agence

Agence Guard — GitHub Marketplace Description

Short description (150 chars max)

Gate every AI agent command through MLS policy enforcement. Classify, log, and block shell commands before they run — fail-closed by design.

Long description

Agence Guard is the command-gating layer of the Agence AI governance framework. It enforces a 4-tier trust model on every shell command an AI agent proposes to execute — before execution.

Trust Tiers

Tier	Gate	Typical Commands
T0	Auto-execute	`git status`, `ls`, `cat`, `git log`
T1	Flag & log	`git add`, unknown read-like commands
T2	Require human approval	`git push`, `git merge`, `terraform plan`
T3	Block	`rm -rf`, `git push --force`, `terraform destroy`

Unknown commands default to T2 (fail-closed). The guard never auto-approves unfamiliar input.

MLS Capability Engine

On top of tier classification, Agence Guard enforces a POSIX-inspired capability model (Bell-LaPadula + Biba). Each agent identity carries a capability set. A command is denied if the agent lacks the required capability — regardless of tier.

Built-in capabilities include:

CAP_EXEC_SHELL — execute shell commands
CAP_MUTATE_GIT — git write operations
CAP_EXEC_INFRA — terraform / docker / k8s
CAP_RED_TEAM — security probing
CAP_PUBLISH — npm publish, gh release

Usage

- name: Gate AI command
  id: guard
  uses: l-agence/agence@v1
  with:
    command: 'git push origin main'
    agent: ci
    fail_on_block: 'true'

- name: Use result
  run: echo "Tier: $"

The Marketplace rollout starts with the reusable GitHub Action first for fast per-repo adoption. The separate GitHub App listing and webhook server can ship afterward for org-level rollout, where one installation can cover many repositories.

For larger organizations, the next onboarding step after the GitHub App is shard bootstrap: giving teams a clean way to initialize separate shards, policies, and knowledge boundaries without manual repo-by-repo setup.

Outputs

Output	Description
`tier`	T0 / T1 / T2 / T3
`action`	allow / flag / escalate / deny
`reason`	Human-readable decision reason
`rule`	Matched policy rule (e.g. `blacklist.linux_shell`)

Custom Policy

Bring your own AIPOLICY.yaml:

- uses: l-agence/agence@v1
  with:
    command: 'terraform apply'
    policy: 'codex/AIPOLICY.yaml'   # relative to your repo root
    fail_on_escalate: 'true'

Why Agence?

Zero databases — state lives in flat files and git
3 runtime dependencies — @modelcontextprotocol/sdk, bun, zod
751 tests with 279 security-specific assertions
9 red-team cycles completed (SEC-008 through SEC-019)
Fail-closed — unknown input is never auto-approved
Marketplace visibility — Action + App listings make adoption easier in both open source and enterprise environments