All kits / Plugin
keel
Plan with a critique gate, build in phases that prove themselves, give each task to the cheapest capable model, see what it cost.
What you'll see
Keel · ≈$0.03 this prompt (sonnet 5.5) · session ≈$0.25 · 95.8k in, 1.5k out · est.Added the login route and its test.
### For you
1. **Add the secret to Vercel.** Run `vercel env add API_KEY`
→ expect: Added Environment Variable API_KEY (blocks the next step)
### Task list
[x] Add the login route
[>] Wire the form to it
[ ] Deploy and smoke testPhases
[x] Phase 0 — MVP
[>] Phase 1 — Core
[ ] Phase 2 — Polish
Now: Phase 1 — Core
Goal: ship the core
Checks to pass:
- `npm test` → green
- login works
Last progress entry
2026-11-14 — Phase 0 passed
Cost this session: ≈$0.42 (1.2M in, 9.1k out), an estimate from list prices.
Keel v0.6.1. Check for updates: https://praxikit.github.io/praxikit/updates/?kit=keel&v=0.6.1#updatesKeel done-check: 2 files changed this prompt and nothing was run to check them afterwards. Run the phase checks or the tests before calling it done.Skills 9
- skill/keel:cost
Show the full token and cost table per model for the last prompt and the whole session, including sub-agents.
- skill/keel:critique
Run the Keel critique gate alone on a plan, an idea, a feature or a PR — is it needed, does it exist, the simplest version, over-engineering, what breaks, cost by tier, attack surface, the counter-...
- skill/keel:delegate
Hand a scoped task to the right Keel agent tier — Scout (Haiku) to read and search, Builder (Sonnet) to implement, Reviewer (Sonnet) to check a diff — with a full brief (goal, files, steps, checks,...
- skill/keel:plan
Start or restart a project plan the Keel way — brief, critique gate, decisions, phases with acceptance checks, security notes — written to docs/plan/ at the project root.
- skill/keel:replan
The idea or the plan changed — take the change back through the Keel critique gate, edit decisions and phases, log the turn, and keep the docs truthful.
- skill/keel:security
Run Keel's security track on a project — wire free scanners into CI, fill the entry-point table from the real code, walk the pre-launch checklist ticking only what has evidence, write the OWASP Top...
- skill/keel:settings
View or change Keel's defaults for this project — the For you section, the Task list section, the cost footer, which models tasks may be distributed among, and which model each tier uses.
- skill/keel:status
One-screen project status for a Keel plan — which phases are done, the current phase and its checks, open decisions, the last progress entry and the cost so far.
- skill/keel:verify
Run a phase's acceptance checks from docs/plan/03-phases.md, paste the real outputs into 05-progress.md, and mark the phase passed, in progress or blocked.
Agents 4
- agentbuilder
The Keel builder. Implements one fully briefed task — goal, exact files, steps, acceptance checks, do-nots — inside its phase, runs the checks, and reports what changed with the check outputs.
- agentcritic
The Keel critic. Argues against a plan, a feature, a change or a PR before it is built — need, prior art, the simplest version, over-engineering, failure modes, cost by tier, attack surface, the co...
- agentreviewer
The Keel reviewer. A second read of a diff or a phase's changes for correctness, security (auth, secrets, uploads, queries, network) and drift outside the phase's file list.
- agentscout
The Keel scout. Finds things and measures things — files, usages, docs, free-tier limits, a number from a page — and returns the answer, not the dump.
What it does on your machine
Runs automatically
SessionStartrunssession-start.mjsPreToolUseonAgent|Taskrunsrequire-model.mjsPostToolUseonEdit|Write|MultiEdit|NotebookEditrunsdrift.mjsStoprunsdone-check.mjsStoprunsfooter.mjs
Listed from the kit's hooks.json, so it matches what is installed.
Reads
- docs/plan/ in your project, for the drift warning and /keel:status
- .keel/config.json in your project, for your settings
- This session's transcript and its sub-agent transcripts under ~/.claude/projects, for token counts only
Writes
- Nothing on its own. Skills write docs/plan/ and .keel/ files only when you run them
- A one-line marker file in your OS temp folder so the drift warning shows once per session
Network
None from the hooks or scripts (a test checks the scripts for network code). The scout and critic agents can search the web when you use them.
Reads, writes and network are declared by the kit's author. Read the source before you rely on them.
Build checks
- Passed: Manifest has every required field
- Passed: Changelog has an entry for v0.6.1
- Passed: Plugin manifest version matches the kit version
- Passed: 9 scripts scanned: no network or child-process code
- Passed: MIT licence file at the repository root
- Passed: 3 test files in the kit; CI runs them on every push
Computed from the files at commit 88be77c, CI results for it, built 2026-11-16. These show the kit is consistent. They are not a security audit and do not mean the kit is safe.
Honest limits
- Not a pentest and not a model classifier.
- Model routing works in Claude Code only; the method and prompts work with any agent.
- The security skill wires free scanners and an evidenced checklist; it is not a pentest and does not prove a project is secure.
- Cost figures are estimates from list prices, not a bill; subscription users do not pay them. Sub-agent usage is read from Claude Code transcript files, whose format is not a public contract.
Changelog
0.6.1 — 2026-11-15
Latest- Changed: the update link
/keel:statusprints now points at the catalog's Updates page (/updates/?kit=keel&v=<version>), which has its own address since the catalog was split into pages. Links printed by 0.6.0 still work; they are redirected.
- Changed: the update link
0.6.0 — 2026-11-15
- New:
/keel:statusends with the installed version and a link that fills in the catalog's "Am I up to date?" form for it (https://praxikit.github.io/praxikit/?kit=keel&v=<version>#updates). A web page cannot read your installed plugins, so the plugin hands the page your version instead.
- New:
0.5.1 — 2026-11-15
- Changed: the manifest now carries examples of what you will see, what the kit reads, writes and sends, and install steps for Cursor and Codex, so the catalog page can show them. No change to how the plugin behaves.
- New: a test checks that Keel's scripts contain no network or child-process code, because the catalog page says they make no network calls.
0.5.0 — 2026-11-15
- New: budget guard. Set
budget.promptand/orbudget.session(estimated USD) in.keel/config.json; when a cap is passed the cost footer ends withOVER prompt cap $X. Warns only, never stops work. No cap by default. - New: cost per phase.
/keel:verifyrunsscripts/phase-cost.mjswhen a phase passes and records its estimated cost in05-progress.md, using a ledger in.keel/ledger.json(difference between marks, all sessions and sub-agents). The first mark covers all recorded history and says so.
- New: budget guard. Set
0.4.0 — 2026-11-15
- New:
/keel:status. One screen with the phases and where you are, the current phase's goal and checks, open decisions, the last progress entry and the session cost. Read-only. - New: done-check. When a prompt changed files (other than docs) and no check-like command ran afterwards, a Stop hook prints one line saying so. A heuristic, never blocks. Turn off with
defaults.doneCheck: false.
- New:
0.3.1 — 2026-11-15
- Changed: a fixed reply layout. The result first, then
### For you(bold action, exact command,→ expect:,(blocks the next step)), then### Task list. Task order and wording stay stable across replies and finished tasks stay on the list until the request is done. A plain question still gets a plain answer. - Changed: task marks default to ASCII
[x] [>] [ ], because the emoji marks drew as empty boxes in some terminal fonts. Setdefaults.marksto"emoji"in.keel/config.jsonto get ✅ 🔄 ⬜. - Changed: the cost footer is one short line (
Keel · ≈$0.03 this prompt (sonnet 5.5) · session ≈$0.25 · 95.8k in, 1.5k out · est.) so it no longer wraps.
- Changed: a fixed reply layout. The result first, then
0.3.0 — 2026-11-15
- New: a standing response format, injected once per session. Replies end with **For you** (numbered, detailed steps for what only the user can do) and **Task list** (✅ done, 🔄 in progress, ⬜ next) when there is something to show. Left out when empty.
- New: a one-line token and cost footer after every prompt (Stop hook): tokens in/cached/out, cost per model, session total, labelled as an estimate. Includes sub-agent usage.
/keel:costprints the full per-model table. - New:
prices.json(list prices with source and date, including Haiku 5.5's tier above 100k prompt tokens). Models with no price show tokens only. Override withpricesin.keel/config.json. - New:
/keel:settingsand.keel/config.json: switch the sections and footer on or off, choose which models tasks may be distributed among, and which model each tier uses./keel:delegatehonours it. - Changed: the per-project tier override moved from
.keel/tiers.jsontotiersin.keel/config.json.
0.2.0 — 2026-11-14
- New: drift hook. After an Edit or Write, warns once per session when the file is outside the current (🟨) phase's Files list in
docs/plan/03-phases.md. Never blocks. - New:
/keel:security. Wires gitleaks, Semgrep and dependency audit into CI (templates/ci/security.yml), fills the entry-point table from code, ticks the pre-launch checklist only with evidence, writes the OWASP review and a pentest recommendation. - New:
tiers.jsonis the single place for model tiers; CI checks the agents match it. A project can override with.keel/tiers.json, and/keel:delegatepasses the model explicitly.
- New: drift hook. After an Edit or Write, warns once per session when the file is outside the current (🟨) phase's Files list in
0.1.1 — 2026-11-14
- Fix: skills read the method and templates from outside the installed plugin folder, so they would have failed after install.
method/andtemplates/now ship inside the plugin. - Moved into the praxikit monorepo (
kits/keel/).
- Fix: skills read the method and templates from outside the installed plugin folder, so they would have failed after install.
0.1.0 — 2026-11-07
- First release: five skills (plan, critique, delegate, verify, replan), four agents, the model-required hook, Cursor and Codex adapters.