fleet-audit

Quick start · Website · How it works · AGENTS.md · llms.txt · Skills · Discussions · All projects

Scan your repositories for every surface coding agents need, and fix them.

Fleet Audit is an agent-native tool: other agents call it through its skills, its MCP server or its command line. Its persona is Muster. A scheduled job runs its evals and proposes fixes. A person makes every change.

Fleet Audit is part of an agent-native play for agent-native systems that can eventually be recursively self-improving and evolving, operated by agent employees.

Find out what a coding agent will find in your repository, and what it will miss. fleet-audit checks a repository against one standard for agent-native projects, scrubs it before it goes public, and writes the standard files it is missing. It is for anyone who publishes repositories that agents are meant to install, read and extend.

Use it from any coding agent

Claude Code: the plugin brings the skills, sub-agents, hooks and MCP server.

/plugin marketplace add naman10parikh/fleet-audit
/plugin install fleet-audit@fleet-audit

Its MCP server, Codex, Cursor, Gemini CLI, one skill and the command line

Its MCP server, in any MCP client:

claude mcp add --transport stdio fleet-audit -- uvx --from git+https://github.com/naman10parikh/fleet-audit fleet-audit-mcp
codex mcp add fleet-audit -- uvx --from git+https://github.com/naman10parikh/fleet-audit fleet-audit-mcp

Cursor (.cursor/mcp.json), Gemini CLI (.gemini/settings.json) and most other clients take the same command as JSON:

{
  "mcpServers": {
    "fleet-audit": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/naman10parikh/fleet-audit",
        "fleet-audit-mcp"
      ]
    }
  }
}

Gemini CLI: the extension brings the context and the MCP server.

gemini extensions install https://github.com/naman10parikh/fleet-audit

One skill (3 in all, in skills/), in any agent that reads Agent Skills:

gemini skills install https://github.com/naman10parikh/fleet-audit --path skills/audit-a-repo
git clone --depth 1 https://github.com/naman10parikh/fleet-audit /tmp/fleet-audit && cp -r /tmp/fleet-audit/skills/audit-a-repo ~/.claude/skills/

The command line:

uvx --from git+https://github.com/naman10parikh/fleet-audit fleet-audit --help

Codex, Cursor, OpenCode, GitHub Copilot and other agents read AGENTS.md; Claude Code reads CLAUDE.md and Gemini CLI reads GEMINI.md, which import it. llms.txt maps every surface for agents that read first.

Quick start

No clone needed: uv builds fleet-audit from GitHub on first use. Run this inside any git repository:

uvx --from git+https://github.com/naman10parikh/fleet-audit fleet-audit check . --offline

It prints one line per check (pass; MISSING, with the fix; or unknown) and ends with a tally. Run on its own repository it ends with fleet-audit: 22 pass, 0 missing, 4 unknown: offline, the description check needs gh, the scrub check needs a patterns file, and two checks need a facts file.

Before a repository goes public, scrub it with your own patterns, kept outside every repository (scrub-patterns.example.txt shows the format):

fleet-audit scrub . --patterns ~/scrub-patterns.txt

Write what is missing from a facts file (facts.example.json shows the keys), with a dry run first:

fleet-audit kit . --facts facts.json --owner <github-owner> --author "<your name>" --dry-run

In a clone, the tests and the evals need only Python and git:

python3 -m unittest && python3 evals/run.py

The evals end with 18 passed, 0 failed.

What it checks

Group	Checks
Context for every agent	AGENTS.md · llms.txt · CLAUDE.md · GEMINI.md · a README with a quick start
One-step surfaces	3 or more skills · 2 or more sub-agents · a hook · an MCP server · a CLI · plugin.json · a marketplace · a Gemini CLI extension · an A2A agent card · it runs from GitHub
It improves itself	dsip.json · CI · a scheduled loop · memory (MEMORY.md and LEARNINGS.md) · no absent component
A mature project	a licence · SECURITY.md · CONTRIBUTING.md · a description and topics · a clean scrub

A tool, a component, an employee or a product gets all 26 checks. A list gets the 8 that fit it (the context files, the README, the hygiene files, the description and the scrub).

How it works

Committed state, not memory. The scan reads a git ref: the remote's default branch, or the local HEAD with --head. Uncommitted files do not count, because a stranger will never see them.

Pass, missing or unknown. A check that needs something fleet-audit does not have (a facts file, gh, a scrub) says unknown. It is never quietly counted as a pass, and never as a failure.

A scrub that keeps its findings to itself. Your patterns run over the tree with git grep -P, over the whole history with git log -G (translated for the BSD regex library on macOS) and over every commit message; gitleaks covers secrets. Every result names a pattern by its number and never repeats the matched text.

Additive fixes. The kit creates only missing files and replaces only its own marked README blocks. It never deletes a file and never commits. In the MCP server, the kit is a dry run until it is called with write: true.

History in a scratch mirror. fleet-audit history clones a bare mirror, drops the internal notes your facts file names, replaces every pattern, and scrubs the result. Pushing that history is left to a person.

The hook. In Claude Code, a pre-push hook scrubs the committed tree with your patterns and blocks the push on a match.

The loop. Every night a scheduled job runs the evals, audits and scrubs this repository, and writes proposals to one issue. It never commits; a person makes every change.

The commands (each answers --help):

fleet-audit check [path] [--facts FILE] [--kind KIND] [--patterns FILE] [--head] [--offline] [--json]
fleet-audit scan [path ...] [--owner OWNER] [--clones DIR] [--out FILE] [--head] [--offline]
fleet-audit report path ... [--facts-dir DIR] [--kind KIND] [--out FILE]
fleet-audit scrub path ... --patterns FILE [--head] [--tree] [--no-leaks] [--out FILE]
fleet-audit kit [path] --facts FILE --owner OWNER --author NAME [--dry-run]
fleet-audit history source ... --patterns FILE [--facts-dir DIR] [--out DIR]
fleet-audit route "TASK"
fleet-audit runs [--last N]

Limits

• The checks find files and read their shape. They cannot tell a thoughtful AGENTS.md from a thin one; the readiness-auditor sub-agent opens the files to judge that.
• GitHub's description and topics need gh signed in. Without it, that check is unknown.
• A scrub is only as good as its patterns. Test each new pattern on a known hit. gitleaks and git-filter-repo are optional, and fleet-audit says when one is missing.
• The kit writes an MIT licence when a repository has none. Add your own first if you want another.

What it is made of

Every project in this fleet is graded on the same 11 harness components and on self-improvement. Each mark is backed by a file you can open.

Component	Grade	Where
Identity	✅	identity/SOUL.md:7 + identity/BRAND.md:5
Memory	✅	memory/MEMORY.md:1 + scripts/memory_search.py:49
Skills	✅	skills/audit-a-repo/SKILL.md:1
Tools/CLIs	✅	pyproject.toml:25 + src/fleet_audit/cli.py:34
Hooks	✅	.claude/settings.json:4 + hooks/hooks.json:4
Sub-agents	✅	.claude/agents/readiness-auditor.md:1
MCPs	✅	src/fleet_audit/mcp_server.py:35 + pyproject.toml:25
Dispatch	✅	src/fleet_audit/route.py:37 + .github/workflows/loop.yml:5
Evals	✅	evals/run.py:1 + .github/workflows/ci.yml:24
Observability	✅	src/fleet_audit/runlog.py:27
Sandbox/Infra	✅	sandbox/run_in_sandbox.py:44
Self-improvement	✅	dsip.json:1 + .github/workflows/loop.yml:41

✅ real and tested · 🟡 partial, with the reason in the row · ❌ absent

How it improves itself

It carries its own evals and a self-improvement contract, dsip.json, which says what it may change and how a change is judged. A scheduled job runs its evals and proposes fixes. A person makes every change. Lessons it learns are kept as rules in its memory files.

Part of a fleet

One of the projects by Naman Parikh. Each owns one part of an agent's harness, or of the product around an agent, and all of them follow one standard.

Related project	What it owns
AgentDial	Identity
WikiMem	Memory
Skillsmith	Skills
Agent Booster	Tools / CLIs
Hooksmith	Hooks
AgentCast	Sub-agents
Plugsmith	Plugins
Armory	MCPs and the component registry
AgentGrid	Dispatch
AgentSwarm	Dispatch
AgentBench	Evals
AgentMoney	Observability
Sandforge	Sandbox / Infra
Claude Harness	The whole harness, as a template

Community

Questions and ideas go to Discussions; bugs go to issues. Follow the work on X and LinkedIn.

Contributing and security

Read CONTRIBUTING.md. AI-assisted pull requests are welcome. Report a vulnerability privately, as SECURITY.md explains.

License

MIT, © 2026 Naman Parikh. See LICENSE and NOTICE.

Star history

If it is useful to you, a star helps other people find it.

Contributors