Connect the Codex and Claude Code CLIs you are already logged in to, plus OpenCode's free models (no login needed), to agent-use. Install once, with no handwritten configuration, then hand tasks to them with one command; if one is unavailable, the next one takes over.
- The agent-use core ships only demo adapters. To use real models you would write your own registry and learn each CLI's arguments and failure modes.
- Calling these CLIs directly from scripts is risky. It is easy to leave tool access on, write files into your working directory, or pile up session history.
- One install command, three CLIs ready. Logged-in Codex and Claude Code work right away; OpenCode's free models need no login.
- If one is unavailable, the next takes over. The default order is Codex → Claude Code → OpenCode, or you can use just one. If you interrupt a run, the whole chain stops instead of trying the next one.
- Conservative defaults. Every task runs in an empty temporary directory: Codex uses a read-only sandbox, Claude Code has all tools disabled, OpenCode allows only free models and denies tool calls; sessions are not kept (best-effort for OpenCode).
- Consistent failure reasons. Missing login, exhausted quota, rate limits, timeouts and more each have a fixed reason code, so callers can decide what to do next.
Does it cost money? Calls use the selected CLI's own account and quota. OpenCode allows only -free models; if the upstream requires payment, the call fails instead of switching to a paid model.
Does it touch my credentials? No. It never reads or stores them; each CLI handles its own authentication.
Where do tasks go? To the selected CLI and its upstream service. Tasks and answers are saved as plain text on your machine.
Will it overwrite my existing configuration? No. Repeating installation is safe, and an existing different file or symlink is never overwritten.
Requires Python 3.11+, Bash, and POSIX (macOS/Linux). Install the official CLIs first. Existing official CLI logins work directly; no additional Python packages are needed.
Clone both repositories next to each other:
git clone https://github.com/Jerryx333/agent-use.git
git clone https://github.com/Jerryx333/agent-use-adapters.git
cd agent-use-adapters
bash scripts/install.sh
cd ../agent-use
python3 -B scripts/agent_use.py list
printf '%s' '{"version":1,"task":"Explain what a CLI is in one sentence.","providers":["codex"],"output":"runs/first"}' | python3 -B scripts/agent_use.py runAuthenticate first if needed: codex login, or start claude and use /login.
OpenCode free Zen models need no login. A model run uses the selected CLI's quota.
Successful output is a draft: state: staged_for_review, accepted: false.
Inspect runs/first/response.txt and runs/first/status.json; use a fresh output
directory for a new task or retry.
Omit providers to try the default order: Codex → Claude → OpenCode, falling back
when a route fails (except after an interruption). Named profiles are default, codex-only, claude-only, and
opencode-only; use "profile":"opencode-only" instead of providers to select one.
Where the installer puts the registry and how to use your own: see the guide.
| Adapter | Requirements / official login | Safety measures | Supported options |
|---|---|---|---|
codex |
Codex CLI; codex login |
Empty temporary directory, read-only sandbox, ephemeral session | effort: minimal, low (default), medium, high, xhigh |
claude |
Claude Code; /login inside claude |
Empty temporary directory, all tools disabled, one turn, no session persistence | effort: low (default), medium, high, max |
opencode |
OpenCode CLI with run --pure; no login for free Zen models |
Only -free models; tool calls require authorization and are denied; deletes observed sessions after execution; tries fallback_models in order within a shared time budget |
fallback_models: comma-separated -free model names |
Defaults: gpt-6.1-sol, claude-opus-5-5, and
muse-spark-1.3-contributor-free. OpenCode's default backups are
mimo-v2.6-flash-free,longcat-2.5-preview-free. Model availability depends on the CLI
and upstream service. Override executable locations with CODEX_BIN, CLAUDE_BIN,
or OPENCODE_BIN if needed.
- Usage guide: installation details, failure reasons, trust model.
- The agent-use core's dispatcher protocol.
Run offline checks with bash scripts/test.sh. Submit changes through PRs; see CONTRIBUTING.md for development. License: Apache-2.0.