3 agents working
The open-source control plane for autonomous coding agents. Budgets, audit, approval and the kill switch live outside the runner, so they hold across every agent, cloud or local.
| agent/issue-1083 | sonnet | ▒▒▒▒▒▒▒▒▒▒ | step 0/12 | $0.00 | running |
| agent/issue-1068 | deepseek | ▒▒▒▒▒▒▒▒▒▒ | step 0/12 | $0.00 | running |
| agent/issue-1059 | qwen3-30b | ▒▒▒▒▒▒▒▒▒▒ | step 0/12 | $0.00 | running |
How it works
kew runs the loop from a scoped issue to a merged pull request. The machine does the work; the decisions that cost money or change main stay with you. Scroll the stages and the console follows. Everything in it is live; touch anything.
Stage 01
You
Nothing dispatches until a person says it is well-formed. Label an issue ready and it enters the queue. That label is the whole contract.
Try it: label #1101 ready and watch a p0 take the front of the queue.
The queue is ordered by priority label (p0-critical through p3-low), then by complexity (simple / medium / complex) which sets the step budget.
An issue carrying blocked stalls the driver rather than dispatching. Dependencies are declared by kew itself, never by the agent.
Stage 02
kew
Every agent gets its own git worktree and its own branch. Nothing touches your working tree, and two agents can never collide in the same checkout.
Worktree per issue, branch agent/issue-N, and optionally a container with an egress allowlist. A stale branch with no unique commits is auto-cleaned on re-dispatch.
Budget caps, the audit log and the kill switch live outside the runner, so they hold identically whether the agent is Claude Code, OpenCode, pi or a local model.
Stage 03
kew
Spend is metered per issue, per day and per week. Cross a cap and the circuit breaker trips. The fleet halts rather than quietly burning your budget.
Try it: drag the cap below the spend.
The breaker halts dispatch; it never kills mid-flight work silently. Issues estimated at or above the approval threshold are held with needs-approval instead of running.
Cost is API-equivalent list price, so a local model shows $0.00 and still reports the work it did.
Stage 04
kew
The reviewer is a fresh context with no memory of writing the code, so it has nothing to defend. What it reviews is bound to what merges.
The reviewed commit and the merged commit are asserted equal. Without that, a review is a statement about code that no longer exists.
Findings come back as review comments; kew review <pr> dispatches an agent to address them, and the verdict is re-run against the new head.
Stage 05
You
What arrives is a pull request, not a merge. Small enough to actually read. The gate is the point. This is the step that separates offloading from surrender.
Try it: approve or deny.
From the TUI, the web inbox, or kew approve. The fleet kill switch is always live. It sits in the console header on this page too, at every stage.
Denial returns the issue to the queue with the reason attached, so the next dispatch starts from your objection rather than from scratch.
Stage 06
kew
Each event is hash-chained to the one before it. Change any record after the fact and every link downstream stops verifying.
Try it: tamper with an event. These are real SHA-256 chains, computed in your browser.
Dispatches, approvals, denials, spend, policy decisions and merges. Anchors are published periodically so the chain cannot be silently rebuilt from scratch.
kew audit verify --anchors checks the whole chain and exits non-zero on a break. An evidence pack exports it as one self-contained file.
queue · ordered by priority
backlog · not dispatchable
Runners
Governance that lives inside one vendor’s console governs one vendor. kew keeps budgets, audit, approval and the kill switch outside the runner, so they hold across every agent, cloud or local.
kew
drives runners claude-code
which reach
The default. Deepest tool use and the most reliable long-horizon runs.
| executes | the claude CLI |
| providers | Anthropic |
| credential | ANTHROPIC_API_KEY or a Claude Code OAuth token |
| cost | reported per turn by the CLI |
| routes | claude-* |
unchanged, whichever you pick
| worktree | isolated per issue |
| budget | per-issue, daily, weekly caps |
| breaker | halts the fleet on breach |
| approval | held above threshold |
| kill switch | kew stop, always live |
| audit | hash-chained + anchored |
What that actually cost, on kew’s own repo, 1 Jul 2026 to 26 Jul 2026
| Model | Runner | Runs | Merged | Merge rate | Spend | Per merged PR |
|---|---|---|---|---|---|---|
| claude-opus-4-8 | claude-code | 299 | 159 | 100% | $1128.32 | $7.10 |
| claude-sonnet-4-6 | claude-code | 54 | 52 | 100% | $110.89 | $2.13 |
| Qwen3-Coder-30B (local) | lemonade | 11 | 7 | 88% | $0.00 | $0.00 |
A frontier model merged 159 pull requests at $7.10 each. Sonnet merged 52 at $2.13. A local 30B merged 7 at nothing at all, and opened fewer of the runs it started. That is the honest trade, and all three ran under the same caps, the same approval gate and the same chain.
1853 events · chain verified ·
13/13 anchors matched ·
exported by kew report export
Integrations
kew does not ask you to move it. The tracker is a port with two implementations, so an epic is a label convention in one and a native Project in the other, and the driver above them cannot tell the difference.
complete
| work items | issues, labelled ready to enter the queue |
| priority | p0-critical through p3-low |
| sizing | simple, medium, complex sets the step budget |
| output | a pull request on agent/issue-N |
| review | PR review comments, addressed by kew review |
| ci | Actions runs, repaired by kew ci-fix |
read surface complete, writes are a deliberate slice
| work items | issues in the project |
| statuses | DONE, BLOCKED, IN_REVIEW, TODO, provider-derived only |
| archived | included; a completed archived dependency is satisfied |
| cancelled | maps to BLOCKED, so it can never read as satisfied |
| delegation | assign to kew, then kew delegate watch claims it |
| writes today | list ready, claim, comment, close |
True on either, because neither owns the control plane
Linear’s read surface is complete: projects, native relations, statuses and webhooks all drive the graph. Its write vertical is intentionally narrow for now, and it is marked that way above rather than rounded up to parity.
Surfaces
Four ways into the same control plane, not four skins on one screen. Each matches a different mode of attention, and the state behind them is the same state.
$ kew status
Thirty commands, every one scriptable. The whole control plane is reachable without a UI, which is what makes it usable from CI and from a machine you have only SSH’d into.
reach for it when You are already in the terminal and want an answer without changing context.
docs →Command index
Everything the dashboard and the web inbox can do, the terminal can do, because they are the same control plane. All of it scripts, which is what makes kew usable from CI and from a machine you have only SSH’d into.
press / to filter
Dispatchput work in flight19
Observesee what is happening8
Governkeep the decisions12
Operaterun the control plane19
Built with itself
kew is not demonstrated on a toy project. It is the tool that built itself, under its own budgets, its own approval gates and its own audit chain. Every figure below comes from the repository you are being asked to install.
thossullivan/kew · captured 2026-07-27 · agent PRs identified by kew’s agent/issue-N branch convention
One of those epics
Eight issues, nine dependencies, resolved by kew rather than by a person holding the order in their head. Watch it run. It will stop on its own when a node costs enough to need you.
Receipts
Every engagement ends with one self-contained file: what ran, what it cost, what merged, and a cryptographic attestation that the record was not edited afterwards. Below is a real one, covering the 25 days to 26 Jul 2026.
Where the money went
| Run kind | Runs | Success | Cost |
|---|---|---|---|
| dispatch | 304 | 86% | $1103.66 |
| review | 59 | 86% | $151.59 |
| adversarial_review | 62 | 74% | $91.96 |
| review_remediation | 10 | 100% | $50.89 |
| planning | 12 | 100% | $14.57 |
Which model earned it
| Model | Runs | Merged | Per PR |
|---|---|---|---|
| claude-opus-4-8 | 299 | 159 | $7.10 |
| claude-opus-5 | 43 | 1 | $169.51 |
| claude-sonnet-4-6 | 54 | 52 | $2.13 |
| claude-opus-4-6 | 2 | 0 | n/a |
| deepseek-v4-flash | 3 | 1 | $0.17 |
| Qwen3-Coder-30B (local) | 11 | 7 | $0.00 |
| unknown | 35 | 0 | n/a |
35 runs carry no recorded model. They are in the totals rather than dropped from them, so the cost per merged PR is not flattered by hiding them.
Verify it here
These are kew’s real anchor records. Your browser just recomputed each one with the
same algorithm kew uses, and checked that every prev binds the entire previous
record. Alter a single value and watch the chain break.
Anchors are committed to git daily, out of band. Rewriting the audit log would mean rewriting that history too, which is what makes a silent re-chain detectable rather than merely unlikely.
What you can check from this page
What you cannot
kew’s record is tamper-evident for whoever holds the database, not verifiable by a stranger from a summary. That distinction is the whole point of publishing the algorithm and the anchors rather than only the verdict.
Sixty seconds
Issue to merged pull request, with the cost showing the whole way.
Prove it
Ten epics, sixty merged PRs, every decision hash-chained and verifiable.
Read it
Apache 2.0. The governance is the product, so none of it is hidden.
pip install kew-cli Python 3.12+ · requires git, gh and at least one runner
Licence
kew is released under the Apache License 2.0. You may use it commercially, modify it, distribute it and run it privately, with a patent grant and no obligation to publish your changes.
Budgets, the audit chain, approval gates and the kill switch are in the licensed source. They are the product, not an upsell. Governance you cannot inspect is not governance.