anumati

Stop approving the same Bash commands over and over. anumati is a deterministic auto-approver for AI coding agents. From a small config of named matchers it auto-allows safe shell commands — so git status, npx tsc, cargo test, jq, and friends just run, while genuinely risky commands still get a prompt. It's agent-agnostic: one shared config drives both Claude Code (via its PreToolUse hook) and OpenAI Codex (via its PermissionRequest hook).

When something falls through, anumati tells you the exact one-liner to allow it next time — so your config builds itself from real usage.

$ git status && cargo test | tail -20
  ✓ auto-approved (no prompt)

$ terraform apply
  ⤳ prompt shown  ·  💡 anumati: no matcher covers "terraform"

Why anumati

• Deterministic, not vibes. Approval is decided by explicit matchers with a strict grammar — not an LLM guessing whether a command is safe. The same command always gets the same answer.
• Safe by construction. Matchers allow only read-only / build / test shapes. Redirects that write files, $(...) substitution, network curl to unlisted domains, git push, rm — all fall through to a real prompt. anumati is allow-only: it can approve a call or step aside, but never blocks anything itself, and never widens what the agent would otherwise refuse.
• Composes across a whole command line. git status && cargo build | tail is approved only if every piece is independently safe — you can't smuggle rm -rf / in by chaining it onto an allowed command.
• Self-building config. Every passthrough comes with a verified suggestion (anumati add …) and a logged reason, so you extend coverage from what you actually run.
• Bash-only by design. anumati vets Bash — the hard problem. Read / Write / Edit stay with the agent's own permission flow.

How it works

Every time the agent is about to run a Bash command, anumati checks it against your allow rules:

1. A rule matches → auto-approved, no prompt.
2. No rule matches → the agent shows its normal permission prompt, and anumati prints a 💡 suggestion for allowing it next time.

A command is approved one of two ways: a single matcher accepts the whole thing, or — failing that — anumati splits it at top-level &&, ;, ||, &, and newlines and approves only if every sub-command is independently accepted.

flowchart TD
    A[Bash command] --> B{"A single rule's matcher<br/>accepts the whole command?"}
    B -- yes --> ALLOW([✅ allow])
    B -- no --> D["Split at top-level && ; || & and newlines<br/>pipes stay glued to their segment"]
    D --> F["For each sub-command:<br/>does some rule accept it?"]
    F --> G{"Every sub-command<br/>approved?"}
    G -- yes --> ALLOW
    G -- no --> PASS([⤳ passthrough])

A disallowed sub-command always fails its own check, so chaining a bad command onto a good one can't sneak it through. Pipes are never split across rules (a pipe feeds data into the next command, so only the matcher owning the pipeline can judge it). Configs cascade: a project config at <cwd>/.anumati/permissions.json is checked before your global ~/.anumati/permissions.json (the legacy ~/.claude/ locations are still honored, so existing setups keep working).

The full model — matchers, composition rules, and safety guarantees — is in docs/CONFIGURATION.md.

Install

npm install -g anumati

Or run without installing via npx anumati ~/.claude/permissions.json.

Update to the latest version any time with anumati update (it checks the published version and, if newer, runs the global reinstall for you; the hook picks it up on the next command). anumati update --check just reports whether a newer version exists.

Quick start

One command sets up everything:

anumati init

It prompts for which agent(s) (Claude Code / Codex / both) and project (this folder) or root (global) scope, then:

1. Writes a starter config of broadly-useful, low-risk rules — read-only inspection, git reads, cd/sleep/echo/sed/jq, npx tsc, cargo, go, test runners (vitest/pytest/jest), and pure-compute python3/node. Enough to be useful immediately. It also seeds the parameterized matchers (curl, gh, pip3-install, git-write, git-push, node-script) as inert placeholders — empty allowlists that approve nothing until you fill them in, so you can see the matcher exists and which key to populate (anumati add <matcher> …) without hunting the docs.
2. Scaffolds an audit log next to the config.
3. Wires the chosen agent(s), merged non-destructively — Claude Code's PreToolUse hook in settings.json, and/or Codex's PermissionRequest hook in ~/.codex/hooks.json.
4. Adds a SessionStart banner ("⚡ anumati active") so you can see it's on.
5. Writes command-style guidance to the sibling CLAUDE.md, nudging the agent to emit approvable commands.

Reload the agent for it to take effect (Claude Code: /hooks or restart; Codex: approve the anumati hook once when it prompts to review it). Then just work — routine commands stop prompting, and when something new falls through you'll see a 💡 anumati add … suggestion.

Grow your config as you go:

anumati review                              # turn your passthrough log into config changes, interactively
anumati add curl --domain api.github.com   # allow curl to a domain
anumati add git-write --git-ops add,commit  # allow specific git writes
anumati stats                               # see your auto-approve rate
anumati apply --all                         # apply accumulated suggestions

anumati review is the quickest way to close the gap: it reads your passthrough log, shows the exact rule that would auto-approve each fall-through command (and how many commands each covers), and applies the ones you pick — so you don't have to hand-write anumati add incantations. It only ever proposes changes anumati itself verified, and re-runs with --all to apply everything or --dry-run to just preview.

Docs

• Configuration reference — every matcher, rule field, audit/sound/debug option, and CLI subcommand.
• Claude Code + Codex — running anumati across both agents from one shared config: per-agent wiring, the Codex integration, and anumati migrate.
• Command-style guide — how to write commands that land on the auto-approve path (also installed into CLAUDE.md by init).

Contributing

See CONTRIBUTING.md for local setup — including running Claude Code against your local build via npm link — plus how to test, add a matcher, and open a PR.

License

MIT