Bring your own coding agent

Takuto orchestrates the agent — you choose which one. Four are supported today; their names stay in English everywhere.

  • Claude Code
  • Cursor Agent
  • Codex
  • OpenCode self-hosted models only (LM Studio, Ollama, vLLM…)

Claude Code

Anthropic’s coding agent, the default provider. For the headless pipeline, authenticate with an Anthropic API key (or a corporate proxy base URL).

Anthropic now requires an API key to run Claude Code in headless mode (Pro/Max OAuth no longer works unattended), so the author has not been able to test it recently — feedback especially welcome.

Cursor Agent

Cursor’s autonomous agent. Authenticate with an API key in your environment, or run its interactive login during setup.

Codex

OpenAI’s coding agent, run as a headless provider in the pipeline. Authenticate with an OpenAI API key in your environment.

Implemented but not yet tested by the author — try it and your feedback is especially welcome.

OpenCode

The open-source agent runner. With Takuto it runs against self-hosted models only — point it at your own LM Studio, Ollama, or vLLM endpoint.

Self-hosted only — tested, but not yet completed end-to-end (the author’s machine can’t run a capable model).

Provider authentication

Authenticate during the initial setup, or later from the dashboard’s Configuration page. Your tokens live in your own environment, on your own infrastructure — under your control. At a high level you have three options:

API key

Provide a provider API key via your environment — the simplest path for most setups.

OAuth login

Use a subscription login (for example Claude Pro/Max OAuth) instead of per-token API billing, where the provider supports it.

Corporate proxy

Point the agent at a corporate proxy base URL when your organisation routes AI traffic through its own gateway.

See the docs for exact provider configuration and the auth flow. Configuration docs →