dashboard and in Claude Code:
- making limits stricter applies at once, with Undo;
- other changes need your click;
- loosening a limit only appears when you asked for it in your own words, as a red card that
quotes you; allowing Pay also needs you to type "Pay" and expires after 10 minutes; - text on a tested page can never ask for a change;
- in the terminal, the AI can only file a request; you approve it yourself with
fastpilot config approve; - every change is written to an activity log.
- "Remember this as a note": a card in Chat saves a project note from your own words.
Schedules and releases
- Every N hours (1–24) with an active window (for example 08:00–20:00), Run now, and Next / Last
columns. - The Schedules table says why a run failed (which test, and the reason), not only "0 of 1 passed".
- Release watch: after a release, re-runs your tests every 5, 10, 15, 20, 30 or 60 minutes for 1 to
24 hours. The first run is the "before". After that you hear once about each problem that is new
since then (a test that starts failing, or the tests no longer running), on the schedule's
notifications: Mac notification, Slack webhook or GitHub issue. When the time is up it turns itself
off and removes its system-scheduler entry; the Schedules list keeps it with how it went. Start one
from Schedules › New schedule › Release watch, from Chat, or with
fastpilot schedule watch tests/ --every 10 --for 2. - Ship check:
fastpilot ship-check <url>(or ashipCheckstep) looks for console errors,
failed requests, broken links, layout at phone, tablet and desktop width, and basic accessibility,
then gives one verdict: SHIP, SHIP WITH FIXES, DO NOT SHIP or INCONCLUSIVE, always with what was not
checked. It exits with 0 (SHIP or SHIP WITH FIXES), 1 (DO NOT SHIP) or 2 (INCONCLUSIVE). Links are
checked without your cookies unless you ask (--cookies); sign-out, delete, one-time and other
links that change something are never opened. - In Chat, a release check shows as a verdict card, labelled "the bot's summary" (the bot wrote
it from the checks it ran; fastpilot does not verify it). A bug report shows as a preview card
that is not sent anywhere, with Copy report and Copy as Markdown.
Dashboard
- History: runs and saved Chat conversations; delete one, several or all; optional auto-delete
after 7, 30 or 90 days. - Memory: one line per test (Reliable, Unstable, Failing, Set aside, New), with Set aside and
Forget, plus project notes (staging address, test account names, "login uses a code") that
agents read. Notes that look like a password or hide invisible text are refused. Notes are kept in
your own fastpilot folder, not in the project. An agent on another AI provider sends its notes to
that provider. - Device: Fix buttons for "Install Chrome" and "Set up an emulator" (you confirm the size and time;
only a person can start them). Devices are named by type ("Android emulator", "iPhone simulator"). - After
fastpilot initandfastpilot demo: "Open the dashboard now? [Y/n]". - Everything is in English: dashboard, terminal messages and the guide.
Setup
fastpilot setup iosworks without the source code: it builds the iPhone test runner once with
your Xcode into~/.fastpilot/ios-runner.- Android tests fail clearly when no emulator is there, instead of being skipped. They are also
more stable on a busy machine. - fastpilot checks your Claude Code version in
init,doctor, Chat andfastpilot run. It needs
2.1.280 or newer and tells you to runclaude updatewhen it is older. fastpilot setup --helpandfastpilot doctor --helpshow help instead of running.
Upgrading from 0.3.2
- Messages are now in English. New projects get
tests/example.md; an existingtests/contoh.mdis
kept. - Block words that start with
!, the worddefault, and words with a comma are refused from chat
and from agents. Only you can switch the built-in list off, in Settings. - fastpilot writes
.fastpilot/.gitignorein your project. It ignores everything except visual
baselines, and never overwrites a file you already have. - Project notes live in your own fastpilot folder (
~/.fastpilot/projects/), not in the project, so a
downloaded project cannot bring notes of its own.
Fixes
- History serves and deletes only reports in fastpilot's own folders.
- A schedule saved while a run is finishing is no longer lost.
- Session ids are checked before they name a file.
- "Save note" works when you type into the empty form.
Known limits
fastpilot run --workersuses every connected Android device and ignores the device you picked
withFASTPILOT_ANDROID_SERIAL/ANDROID_SERIAL(this was already so in 0.3.x). Until 0.4.1, run
Android tests with one device connected, or without--workers.fastpilot run, schedules and Release watch read secret variables such as${DEMO_PASS}only from
the shell, not from.mcp.json(Chat and Claude Code do read.mcp.json). In the demo folder, run
export FASTPILOT_VARS=DEMO_PASS DEMO_PASS=<the demo password>first, or a saved test fails with
"variable DEMO_PASS is not allowed".- Claude Code in the terminal reads project notes when it starts: restart it after you change a note.
Dashboard Chat picks up changed notes on your next message. - Ship check reads the page from inside the page, so a hostile page can fake its own layout and
accessibility results (not which links are opened). Afterwards the screen size is put back from the
page's own numbers, within safe bounds.
Not in this release
- Jira and the Apps page (planned for 0.5).
- Not tested yet: models other than Claude, Linux, physical Android phones and iPhones.
fastpilot v0.3.2
What's new in 0.3.2
Easier to start, safe by default.
New
fastpilot demo: writes a demo folder (demo pages, three example tests,.mcp.jsonwith the
public demo password, and the fastpilot Skill) from the binary alone, so a first test runs without
the repository:fastpilot demo && cd fastpilot-demo && claude, then type
"Pakai fastpilot, jalankan tests/login.md lalu report".--androidinstalls the demo app (built
into the binary) on the attached device or emulator. An existing folder is never overwritten
(--forcewrites it again and keeps your own.mcp.json, saving the old one as.mcp.json.bak).fastpilot init: sets up a project folder in one command: checks the machine, registers
fastpilot with Claude Code (claude mcp add), installs the Skill, addstests/contoh.md, and shows
the limits in force. Every change asks first (--yesagrees to all); outside a terminal it only
prints the plan; a second run changes nothing.--globaluses the user scope.- Safe by default: on a new install, buttons for payment and account deletion are always blocked
(block_actions: bayar, pay, beli, buy, checkout, purchase, transfer, buat pesanan, place order,
hapus akun, delete account; matched as whole words, also through spacing, full-width and zero-width
tricks, e.g. "Pay now", "Delete my account", "Place your order"), and plain deletes need
confirm: true(confirm_actions: hapus, delete, remove), so "Hapus filter" still works.
Pengaturan › Batasan & izin shows the built-in list with Kosongkan and Kembalikan ke bawaan.
The AI can only add to these lists, never empty or replace them. - README in English, simpler: three steps (install → demo → init); the full guide moved to
docs/guide.md(still Indonesian for now). Messages in the terminal for people (doctor,setup,
init,demo,config list,help) are in Indonesian; what the AI reads (tool results, Skill)
stays English.
Upgrading from 0.3.1: what changes for existing tests
- Flows that tap Bayar / Pay / Buy / Checkout / Purchase / Transfer / Place order / Hapus akun /
Delete account are now blocked, also withconfirm: true, until you change the list. Steps that
tap "Hapus" / "Delete" / "Remove" needconfirm: true. evalScriptis off while a block list is active (the built-in list counts), because a script
could press a blocked button. To allow it:fastpilot config set allow_eval_script true(only you
can; the AI cannot), or the "Jalankan JavaScript bebas" switch on the Batasan page.- To go back to no blocking, in one line:
fastpilot config set block_actions ""(or the
Kosongkan button), then restart the MCP server.fastpilot config unset block_actionsrestores the
built-in list. An empty environment variable does not switch the built-in list off. - Words are matched whole: "Pembayaran" (or "Paypal") is not blocked by "bayar"/"pay". Add your
own words if your app needs them (default,<word>keeps the built-in list).
Known limitations
- The built-in list is a list, not a guarantee: text written with invisible marks inside a word
(left-to-right marks, combining characters), look-alike letters or math letters can slip past it. - An environment value such as
FASTPILOT_BLOCK_ACTIONS=","still empties the list; environment
and.mcp.jsonare trusted (whoever can write them can also run other programs). demo --forceon a.mcp.jsonthat is a symlink copies its target into.mcp.json.bakbefore the
write is refused;demo --androiddoes not checkallow_physical_devices;initcounts any
fastpilotentry in.mcp.jsonas registered.demo --androidfrom a source build needs the demo app built first (release builds include it).config setvalidation errors are still in English;setup --helpanddoctor --helprun the
command instead of showing help;config listshows a mixed "(default (bawaan))" label.- Carried over: a refused takeover from another installation sometimes shows a misleading message;
Android WebView and system-dialog tests can be flaky on a heavily loaded machine. - Not tested: real Claude Code answering
y/Nin a terminal,npxdetection from the npm registry,
Linux, physical phones, real hosted providers and GitHub posting, the real Keychain, iOS demo
(comes withfastpilot setup ios, next release).
Install (macOS arm64/x86_64, Linux x86_64):
curl -fsSL https://raw.githubusercontent.com/autobot-lab/fastpilot/main/scripts/install.sh | sh
npm install -g fastpilotEvery archive has a .sha256 next to it; the install script and the npm package check it.
Earlier releases: https://github.com/autobot-lab/fastpilot/releases (0.3.1: takeover only from the same installation and allow_takeover, provider key and redirect fixes; 0.3.0: Live Android/iOS + Ambil alih, Jadwal, Chat, OpenAI-compatible providers).
fastpilot v0.3.1
What's new in 0.3.1
A fix release for the known limitations of 0.3.0.
Fixes
- Ambil alih (takeover): the control socket now serves only a
fastpilot uirunning the same
program file as the fastpilot process that owns the session (peer pid and executable path are
checked); a program merely namedfastpilotis refused. New limitallow_takeover(default on,
tighten-only, pinned at start; Pengaturan › Batasan): turn it off when your agent has a shell. Run
fastpilot uithe same way as its MCP server (e.g.npx -y fastpilot uiwithnpx -y fastpilot mcp). - PR comments (
--report-github): only a marked comment written by the token's own account is
edited; a copied marker in someone else's comment is left alone. - Provider keys: a key saved by 0.2.0 (no server recorded) is not sent to a custom or local server;
runfastpilot config set-key <provider>once. Provider redirects are no longer followed (the key and
the request body never go to the redirect target). - Flows with their own
sessionstep run without anopenfirst (unknown session default, #9). - Limits: a tap by text,
id,reforrolewaits for its element like the tap itself and is
refused while a rule is on and the element cannot be seen (fail closed), so a "Bayar" button that
appears late is still blocked. - Reports: the video player iframe is sandboxed (
allow-scripts, no same origin). - UI: Jadwal messages are all Indonesian; after "Kembalikan ke AI" focus goes to the page title;
Jadwal field errors are tied to their fields; Chat namesvarsas "Nama secret"; a stopped
fastpilot uiis reported in plain words instead of "Failed to fetch".
Known limitations
- Takeover is still not a security boundary against an agent that can run the real
fastpilot ui
(it has a shell): turnallow_takeoveroff for such agents. Withallow_takeover = false, watching
an Android/iOS screen in Live still works; only steering is off. - After upgrading on Linux, a fastpilot process started before the upgrade refuses takeover until it
is restarted. --report-githubwith a GitHub App installation token can post a new comment on every run (use the
workflow'sGITHUB_TOKENor a personal token).- A refused takeover from another installation sometimes shows "proses fastpilot itu tidak menerima
kendali" instead of the "same installation" message. - With an OpenAI-compatible provider, what the model sees (snapshots, step results, local paths) goes to
that provider;${NAME}values stay masked. Use Claude Code or a local Ollama when data must stay on
the machine. - Android system-dialog and WebView tests can be flaky on a heavily loaded machine (as in 0.3.0).
- Not tested on: Linux, physical phones, real hosted providers and Ollama, real GitHub posting.
Install (macOS arm64/x86_64, Linux x86_64):
curl -fsSL https://raw.githubusercontent.com/autobot-lab/fastpilot/main/scripts/install.sh | sh
npm install -g fastpilotEvery archive has a .sha256 next to it; the install script and the npm package check it.
Earlier releases: https://github.com/autobot-lab/fastpilot/releases (0.3.0: Live Android/iOS + Ambil alih, Jadwal, Chat, OpenAI-compatible providers, --reset, video per flow, --report-github).
fastpilot v0.3.0
What's new in 0.3.0
New
- Live for Android and iOS, and "Ambil alih" (take over) in
fastpilot ui: Android and iOS screens
(screenshots from the device, only while the panel is open), and for web, Android and iOS you can take
a session over, tap, swipe and type yourself while the AI's steps wait, then give it back. The AI's
next result says the user stepped in, and runs withtest_idrecord it. A yellow banner shows a
takeover on every page; closing the large view gives the session back. - Jadwal (schedules) + notifications: run a test folder or file every day, on weekdays or weekly,
through the OS scheduler (launchd on macOS, a systemd user timer or cron on Linux), registered only
when you save. On failure: a Slack-compatible https webhook (the URL is kept as a secret), a
notification on this computer, and/or a GitHub issue.fastpilot schedule list | run | remove. - Chat in
fastpilot ui: talk to a fastpilot bot driven by your own Claude Code (claude -p,
no API key) with only the fastpilot tools. It can only propose schedules and stricter limits, as
cards in plain Indonesian; nothing changes until you click, and loosening is refused. - OpenAI-compatible providers (opt-in):
fastpilot runcan be driven by OpenAI, OpenRouter, Groq,
Ollama or a custom chat-completions API with tool calling (provider,provider_model,
config set-key,config test-provider, or Pengaturan › Model AI). Claude Code stays the default.
A key is stored in the Keychain / secret service (or an owner-only file) and only sent to the
server it was saved for; a provider that cannot be used fails the test with the reason instead of
falling back silently. --reset: every mobile session starts its app fresh (data, permissions, orientation), also when
the app is opened withlaunchAppaftersession start.- Video per flow in the HTML report (
run --record): markers per step (red where a step failed),
a button per flow, and the final screen of passing flows. --report-github: the results as one comment on the pull request (updated, not repeated).action_timeoutsetting: how long an action waits for its element; text input on phones waits
longer for the keyboard.- Docs: running fastpilot on another machine over SSH, and lighter iOS simulators with simslim
(docs/remote-machine.md; the SSH parts are marked untested).
Security fixes
- The limits gaps listed for 0.2.0 are closed: a
pointtap uses the same coordinates as the
snapshot and checks every element under the point;pressKeyEnter andevalScriptobey
block_actions(evalScript is off whileblock_actionsis set); the Chrome extension never
connects without a relay token (extension 0.2.1). - Takeover socket (
~/.fastpilot/live/<pid>.sock): served only to afastpilot uiprocess, a random
per-takeover token kept in memory is needed for every input, and a takeover ends by itself when the
page goes away (±45 s) or after 30 minutes. - Switching provider no longer sends a key to the previous provider's server.
fastpilot uikeeps running when whoever started it stops reading its output.
Privacy: third-party providers
With an OpenAI-compatible provider, what the model sees goes to that provider: screen snapshots,
step results and errors, which can include local paths such as file:///Users/<name>/…. ${NAME}
secret values stay masked. Use Claude Code (default) or a local Ollama when data must not leave the
machine.
Known limitations
- Takeover is not a security boundary against code that already runs as you: an agent with shell
access can start its ownfastpilot ui(or a program namedfastpilotrun withui) and take a
session over without a human click; takeover input is not checked by the limits. fastpilot's limits
protect against an AI that only has fastpilot's tools. A renamed fastpilot binary cannot use
takeover ("only fastpilot ui may use this socket"). - With
--report-github, if someone posts a comment carrying fastpilot's marker before the bot does,
the report can be written into the wrong thread (Low). - A provider key stored by 0.2.0 has no server recorded; set it again with
config set-keyso it is
bound to its server. Provider requests follow HTTP redirects (Low). - A flow with its own
sessionstep but no session started yet can still fail with
unknown session default(#9). - A non-point selector is not checked against the limits when the snapshot fails; a point covered
only by nameless elements passes (Low). - The report's video player sits in an
<iframe srcdoc>withoutsandbox(Low). - Android system-dialog tests can be flaky on a heavily loaded machine.
- Not tested on: Linux (systemd/cron, secret-tool, peer check), physical phones for takeover, real
hosted providers and Ollama, SSH, the Chrome extension in daily Chrome, VoiceOver.
Install (macOS arm64/x86_64, Linux x86_64):
curl -fsSL https://raw.githubusercontent.com/autobot-lab/fastpilot/main/scripts/install.sh | sh
npm install -g fastpilotEvery archive has a .sha256 next to it; the install script and the npm package check it.
Earlier releases: https://github.com/autobot-lab/fastpilot/releases (0.2.0: fastpilot ui, new limits, launchApp arguments, remote devices, Agent Skill).
fastpilot v0.2.0
What's new in 0.2.0
Upgrading from the Chrome extension 0.1.0: the extension now has a fixed ID and gets its relay
token through native messaging. Reload it once (chrome://extensions → fastpilot → reload).
New
fastpilot ui: a local page (127.0.0.1 only) with Live, Settings and Device.- The printed link works once and for 10 minutes; it is traded for a session cookie, and it is
never passed in another process' arguments. Fonts and assets are built into the binary (no CDN). - Settings: AI models for
fastpilot run, limits and permissions, secret names (values never shown).
A brokenconfig.tomlis shown with its line and column, and nothing is saved until it is fixed. - Device:
doctorresults with a re-check button. - Live: the running sessions of
fastpilot mcp, the shell CLI andfastpilot run, each with its
steps (done, running, failed, blocked, and what a limit will block), and the screen of web
sessions on fastpilot's own Chrome, read-only and only while the panel is open.
- The printed link works once and for 10 minutes; it is traded for a session cookie, and it is
- New limits, enforced by the engine (an agent can only make them stricter):
block_actions(taps on text containing a word are always refused, even withconfirm: true),
allow_app_install,allow_app_reset,allow_physical_devices,allow_device_settings,
allow_eval_script,upload_dirs,max_stepsandtest_timeoutforfastpilot run.
New stepsinstallApp/uninstallApp. launchAppwithargumentsandenv(E1): Android intent extras, iOS launch arguments and
environment. Env values always show as***; names such asDYLD_*andFASTPILOT_*are refused.- Ambiguous selectors (E2): an
idor a namelessrolethat matches several elements fails with
the candidates instead of picking the first;index: Npicks one. - Devices and Chrome on another machine (D2):
adb_server/FASTPILOT_ADB_SERVERand
chrome_endpoint/FASTPILOT_CHROME_ENDPOINT, through an SSH tunnel (see the README). - Agent Skill (B4):
skills/fastpilot/SKILL.mdteaches an AI agent how to test with fastpilot. - Recovery: when fastpilot's Chrome crashes or is closed, its sessions end with a clear message
and the next start launches a new Chrome. fastpilot doctorexits non-zero when web is not ready;fastpilot ui --help; clearer errors for
ports and settings.
Security fixes
- Android: values passed to the device shell (
appId,launchApp,uninstallApp,setLocale,
setPermissions,pushNotification) are quoted or refused; shell injection is no longer possible. tapOn: { point: [x, y] }is checked againstblock_actionsandconfirm_actionslike a named tap.- Chrome extension relay: a random 256-bit token per session is required on both routes; the token
file only exists while waiting for the extension. - Guard rails, program paths and remote endpoints are pinned when fastpilot starts; a looser value
written to the settings file applies only after a restart. Remote endpoints that are not loopback
get a warning. - Settings values fastpilot would misread stop it with the key and the reason.
Known limitations
- After a page is scrolled, a
tapOn: { point }taken from snapshot coordinates can miss but still
report success (since 0.1.2). Use refs or text selectors. block_actionsfilters taps:pressKey(e.g. Enter on a focused button) andevalScriptare not
covered. Setallow_eval_script = falsewhen you rely onblock_actions.- A
pointtap on an element without a name is refused whileblock_actionsorconfirm_actions
is set (it cannot be told what it is). - The Chrome extension still connects when the relay gives no token (Low).
- On a slow emulator the soft keyboard can take longer than the 5 s action limit; run emulators
headless (emulator -avd <name> -no-window).
Install (macOS arm64/x86_64, Linux x86_64):
curl -fsSL https://raw.githubusercontent.com/autobot-lab/fastpilot/main/scripts/install.sh | sh
npm install -g fastpilotEvery archive has a .sha256 next to it; the install script and the npm package check it.
fastpilot v0.1.2
What's new in 0.1.2
fastpilot doctorchecks web, Android, iOS and Claude Code, with a fix hint for each ✗;
fastpilot setupapplies the fixes it can, asking before each one.- Settings file
~/.fastpilot/config.toml,fastpilot config list|get|set|unset|path, and a
configMCP tool. Environment variables still win. Guard rails (allowed_domains,
confirm_actions,vars,state_expire_days) are pinned when fastpilot starts; an agent can only
make them stricter. - README rewritten for the npm / curl install.
Install (macOS arm64/x86_64, Linux x86_64):
curl -fsSL https://raw.githubusercontent.com/autobot-lab/fastpilot/main/scripts/install.sh | sh
npm install -g fastpilotEvery archive has a .sha256 next to it; the install script and the npm package check it.