If you landed here from HN, the pitch is simple: keep one repo-owned source of truth at .agentkit/source.md, project it into AGENTS.md, CLAUDE.md, GEMINI.md, COPILOT.md, AGENT.md, and llms.txt, then score whether the repo is actually ready for agent work.
Keep the canonical instructions in .agentkit/source.md, then project the filenames your tools already read.
Run source --init and project --write to add real files to the repo instead of hiding setup in another hosted UI.
Use quickstart for the first read, score for the repo-owned setup, and gate once you want CI enforcement.
Real installs, real scored examples, and a workflow that extends past the install command.
PyPI stats source: pypistats.org, updated 2026-04-21.
The first visible output is a repo diff. agentkit does not ask you to trust a hosted dashboard before it changes anything useful in the repo itself.
repo/
README.md
pyproject.toml
src/
tests/Normal codebase, but no shared agent instructions. Each tool starts from a different file or from no file at all.
source --init + project --write
repo/
.agentkit/source.md
AGENTS.md
CLAUDE.md
GEMINI.md
COPILOT.md
AGENT.md
llms.txt
README.md
pyproject.toml
src/
tests/One repo-owned source fans out into the filenames Claude Code, Codex, Gemini, Copilot, and other agent surfaces already look for.
The setup becomes a normal repo change that maintainers can diff, review, and revert.
You edit the source once instead of hand-syncing separate agent files every time the repo rules change.
Run agentkit score or agentkit gate after the files exist so the workflow stays inspectable.
There are two honest entry points: use quickstart if you want a score before editing files, or use the canonical-source flow if you already want repo-owned agent instructions in git.
Best for skeptical first runs. You learn where the repo stands before adopting any workflow changes.
pip install agentkit-cli
agentkit quickstart .quickstart runs doctor plus a fast composite score on the current repo.Best when you already know the repo needs one shared source of instructions across multiple agent surfaces.
agentkit source --init
agentkit project --write
agentkit scoresource --init creates .agentkit/source.md as the file you actually maintain.project --write writes the projected agent files next to the code.score tells you if the repo is ready for gates and repeatable agent work.The front door is a repo workflow, not just a leaderboard. Start with a canonical source, project it into agent-specific files, and put measurable guardrails around the whole loop.
Initialize .agentkit/source.md so your team edits one file instead of manually keeping AGENTS.md, CLAUDE.md, and friends in sync.
Generate the filenames existing tools expect, so Claude Code, Codex, Gemini, Copilot, and MCP flows inherit the same context.
Run score, fail CI with gate, benchmark execution quality, and inspect transcripts so the workflow gets better over time.
This is most useful once agent work is a repo problem, not just a personal prompt problem. The value starts when the instructions need to live in git and stay consistent across tools or teammates.
You already have Claude Code, Codex, Gemini, Copilot, or MCP tools touching the same codebase and you do not want their instructions drifting apart.
You want the setup to be a normal change in git, with files maintainers can review, diff, and enforce in CI.
You want to score real repos, inspect public examples, and measure readiness before telling a team to standardize on a workflow.
The value here is not just the thesis. There is a shipped toolchain behind it, public scored examples, and enough surface area to show the workflow has been exercised.
102 released versions and about 51.1k PyPI downloads show this is a real shipped toolchain, not a one-weekend demo.
20 public repositories are already scored on this site across 4 ecosystems, so visitors can inspect what the workflow measures.
The toolkit already spans quickstart, source management, projection, scoring, linting, benchmarking, transcript review, and MCP access across 6 packages, with 5.1k downloads in the last 30 days alone.
agentkit-cli is the entry point, but the product story is the full workflow around source management, projection, scoring, linting, benchmarking, and transcript learning.
The orchestration entry point for quickstart, source, project, score, gate, and guard flows.
Generate or tighten agent-facing context files from the canonical source before they drift.
Catch stale paths, year-rot, missing references, and bloated instructions before they mislead agents.
Benchmark AI coding agents on your own tasks instead of trusting generic benchmark marketing.
Read transcript history to see where agent runs burned budget, stalled, or repeated avoidable mistakes.
Expose the toolkit through MCP when you want the same workflow available inside other agent surfaces.
Start here if you want evidence before install. These scorecards show what the toolkit measures against real repositories, not toy demos.
Loading...
If you only try one path first, use the setup flow. The rest of the commands help once you want analysis, sharing, transcript review, or CI enforcement.
| Command | Description |
|---|---|
| agentkit quickstart . | Run the fastest honest first pass: doctor plus a fast composite score on the current repo |
| agentkit source --init | Create the dedicated canonical source file at .agentkit/source.md |
| agentkit project --write | Write AGENTS.md, CLAUDE.md, GEMINI.md, COPILOT.md, AGENT.md, and llms.txt from that source |
| agentkit score | Compute the repo's composite agent-readiness score |
| agentkit analyze github:owner/repo --share | Analyze a public repo and publish a shareable scorecard |
| agentkit burn --path ./transcripts | Inspect local coding-agent transcript spend and waste patterns |
| agentkit gate --min-score 80 | Fail CI when the setup drops below your quality bar |
Browse ecosystem scoreboards and repo pages first. It is the fastest way to understand what agent-readiness means in practice.
Rank all public repos in a GitHub org to spot where agent workflows are already strong and where they need cleanup.
agentkit org github:ORGGenerate a public profile card when you want a simple external artifact for a person or repository set.
agentkit user-scorecard github:USERNAMEA sample of currently scored repositories. The rankings are still useful, but they now support the broader workflow story instead of pretending to be the whole product.
| Repository | Score | Grade |
|---|---|---|
| langchain-ai/langchain | 50 | C |
| fastapi/fastapi | 50 | C |
| django/django | 50 | C |
| psf/requests | 50 | C |
| pallets/flask | 50 | C |
| vercel/next.js | 50 | C |
| microsoft/vscode | 50 | C |
| nestjs/nest | 50 | C |
| supabase/supabase | 50 | C |
| trpc/trpc | 50 | C |
| rust-lang/rust | 50 | C |
| tokio-rs/tokio | 50 | C |
| actix/actix-web | 50 | C |
| hyperium/hyper | 50 | C |
| serde-rs/serde | 45 | D |
| golang/go | 50 | C |
| gin-gonic/gin | 50 | C |
| gofiber/fiber | 50 | C |
| spf13/cobra | 50 | C |
| kubernetes/kubernetes | 50 | C |