Repository navigation
Releases
keel v0.33.0
keel v0.33.0
keel's code graph (its RAG) is now on everywhere.
What changes for you
- Every agent and KeelBot use the code graph. Before, it was off for all of them, so agents found code only with Grep and often read whole files. Now the code graph's search sits next to Grep: the agent asks "where is X?" and gets file:line, then reads only those lines. Its other tools (callers, callees, impact) load when an agent needs them.
- The agent calls the search directly. No extra LoadTools step first.
- You can still turn it off per agent (Agents page › an agent › Knowledge › Code graph tools).
- The code graph needs CodeGraph, which is already in the desktop app and the Docker image, and a ready index (Code page).
Measured (GitHub Copilot auto — the test machine's Copilot Free plan has no other model; KeelBot; one run per row unless noted)
"hi" (KeelBot, Edit mode): code graph off 2,030 new + 3,537 cached; on 2,114–2,126 new + 3,819 cached. The one extra base tool costs about 0.3k tokens, almost all from the cache.
Small demo project (5 files):
| question | off (2 runs) | on, fixed prompt |
|---|---|---|
| where is the price with tax computed, which tests | 8.3k new / 13.7k cached; 2.3k / 14.8k | 2.7k new / 15.9k cached, 1 codegraph_search |
where is enough defined, how many versions |
7.3k / 9.1k; 4.4k / 23.0k | 3.0k / 22.5k, 1 search + 2 Grep |
An open-source Python project (aider, 121 files, 2.0k symbols):
| question | off | on, fixed prompt |
|---|---|---|
| where is the repo map built, who calls it | 11.3k new / 78.8k cached, 8 calls, 5 Grep | 15.0k / 48.0k, 6 calls, search + callers |
| how is an edit block applied | 2.3k / 9.1k — no answer: the search was blocked by the secrets guard | 6.5k / 44.0k, a full answer with file:line |
| which function decides a model's edit format | 8.7k / 48.7k, 7 calls | 9.8k / 91.3k, 11 calls, a more exact answer |
What it shows: with the graph on, KeelBot finds code with codegraph_search first and gives answers with the right file:line. Tokens are lower on the small project and mixed on the larger one (one run each, so noisy); the answers were as good or better each time, and one question was only answered with the graph on.
The first "on" pass (before the prompt fix) showed the model calling LoadTools for the search tool it already had (one wasted model call per question); the fixed prompt removed that call in every answer.
Not measured: Claude or OpenAI keys (none on the test machine), flows (only KeelBot), a paid Copilot plan's models.
JetBrains plugin: keel-jetbrains-0.5.1.zip below (Settings › Plugins › ⚙ › Install Plugin from Disk).
keel v0.32.0
keel v0.32.0
The next step of making keel's agent harness cheaper, from the research report on the best methods (Claude Code, Codex, OpenCode, Anthropic and OpenAI guidance).
What changes for you
- Fewer tokens in KeelBot chats. Old tool results are cleared (a one-line note stays), every tool's output has one size limit (the full output is saved to a file the agent can read in parts), and long chats keep a recent part of about 16k tokens when they are summarised.
- The cache is shown. Under each model call, "Show what it did" says how many tokens were new, written to the cache and read from it, and warns with "cache break" when the cache was lost.
- GitHub Copilot: each KeelBot answer shows about how many AI credits it cost, and the model list shows each model's price multiplier (for example ×9).
- Claude (Anthropic API): tools load on demand without breaking the cache (Anthropic's own tool search), and KeelBot chats keep their cache for an hour.
- Long flows: a summary now keeps keel's own progress (the task list, files changed, the last files read, the project's notes), and helper agents give short answers.
- Code graph: a script measures KeelBot with and without it (scripts/measure-code-graph.py). It stays off until measured.
Measured (GitHub Copilot auto, KeelBot in Edit mode, 8 plugins on, a demo project)
| 0.31.0 | 0.32.0 | |
|---|---|---|
| "hi" (new tokens) | 5,568 | 2,031 (+3,537 from cache) |
| a code question | 4,147 new | 3,635 new |
| a follow-up in the same chat | 2,132 new | 2,136 new |
| the whole chat (keel's count) | about 16.6k | about 12.7k |
GitHub Copilot's own tool search was measured too and was more expensive (a "hi" cost 9.5k tokens with it), so Copilot keeps keel's own on-demand tools.
Only tested with test doubles (no key on the test machine)
- Anthropic's tool search, the 1-hour cache and server-side compaction counts.
- OpenAI's cache key and cache-write counts.
- A paid Copilot plan's price multipliers (the test machine has Copilot Free).
JetBrains plugin: keel-jetbrains-0.5.1.zip below (Settings › Plugins › ⚙ › Install Plugin from Disk).
keel v0.31.0
keel v0.31.0
Far fewer tokens per message
keel now works like Claude Code: a small set of tools first and the rest loaded only when needed (LoadTools), one live GitHub Copilot session per KeelBot chat (the next message goes on where Copilot is, mostly from its cache), a short KeelBot prompt unless you ask about keel's own work, and earlier compaction of long chats.
Measured on GitHub Copilot (auto), KeelBot in Edit mode with 8 plugins on, the same demo project:
| Message | 0.30.0 | 0.31.0 |
|---|---|---|
| "hi" | 23.1k tokens | 5.6k |
| a code question | 65.7k | 7.0k |
| a follow-up in the same chat | 60.3k | 4.4k |
| the whole chat | about 149k | about 17k |
"Show what it did" now has a context: line per model call (system, tools, chat, new; new and cached tokens), and each KeelBot answer shows its new / cached / output tokens.
GitHub Copilot (your plan)
- The model list: Settings › Models & keys › GitHub Copilot › Test. The picker shows your plan's models first, with their efforts, then every other model Copilot knows ("not in your plan's list"; if your plan lacks it, the step runs on auto and says so).
- The effort can be picked for each model that takes one.
- The row is Ready with this computer's own GitHub login (gh, the Copilot CLI) after a passing Test.
- Cached tokens are no longer counted twice.
keel Product in every keel
Product is in the normal image, the native install and the keel app, off until you turn it on in Add features (Beta). There is no separate product image any more; keel2 update moves a :product-beta keel to the normal one and keeps Product on.
Fixes from an end-to-end walk
- Start a flow: the "Run init first / Start anyway" question and a refused start stay in view.
- KeelBot says "KeelBot has no model yet" instead of answering with the fake model.
- A keel without CodeGraph is no "Index failed" alarm; the keel app now carries CodeGraph.
- A plugin that is off for a project turns on from its own page.
- Messages name Settings › Plugins / Connections / Databases.
JetBrains plugin: keel-jetbrains-0.5.1.zip below.
keel v0.30.0
keel v0.30.0
GitHub Copilot (your plan) (#146)
GitHub retired GitHub Models; its row is gone. The new row GitHub Copilot (your plan) runs keel's harness on GitHub's official Copilot SDK with your own Copilot plan. Copilot's own tools are off: keel's tools and rules check every call. Sign in with GitHub or paste a token (gho_, ghu_, github_pat_). Models come from your plan; auto is always there.
Connected session without a key (#147)
A Connected-session flow now starts on a keel with no API key (it was refused with "would use the fake model").
keel desktop
- MCP for the app's keel (#148): Help › Connect Claude Code or OpenCode… shows and copies the commands.
- Keeps running when the window closes (#149): tray / menu-bar icon, "waits for you" notifications, Start at login, Quit asks while flows run.
- Carries the Copilot runtime (no download at run time).
JetBrains plugin: keel-jetbrains-0.5.1.zip below (Settings › Plugins › ⚙ › Install Plugin from Disk).
keel v0.29.1
keel v0.29.1
JetBrains plugin 0.5.1 (download keel-jetbrains-0.5.1.zip below; Settings › Plugins › ⚙ › Install Plugin from Disk)
- Review: comment on any line of a PR file or on a selection (right-click › keel: Comment on This Line / Selection, or Ctrl+Alt+Shift+C). A clear message when no PR is open; Show Pending after saving. (#140)
- Review › Pending: click a comment to go to its line, Edit, Delete. Loading states in the PR list and PR view. (#140)
- Review › Changed files: search (words, or
*.php) and folders you can fold. (#142) - KeelBot answers as Markdown: code blocks, lists, bold, links;
file:lineopens the file. (#141) - Flows open every time, with a clear error and Try again. (#141)
- The plugin page links to github.com/MiladNalbandi/keel-v2.
keel
keel v2 0.29.0
keel v2 0.29.0
Your own coding agent can do keel's work. Run the harness flow with the Claude Code or OpenCode session you already have open and connected to keel's MCP server: keel sends it work orders and keeps the gates, the tests and the commits. A project can also turn the whole test suite before push off when CI runs it. Plus fixes for the Inbox, the Review page and GitHub Models.
New
- Run with a connected session (#136)
- Start a flow › Run with: keel's harness · API key or Connected session · claude-code (shown when a session is connected).
- Connect once:
claude mcp add keel -- keel2 mcp --write(OpenCode: see Help › Start here). Then tell the session: "Work for keel: call keel_session_open, then keel_session_next, and follow its work orders." - keel → session: plan this → the session reports its plan and todo list → you approve the plan in keel → build the plan → the todo list updates live on the Flow page → keel checks the change, runs the tests (a failure goes back as fix), you look, keel commits.
- The session runs its own tools, so keel checks its change after each step and puts back what the step may not change. keel never starts the app and never sees its login.
- A session gone for 10 minutes: keep waiting, run this step on keel's harness, or stop.
- Whole test suite before push: required / off (#137)
- Settings › Flow and gates, per project (
.keel/config.ymlrelease: {required: false}). - Off: ship skips verify release, the push is not blocked, and keel shows a warning instead.
- From MCP (
keel_release_check), turning it off waits for your approval in the Inbox; an agent cannot switch it off alone.
- Settings › Flow and gates, per project (
Fixed
- Inbox (#133): on a mid-size window the selected item no longer slides sideways and the list keeps its width; long paths wrap.
- Review (#134): the PR detail no longer covers the list, the search icon has room, the detail uses the page's width, the overview is a quiet fold, and the tabs wrap.
- GitHub Models (#135): your GitHub sign-in now comes before an old Copilot login's token; a Copilot-only model and GitHub Models' 401 / 403 / 404 now say what to do. GitHub Models is not Copilot: a Copilot plan's models work only in GitHub's own apps.
- Config changes from the Rules page and the project notes switch reach the engine again (#138).
Good to know
- Your data stays as it is. Projects keep the whole test suite required unless you turn it off.
keel v2 0.28.0
keel v2 0.28.0
keel's new look. A calmer menu, a Home page that asks KeelBot first, and every page redesigned to be handier: a clear page head, calm cards and tables, quiet motion, and the keys shown on the buttons. Light and dark, phone width, and reduced motion everywhere.
New
- A calmer menu and a top bar (#122)
- The menu keeps six main places: Home, Inbox, Flow, Code, KeelBot and Live agents. Below them: your pinned pages and the recent flows.
- All pages (the menu's foot, or ⌘K) lists every page by group, with search and a pin on each.
- The top bar shows where you are (project › page) with a Pin toggle, the budget, search and the bell.
- N (or the menu's New flow) opens Start a flow from anywhere.
- Home (#122): one box that asks KeelBot first; Tab switches it to Build a flow. Below: what waits for you, what runs now and the recent flows.
- Add features (#122): keel's plugins as cards by kind, where each is on (off / this project / all projects), Open, Pin to menu, and Build your own feature (a flow with the plugin workflow).
- Flow (#125)
- The coloured steps sit in a rail that folds («). J / K follow the steps, G jumps to the current one.
- The step's work is a calm feed: one turn per agent call; each tool call or edit is one line that opens the full step.
- The gate card is in the feed with its keys: Approve A, Send back R, Read the change O. Your answer shows in the feed, with a way back to a checkpoint.
- Now, Step, Decisions, Checkpoints, Budget, Criteria, Before ship and the add-ons' panels are tabs in one Details drawer that D folds.
- Inbox (#126): a list and the selected item's card. J/K move, A approves, R goes to the reason, 1–9 pick a choice, O opens the flow, L is later on tasks. After an answer the next item is selected. Filters show their counts.
- Code (#124): a calm reader with the outline, "This file in keel", a keys strip and a clean ⌘P.
- keel desktop (#123): the new look on its own pages; File › New Flow… and a Go menu.
Changed
- Watch pages (#127): KeelBot is a calm chat with one compose box; Live agents, Jobs, Flow doctor (what waits for you first, Copy C), Record (search and status chips, Export E), Cluster and All projects (Add A).
- Work pages (#128): Tasks, Goals, Org, Desk, Review, Map, Graph, Wiki and Library, with their keys (N to add, M for a mission on Org, R to rebuild the Map). Review's thread check words stay the same.
- Build pages (#129): Workflows, Agents, Skills, Stacks, Tools, Safety & History, Quality, Settings and Start here, with their keys. API keys are still never shown again.
- Product, Pipelines and Supervisor (#130): card-based for people who plan work. Pipelines' Show filter gains Passed and counts.
Fixed
- Jobs › Pipelines and Jobs › Supervisor no longer crash in newer Chromium ("destroy is not a function") (#131).
Good to know
- keel's api and engine did not change. Your data stays as it is.
- Answers cannot be undone (keel's api has no way to take an answer back); the way back is a checkpoint.
keel v2 0.27.1
keel v2 0.27.1
keel's other apps catch up with 0.26. The Android app, the JetBrains plugin and keel TUI now show what keel 0.26 added: API keys only, the flow doctor and Code Review's thread check. The web now shows the right choices when a step crashes.
Changed
- Web: a crashed step's pause (#116)
- When a step that is not an agent step crashes and keel cannot repair it, the flow doctor pauses the flow.
- The Inbox card says "step failed" and shows keel's own choices: Retry the step / Stop the flow. A note is optional. Before, it showed Approve / Send back.
- The flow page's gate shows the same choices.
- In auto mode, the card says why it still waits: only you decide to retry the step or stop.
- Android: API keys only and the flow doctor (#111)
- The model sheet shows how keel reaches the provider: API key, API key (GitHub Models), Google Cloud or Local server.
- A failed step shows as "Step failed" in the Inbox, the Flows list and the flow. It answers with keel's choices: Try again / Stop the flow, or Retry the step / Stop the flow for a crash.
- Android: Code Review's thread check (#112). The same words as the web:
- "Reviewer (bo) is right"
- "Same point as the thread on line N"
- how many files the check read
- "Fixed in abc1234" with Resolve, or "Still open"
- Android: motion that follows Material 3 (#113)
- Screens slide in and out, and the back gesture follows your finger (predictive back).
- Tabs fade through. List cards fade and slide into place.
- The system's "Remove animations" setting turns it all off.
- JetBrains plugin 0.5.0: Code Review's thread check (#114), in the editor's threads and the review panel, with the web's words.
- keel TUI:
/doctor(#115) shows the flow doctor's report. - Docs: KeelBot keeps its chat's session for every model.
Good to know
- keel's api and engine code did not change. Your data stays as it is.
- keel desktop 0.27.1 has these web changes too. The files appear here about 30 minutes after the release.
- The JetBrains plugin zip is built by the
jetbrainsworkflow, as before.
keel v2 0.27.0
keel v2 0.27.0
keel desktop (preview): keel without Docker, in one download. Open the app and keel runs. A small Java, a Python with keel's packages and ripgrep are inside; your computer only needs git.
Download
| Computer | File |
|---|---|
| Linux (x64) | keel-desktop-0.27.0-linux-x86_64.AppImage: make it executable (chmod +x), then open it |
| macOS (Apple chips) | keel-desktop-0.27.0-mac-arm64.dmg: drag keel to Applications |
The macOS app is not signed by Apple yet. The first time, macOS says it cannot check it: open System Settings › Privacy & Security and choose Open Anyway, or run xattr -dr com.apple.quarantine /Applications/keel.app.
The files appear here about 30 minutes after the release (they are built and tested first). SHA256SUMS-desktop-*.txt has their checksums.
New
- keel desktop (#118)
- First start: keel says that its agents run as you (your files, your SSH keys, your network). Then choose a project folder, or try the demo project.
- keel opens in the app's window, already signed in. File › Open in Browser opens it in your own browser.
- File › Add Project Folder adds more folders; every git repository in them becomes a project.
- Agents use your own programs and your own PATH (npm, uv, pytest …), never the app's Java or Python.
- Quit stops keel. When agents are working, the app asks first.
- Data stays outside the app:
~/Library/Application Support/keelon macOS,~/.local/share/keelon Linux. A new version is a new download; your data stays.
Changed
- Native mode runs the Python api in front (#117), as the Docker image does:
keel2 start --nativeinstalls it, and the Kotlin api moves to a free port behind it. - In the desktop app, the update notice says Download the new app instead of a
keel2command.
Good to know
- keel desktop is a preview. Not yet: Windows, Intel Macs, self-update, a tray icon, start at login,
keel2commands for the app's keel. - The AppImage is about 290 MB: keel, Java and Python are inside.
- keel Docker and
keel2work as before. The app keeps its own data, apart fromkeel2's keels.
keel v2 0.26.1
keel v2 0.26.1
Code Review threads, clearer and cheaper. keel's check of a review thread now:
- says who is right by role and name
- checks the same point once
- reads the files it needs
- offers Resolve when a fix really landed
Changed
-
Who is right, by role and name (#106). Both review views use the same words:
- "Reviewer (bo) is right"
- "Author (you) is right"
- "Reviewer (bo) is partly right"
- "Unclear: keel could not decide"
"YOU'RE RIGHT" and "THEY'RE RIGHT" are gone. What keel hands to KeelBot uses the same words.
-
One check per point (#107).
- Two threads about the same problem form one group: same file, lines close together, the same claim. keel checks the group once.
- Each thread says "Same point as the thread on line N (alex)". "Use as my reply on both" fills both reply boxes and posts nothing.
- "Check all open" counts points, not threads. Nothing is merged on GitHub or GitLab.
-
The check reads the files it needs (#108). Before, it read only the thread's own file. Now it also reads:
- the files the comments name
- the tests the author points to ("covered by test_refunds")
- the changed file's own test (
money.py↔test_money.py,Foo.kt↔FooTest.kt)
It reads at most 4 files, within the token cap, and says how many it read. Far fewer answers end in "Unclear".
New
- Fixed → Resolve (#109).
- When the author answers "fixed", "done" or "addressed" and pushes a newer commit, keel checks that thread once against the new code.
- It then shows "Fixed in 1a2b3c4 · Resolve", or "Still open: …" with what is still missing.
- Resolve is one click on the button that was there before. keel never resolves a thread by itself.
Good to know
- The fix check starts by itself when you open the review or send a reply. It is one short model call per thread that qualifies, within the token cap, and it runs once per reply and commit.
- A "newer commit" is found by its date. A rebased commit that keeps an old date is not seen.
- Grouping is by matching words, and you cannot yet split a group keel made by mistake.