agf-integrator

What is agf-integrator?

agf-integrator is a Claude Code skill — not an SDK, not a library your agent calls. Point it at an existing AI-agent codebase and it discovers where the codebase exposes agent/tool actions, maps what's already governed against the AAP-Core object model, reports exactly what's missing, and — only after you approve a concrete plan — implements and verifies real AGF wiring.

It never modifies a target repository without an explicitly approved plan, and it never claims more coverage than it can actually trace and verify. That's the whole promise: discover, integrate, and verify — not "make the agent AGF compliant."

Installation

As a Claude Code plugin (recommended) — agf-integrator is a self-hosted plugin marketplace with one entry, itself:

/plugin marketplace add agent-governance-foundation/agf-integrator
/plugin install agf-integrator@agf-integrator

# Clone or symlink into a .claude/skills/ folder:
#   ~/.claude/skills/agf-integrator                (personal, all projects)
#   <project>/.claude/skills/agf-integrator         (one project only)
git clone https://github.com/agent-governance-foundation/agf-integrator \
  ~/.claude/skills/agf-integrator

• Either way it becomes invokable as a /agf-integrator slash command. It's also just a set of markdown instructions — read them directly if you'd rather follow the pipeline by hand.

The 8-step pipeline

Every run walks the same eight steps, in order. Steps 0–4 are read-only. Step 5 is a hard stop for your explicit approval. Only steps 6–7 touch the target repo, and only after you've approved exactly what they'll change.

• 0 — Readiness: checks whether agf-sdk, a runtime URL, an agent identity, and a credential source are available — without ever inventing or writing a secret.
• 1 — Discover: finds agent/tool entrypoints, existing auth logic, side effects, and audit trails.
• 2 — Map: labels what's present against the AAP-Core object model — Actor, Authority, Action, Decision, Receipt, Invalidation.
• 3 — Gap Analysis: reports Present / Partial / Missing per action, per object.
• 4 — Classify: confirms the repo matches a supported integration profile.
• 5 — Plan: proposes exact file:line changes and real agf-sdk calls — stops for explicit approval.
• 6 — Implement: applies the approved plan on a dedicated branch, one file at a time, with a diff shown for each — no commit made.
• 7 — Verify: reports what was actually wired, with evidence — not a claim.

Core principle

“Never claim AGF compliance merely because AGF libraries are installed. Compliance requires identifying governed actions, enforcing authorization at the execution boundary, producing evidence, and passing verification.”

Every outcome uses exactly one of four words — never “compliant”:

• UNSUPPORTED PROFILE — Step 4 found no matching integration profile for this repo.
• NOT INTEGRATED — a profile matched, but nothing has been wired yet.
• PARTIAL — some governed actions have Decision enforced; others don't.
• FULL — every discovered action has Actor, Authority, Decision, Execution-validation, and Receipt all enforced, traced, and tested.

Supported profiles

Five integration profiles today, each self-tested against a synthetic fixture, live-tested against a real local agf-runtime, and validated end-to-end against a real third-party repository.