Repository navigation
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).