Agent Assurance

Declare what your agent may do. Verify it on every edit, every PR, every release.

Your repo promises what an AI agent is allowed to do (agent-assurance.yaml: read the CRM, no external send, a human approves actions). Agent Assurance observes what the configuration actually grants — MCP servers, Claude Code permissions, tool definitions — and fails the change when the promise is broken, pointing at the file and line that broke it. Then it leaves a signed record of what the agent could do, and when.

No dashboard, no backend, no LLM in the verdict, no network calls, nothing executed. Rules you can read; numbers you can reproduce by hand.

pipx install agent-assurance
agent-assurance scan .        # what does this repo let the agent do — and does it match the promise?

▶ 22-second launch video

What it catches

Live: the demo PR stays red on purpose. Real repos, scanned as-is: docs/real-world.md.

Why this and not a scanner

Vulnerability scanners look for poisoned tools and leaked secrets. Permission-diff bots show what changed. Neither answers the governance question: does this agent still do only what we said it does?

Checked 2026-09-17; details and sources in docs/landscape.md. This is the governance layer under a scanner, not a replacement for one.

Not for you if you want prompt-injection or tool-poisoning detection (use a scanner), runtime interception of tool calls (a different product), or a score for agents that have no configuration in a repo (there is nothing to observe).

One promise, three enforcement points

Same engine, same words, everywhere.

Two inputs, two checks

Declared — the promise, written by the team in agent-assurance.yaml: capability classes, systems, data, autonomy. Observed — what the configuration actually grants: MCP configs for Claude Code, Cursor, Gemini CLI, VS Code and Codex; Claude Code permissions (allow = no human in the loop); serialized tool definitions (OpenAI, MCP, Claude, LangChain, CrewAI). Unknown is UNKNOWN, never guessed.

Details, weights and the organisation policy file: docs/how-it-works.md.

Use it

# .github/workflows/agent-assurance.yml
permissions: { contents: read, pull-requests: write }
steps:
  - uses: actions/checkout@v4
  - uses: kunko-ai-labs/agent-assurance@v0.5
    with:
      mode: diff                 # on pull_request; use scan on push
      manifest: agent-assurance.yaml
      card: aa-card.html         # optional: shareable capability card
      attest: write              # optional: in-toto evidence (sign = Sigstore)

CLI (scan, diff, check, attest, validate; formats md, json, sarif, html), pre-commit, MCP server, Claude Code hook and the Python API: docs/integrations.md.

Exit codes are a contract: 0 pass · 1 gate tripped · 2 usage error.

Running third-party code in your CI — what you should check

• Pin by commit SHA, not by tag: uses: kunko-ai-labs/agent-assurance@<sha> # v0.5.0. Tags can move; a SHA cannot. Dependabot keeps the comment and the SHA in step. Our own workflows pin every action the same way.
• What the Action does: pip install of this repository at that SHA, then runs the CLI on your files. It makes no network calls of its own (the only outbound traffic is pip and, if you opt in, upload-sarif / attest to GitHub). It executes nothing from your repo, never starts an MCP server, never reads a secret's value.
• Least privilege: contents: read is enough for check and scan; add pull-requests: write only for the diff comment, security-events: write only for SARIF upload, id-token: write + attestations: write only for signing.
• Verify what you get: every release is signed; gh attestation verify on the artifacts, and the OpenSSF Scorecard of this repo is public.

The capability card

--format html produces a single self-contained file — the agent's nutrition label — for auditors, customers or your manager. Light or dark, follows the viewer.

Rules you can rely on

• Deterministic. No LLM anywhere in the verdict.
• Nothing executed, nothing sent. Config files are read; MCP servers never started; secret values never read; no network.
• Unknown is a result, not a silence. And a guess (an inferred class) never fails a build on its own.
• Honest mapping. OWASP ASI controls are maps; anything borrowed (OWASP APTS, EU AI Act art. 12) is adapted. No conformance claimed.
• Tested contracts. Exit codes, SARIF, Action YAML and every fixture's verdict, on Python 3.10–3.12.

FAQ

Does it read my secrets? No. It looks at the names of environment variables and headers (GITHUB_TOKEN, Authorization) to know a credential is in play; values are never read, logged or hashed.

Does it work without the manifest? Yes: scan without agent-assurance.yaml reports the observed blast radius and a "Sources scanned" table. Add the manifest to turn reach into a promise (AA-002).

What about our private MCP servers? They come out UNKNOWN: scored conservatively, and a promise cannot be called kept over them. Add them to your organisation policy catalogue — one entry with its source — and they are classified like any other.

Is the risk score "AI"? No. It is a table of weights in one file, every point has a reason, and the report shows the sum. Same input, same number, forever.

Can it stop the agent at runtime? No, by design. It checks what the repo says the agent can do — at edit time, in the PR, at release. Runtime enforcement is a different product.

Roadmap: docs/ROADMAP.md · Contributing: CONTRIBUTING.md · Good first issues: label:good first issue

License

Apache-2.0 © 2026 Kunko AI Labs