ClawKeeper
Governed Claude agent control plane. A deterministic policy engine runs before every LLM call, autonomous spend defaults to zero, and the audit log is enforced by the database.
Proof ledger
- Classification
- Built artifact
- Repository
- Alexi5000/ClawKeeper
- Language
- TypeScript
- Licence
- MIT
- Status
- Active development
- Tests
- 3 recorded gates
- Last verified
- 2026-08-06
- Related write-up
- alexcinovoj.com
What it is
ClawKeeper is a control plane that sits between an operator and a fleet of Claude agents doing finance work. Every proposed action passes a deterministic policy evaluation before any model call is made, so a denied action never reaches a provider. Spend is opt-in rather than opt-out: autonomous budget is zero until an operator raises it. The audit log is append-only at the database layer, not at the application layer.
Architecture
- Policy engine evaluates a proposed tool call against deterministic rules before any provider request is issued.
- Agent clusters are grouped by responsibility: intake, validation, reconciliation, reporting, compliance.
- Postgres with row-level security carries tenant isolation; the audit table is append-only and enforced in the database.
- Bun and Hono on the service side, React on the dashboard side, Docker Compose for local bring-up.
- Skills and steering documents (SKILLS.md, STEERING.md, AGENTS.md) are checked into the repository so agent behaviour is reviewable in diffs.
operator request
|
v
+-----------------+ deny
| policy engine |----------------> audit log (append-only)
| (deterministic) |
+--------+--------+
| allow
v
+-----------------+ +----------------------+
| budget ceiling |---->| provider call |
| (default $0) | | (Claude) |
+-----------------+ +----------+-----------+
|
v
audit log (append-only)Tests and gates
- The README states 147 passing tests; the suite runs in the repository's GitHub Actions workflow (.github/workflows/ci.yml).
- Lint and type checks are wired into the same workflow via eslint.config.js and the package scripts.
- CODEOWNERS, CONTRIBUTING.md, and SECURITY.md are present, so review and disclosure paths are defined.
Limits
- This is an open-source control plane, not a hosted product. There is no managed deployment behind it.
- The policy engine is deterministic by design: it does not reason about intent, so policies must be written explicitly.
- Finance workflows in the repository are reference implementations, not certified accounting logic.
Source artifacts
Patterns applied
Acceptance tests for agent work →
Define done before the agent starts.
Evals that gate releases →
A golden suite in CI, not a spreadsheet.
MCP permission scoping →
Deny by default, declare every tool.
Observability for agent runs →
One trace ID from request to receipt.
Cost controls that actually stop spend →
Ceilings enforced before the call, not after the invoice.
Approval gates that suspend the run →
Nothing downstream executes before a human approves.
Related proof
TechTide Harness Kit →
Trust contracts for agent skills expressed as JSON Schema: risk tiers, tool permissions, deny-by-default egress, approval gates, and prompt-injection fixtures behind a wall of CI validation gates.
CipherClaw →
Deterministic multi-agent debugger with zero LLM calls in the analysis path: causal graphs, weighted root-cause scoring, and failure prediction.
Swarm 357 →
357 Claude agent roles across six business layers, with portable memory, enforced cost caps, a bash policy gate, and durable checkpoints.
FintheFinder →
Deep-research agent whose runs suspend until a human approves, with a deterministic citation audit that keeps AI out of the verification path.
This page is implementation proof only. Engagement scope and pricing live on TechTide AI.
AI Production Systems at TechTide AI→© 2026 Alex Cinovoj · TechTide AI · Columbus, OH