claude-anvil
Evidence-first coding agent for Claude Code CLI. Port of burkeholland/anvil from GitHub Copilot CLI.
Your coding agent should prove its work. This one does.
/anvil verifies every change before you see it — builds, tests, lints, runs IDE diagnostics, then has other AI models (Claude, GPT, Gemini, Ollama) try to break it. Every check is INSERTed into a SQLite ledger; the final Evidence Bundle is a SELECT, not prose. If the INSERT didn't happen, the verification didn't happen.
What you get
- Adversarial review. Every change gets attacked by 1 (Medium task) or 3 (Large / 🔴 task) different models in parallel. Claude runs as a Task subagent; GPT / Gemini / Ollama run via Bash scripts. They disagree with each other. That's the point.
- SQL verification ledger. Every build, test, lint, diagnostics, and reviewer verdict lives in
anvil_checks. The Evidence Bundle is SELECT ... ORDER BY phase — the agent can't fake it.
- Baseline snapshots. Before touching anything, anvil records your project's state. After, it diffs. Regressions are caught before you see the code.
- Pushback. Anvil is a senior engineer, not an order taker. If your request introduces tech debt, duplicates existing code, or has a dangerous edge case, it tells you before writing a single line.
- Session memory. A PostToolUse hook tracks every file anvil edits. Next session, if you touch that file again, anvil recalls past failures and accounts for them.
- Git autopilot. Checks git state, stashes uncommitted work, creates a branch, commits the result with a clean rollback command.
- Context7 docs. Bundled MCP server fetches up-to-date library docs so anvil doesn't hallucinate APIs.
Requirements
- Claude Code CLI (2.1.x or newer; plugin support).
- Python 3.9+ on your
PATH — used for the ledger and external reviewer scripts (stdlib only, no pip install).
- Git (obviously).
- Node.js (only needed for the bundled Context7 MCP server, launched via
npx).
- Optional reviewer credentials (see Reviewer setup below). Anvil gracefully degrades if a reviewer isn't configured.
Install
From GitHub (recommended)
In any Claude Code session:
/plugin marketplace add allut/claude-anvil
/plugin install claude-anvil@allut-claude-anvil
For development / local install
git clone https://github.com/allut/claude-anvil.git
claude --plugin-dir ./claude-anvil
Initialize the ledger once (it'll auto-init on first /anvil run, but this lets you verify the install):
python scripts/anvil-ledger.py init
# -> anvil ledger ready at ~/.claude-anvil/anvil.db
Configure
Run /anvil-setup from any Claude Code session. The wizard will:
- Ask which reviewers to enable (Claude / Gemini / Ollama / OpenAI-compatible).
- Collect API keys, validate each with a small live request, and write them to
~/.claude-anvil/config.json (chmod 0600 on POSIX).
- Pick the Claude reviewer's model (
sonnet / haiku / opus).
- Offer to
ollama pull if Ollama is enabled and the model isn't pulled yet.
- Optionally create bare shortcuts so you can type
/anvil instead of /claude-anvil:anvil (see below).
You can re-run /anvil-setup any time to reconfigure, or /anvil-setup reset to wipe and start over.
Bare command shortcuts (optional)
The wizard's final step offers to write ~/.claude/commands/anvil.md and ~/.claude/commands/anvil-setup.md, which expose the commands without the plugin namespace prefix. After accepting, you can use /anvil and /anvil-setup directly in any conversation.
Re-running /anvil-setup refreshes the shortcuts when the plugin updates. To create them manually:
python "${CLAUDE_PLUGIN_ROOT}/scripts/anvil-config.py" create-shortcuts "${CLAUDE_PLUGIN_ROOT}"
Reviewer options
| Reviewer | Cost | Get a key |
|---|
| Claude (Task subagent) | Covered by your Claude Code plan | Already installed — wizard picks the model (sonnet default, haiku faster/cheaper, opus deepest) |
| Ollama (local) | Free | Install from ollama.com; the wizard offers to ollama pull your chosen model |
| Gemini (Google AI Studio) | Free tier | aistudio.google.com/apikey |
| OpenAI-compatible | Paid for openai.com GPT-5; free via OpenRouter / Groq | The wizard offers OpenAI / OpenRouter / Groq presets, or a custom endpoint |
Default roster after wizard: Medium = first enabled reviewer; Large = up to 3 enabled reviewers (priority claude > gemini > openai > ollama).
Power users: any ANVIL_* env var still wins over config.json, so the legacy .env-based flow keeps working — see .env.example for the full list. The dispatcher (anvil-review.py) reads env first, then config.json, then defaults.
