Choose your agent

grith ships with built-in supervisor profiles for the popular AI agents. A profile is a curated set of routine paths, commands, and network destinations that the agent needs to do its job — paths grith already trusts and routes to auto-allow without quarantining.

You don't have to use a built-in profile. The generic profile gives you strict defaults (project-only file access, no shell, no network) and you can write your own for any agent. But starting from a built-in profile is faster and tends to produce a quieter digest.

Built-in profiles

Agent	Profile flag	Notes
Claude Code	--profile claude-code	Anthropic API + node/cargo/python toolchains + GitHub
OpenAI Codex CLI	--profile codex	Codex's stdlib of editors and exec patterns
Aider	--profile aider	git-heavy workflow + provider API
Goose	--profile goose	Block's agent with native tool calling
GitHub Copilot CLI	--profile copilot	gh-cli integration and suggestions
Cursor (CLI/agent mode)	--profile cursor	Cursor's helper subprocesses
Cline (Claude Dev / extension CLI)	--profile cline	Browser auth, MCP servers
OpenClaw	--profile openclaw	Community fork patterns
Anything else	--profile generic-cli	Standard CLI tools (git, npm, pip, etc.)
Strictest	--profile generic	Project-local files only, nothing else

All profiles extend generic, so any restriction set in generic applies to every profile unless explicitly relaxed.

How a profile shapes scoring

The profile changes which calls go directly to auto-allow versus get scored fully. Concretely, a profile declares:

• routine_paths — read/write paths the agent needs (e.g. ${PROJECT_DIR}/**, ~/.cache/{agent}/**).
• routine_commands — shell commands the agent runs frequently (e.g. git status, npm install, cargo build).
• routine_destinations — network hosts the agent needs (e.g. api.anthropic.com, github.com, registry.npmjs.org).
• routine_exec_roots — directories the agent's helper processes live in.
• readonly_paths — paths it can read but not write.

When a syscall hits a routine entry, the corresponding filter contributes a negative score (i.e. argues for allow) and the call usually auto-allows. Anything outside the routine set runs through the full filter stack at normal weight.

Picking a profile when in doubt

1. You're running a known agent → use its named profile.
2. You're running a wrapper around a known agent → use the wrapped agent's profile.
3. You're shelling around in bash directly → use generic-cli.
4. You're paranoid or testing → use generic and let everything land in the digest.

Custom profiles

If your team has a house tool, an internal agent, or a fork of an upstream agent, you can either:

• Extend an existing profile by copying it into ~/.config/grith/profiles/ and editing.
• Write one from scratch — see Writing a custom profile.

For Pro teams, profiles can be centrally managed and synced — see Team sync.

Auditing a profile

After a session, grith profile audit --profile <name> --trace <syscall-log> rebuilds the call graph and shows which routine entries were exercised, which were never touched, and which calls escaped the routine set and went through full scoring. Useful for tightening a profile that's grown too permissive.

Next

• Connect a model — BYOK setup so your agent has something to talk to.
• Your first supervised session — worked example end-to-end.