Skip to content
MCP server

Keel

MCP server that detects when coding agents tamper with locked tests and tracks their cost

Pricing
Open source
Platforms
CLI, self-hosted, vscode
Version
1.0.0
Listing updated
today

Similar MCP servers

About Keel

Keel is an open-source MCP server for developer tools. It is available on CLI, self-hosted, vscode and API.

Keel is an MCP server that sits alongside a coding agent and gives it tools for long, multi-step work: reloading context at session start, saving progress, logging decisions, and committing checkpoints. Its main concern is verification. The README says the agent runs a project's locked test suite through a verify tool, which reports test results and a separate tampered flag, so a green suite that was rigged by editing the tests still shows up as tampered.

The README describes three layers. The injected CLAUDE.md rules tell the agent to use verify instead of running tests directly. On Claude Code, a PreToolUse hook called keel-guard blocks common direct test runners such as pytest, npm test and go test. Checkpoints are marked certified only when a passing, untampered verify ran against the same working tree, so bypassed milestones appear as uncertified in stats and reports.

Keel is aimed at developers who run agents on long tasks and want to see what the run cost and whether its results can be trusted. Every tool call writes to a local events.jsonl file in the repo, which feeds the stats and milestone reports. The README says the core is MIT licensed, and that a separate paid team dashboard is planned on top of it.

Tools this server provides (12)

Taken from the project documentation; the running server may differ.

init_project
Lays the harness: architecture, constraints, test lock, git, hook, CLAUDE.md.
load_context
Resumption briefing at session start.
save_progress
Writes done, doing and todo state to disk.
record_decision
Append-only log of architectural choices.
checkpoint
Git commit of the step, with certification.
verify
Runs the locked tests and detects tampering.
lock_tests
Re-locks the tests, a human action.
check_drift
Recalls constraints and flags unsaved work.
route_task
Recommends a cost-effective model for a subtask.
log_usage
Records spend of a model call.
stats
Aggregated metrics on cost, cheating, certification and velocity.
report
Writes a markdown milestone report.

What it does

Tamper detection for tests
Verify runs the locked test suite and reports whether any test file changed, separately from whether the tests pass.
Direct test run blocking
On Claude Code, a PreToolUse hook refuses common direct test commands and points the agent to verify.
Certified milestones
Checkpoints are marked certified only when a passing, untampered verify ran on the same working tree.
Session memory and progress
Tools save done, doing and todo state and decision logs so a run can resume after context is lost.
Model cost logging
Records token usage and spend per model call and summarizes cost per model.
Milestone reports
Generates a markdown report of stats, certified milestones and cheating alerts into the .keel folder.

Good for

  • Keeping a coding agent from editing tests to force a green run
  • Reviewing what an agent's long task cost, broken down by model
  • Checking which milestones were verified before accepting agent work
  • Resuming an agent task across sessions without losing decisions or progress

Pricing

Open source

MIT licensed local sensor; a paid team dashboard is planned but not available.

Keel FAQ

Does it work with clients other than Claude Code?

The README says the 12 MCP tools work with any MCP-capable coding agent, including Cursor, Cline and Codex. The keel-guard hook that blocks direct test runs is specific to Claude Code.

What is the license?

The README says this repository is MIT licensed.

Does it send my code to a server?

The README says events are written to a local events.jsonl file in your repo. It also says the planned cloud dashboard would receive metadata only, never code or prompts.

Can an agent bypass the test check?

The README says the hook catches canonical test commands but not every invocation, such as a test run inside a shell script. Any bypass is recorded as an uncertified milestone.

Is Keel free?

Yes. Keel is open source, so you can use and self-host it for free.

What are the best alternatives to Keel?

The closest alternatives to Keel are keelyard, Kestrel, Kodelyth ECC, Tracekit MCP gateway and Prober.

Where this comes from

  • Description: written from the project's README and its MCP Registry entry.

How AIPoweredBot checks servers