Skip to main content

​
Set up Termyte for coding agents

Termyte currently supports these agent names:
  • codex
  • claude
  • claudecode
  • aider
claudecode falls back to the claude executable when needed.

​
Before you launch an agent

Set up repository policy and verify it: 
termyte policy show
termyte policy test "cat .env"
termyte doctor
Add repository-specific rules before giving the agent broader autonomy: 
termyte policy local add "Ask before touching auth or payments" --yes
termyte policy local add "Block destructive database commands" --yes

​
Inspect the launch plan

termyte run --dry-run codex
The plan shows:
  • requested agent;
  • resolved executable or missing-executable warning;
  • selected profile;
  • planned shims and hooks.
Dry run does not launch the agent.

​
Launch an agent

termyte run codex
termyte run claude
termyte run claudecode
termyte run aider
You can pass arguments after the agent name: 
termyte run codex --help

​
What Termyte prepares

Before launching the agent, Termyte:
  • finds the repository root;
  • verifies local state is readable and writable;
  • loads built-in, global, and repository YAML policy;
  • enables repository-local logs and memory;
  • creates a Termyte session ID;
  • passes agent and workspace context through environment variables.

​
Current runtime mode

Agent launching currently uses limited mode. The agent process receives Termyte context, but termyte run <agent> does not guarantee that every command or subprocess created by the agent is intercepted.
Use the stable check and policy workflows to verify rules. Use termyte shell only when you intentionally want to test the experimental intercepted runtime.

​
Missing agent executable

If Termyte cannot find the requested agent, it exits without launching it. 
termyte run --dry-run codex
termyte doctor
Install the agent and ensure its executable is on PATH, then retry.