Executors

agenix agents shows the up-to-date list. The matrix:

Agent	Native configs	Local models	Auth
claude-code	CLAUDE.md, .claude/skills/, .mcp.json	via gateway	subscription / key
codex	AGENTS.md, .codex/config.toml	--oss / providers	subscription / key
cline	.clinerules/	Ollama, LM Studio, …	key
openhands	.openhands/microagents/repo.md	litellm	key
aider	CONVENTIONS.md, .aider.conf.yml	litellm	key
goose	.goosehints	Ollama and others	key
opencode	AGENTS.md, opencode.json	many providers	key
qwen-code	QWEN.md, .qwen/settings.json	local Qwen	subscription / key
mini-swe-agent	KNOWLEDGE.md (+prompt)	litellm	key
gemini-cli	GEMINI.md, .gemini/settings.json	none	subscription / key

How an adapter works

Four operations (src/adapters/base.ts):

• materialize(profile, kb, workspace) — write the native configs;
• buildCommand(profile, prompt) — non-interactive single-task run;
• buildEnv(profile, apiKey) — model/gateway connection variables;
• run(...) — the shared loop: materialization → process → RunResult.

Adding a new executor = one file with these methods + registration in registry.ts.

Live testing status

The full smoke (2/2) has passed for seven executors: claude-code (subscription, no gateway), codex (ChatGPT subscription, model.id: builtin, no gateway), cline, openhands, goose (via the LiteLLM gateway on open-source models), mini-swe-agent and qwen-code (OpenRouter directly, key from the agenix store).

Special cases:

• aider — applies edits flawlessly (hello-file PASS on all models), but does not execute proposed shell commands in headless mode without a TTY — that is by design; don't give it "run a command" tasks;
• opencode — config/models/permissions compile correctly, the model returns tool calls, but headless opencode run (v1.16–1.17) doesn't execute them — an upstream bug, track via opencode releases;
• swe-agent — deferred (the PyPI package is broken, official install is from git); gemini-cli — not exercised (requires a Google login).

Notes

• codex: in subscription mode CODEX_HOME is not overridden (the OAuth token lives there), the model is passed via -m; in key mode CODEX_HOME points at the workspace and the profile fully controls the provider and MCP via config.toml.
• cline (verified live on CLI 3.x): the adapter configures the provider via cline auth into an isolated --data-dir inside the workspace — global state is untouched. Cline sends stream_options, which strict providers reject — go through a LiteLLM proxy with additional_drop_params: ["stream_options"] (example: examples/gateway/litellm-ghmodels.yaml); the proxy's baseUrl must include /v1.
• openhands (verified on CLI SDK v1.21): install the openhands package (not openhands-ai — that one has no CLI binary). The key flag is --override-with-envs: without it LLM_MODEL/LLM_API_KEY/LLM_BASE_URL are silently ignored. Mind the size of the OpenHands system prompt (>8k tokens) — models with a small request limit (e.g. the GitHub Models free tier) won't work.
• aider: install on Python ≤ 3.12 (uv tool install --python 3.12 aider-chat) — on 3.13 it crashes due to the audioop module removed from the stdlib.
• free cloud open-source models: GitHub Models (baseUrl: https://models.github.ai/inference, the key is a regular GitHub token, see profiles/aider-ghmodels.yaml). Mind the strict rate limits of the free tier: single runs pass, a series of back-to-back evals hits 429.
• mini-swe-agent: intentionally minimal (no rule files or MCP) — knowledge is supplied as KNOWLEDGE.md referenced in the prompt; an ideal cheap "test subject" for evolve experiments.