What it does
ControlKeel describes itself as an agent control plane that sits between coding agents and production. It learns a team's intent rules, review taste and delivery habits from observation, findings and evaluation, and turns them into typed memory, policy checks and proof bundles. Its loop is to capture intent and policy (scope, risk, budget), validate agent output with deterministic checks plus optional advisory review, ask a human to approve only high-impact actions, keep evidence so it survives a switch of agent host, and turn recurring failures into regression evidence. The README argues that a rules file only asks the model to behave, whereas ControlKeel enforces the output. It installs as a CLI through Homebrew, npm or a shell script (a PowerShell script is given for Windows) and attaches to a supported agent host with a controlkeel attach command. Local features include a stdio MCP tool set, scanner validation, findings, reviews, budgets and typed memory, with opt-in cloud sync, org membership, webhooks and self-hosting guidance for teams. The README reports its own benchmark on 12 risky scenarios, where the checks caught all 12 and blocked 9 using no provider tokens, and it stresses that this applies only to that suite. It is built with Phoenix and Ecto on SQLite.
Capabilities
- Deterministic validation of agent changes with recorded findings
- Human approval gates for high-impact actions
- Proof bundles, budgets and typed memory that persist across hosts
- Attaches to supported agent hosts, including OpenCode, Codex and Claude
- CLI plus a local stdio MCP tool set
- Benchmark and eval loop built from recurring failures
When to use it
- Your team uses coding agents and wants risky changes checked before they reach main
- You want the same policy and evidence to follow you across different agent hosts
- You need approvals and budgets tracked for agent work
