clibridge

Run unmodified OpenAI, Anthropic, LangChain and LiteLLM code on your local Claude Code, Codex or OpenCode.

Set the model to claude-code, codex or opencode. No import, no proxy, no API key.

pip install "clibridge[all]"

from openai import OpenAI

reply = OpenAI().chat.completions.create(
    model="claude-code",                      # 👈 the only change
    messages=[{"role": "user", "content": "Hello!"}],
)
print(reply.choices[0].message.content)

Behind the scenes, clibridge runs the claude, codex or opencode command on your computer, so the request uses your existing login instead of an API key. Good for demos, testing and side projects. The chat UI above is in examples/chat.

---

Setup (once)

1. Install and log in to at least one CLI:

CLI	Install	Log in
Claude Code	npm i -g @anthropic-ai/claude-code	run claude once
Codex	npm i -g @openai/codex	codex login
OpenCode	see opencode.ai/install	opencode auth login

2. Install clibridge into the same Python environment as your code:

pip install "clibridge[all]"

3. Check it works:

clibridge doctor --ping

Or check your CLIs without installing anything: uvx clibridge doctor --ping.

That's it. No import, no server, no API key.

---

Model names

Model name	What runs
claude-code	Claude Code with your default model
claude-code/fable	Latest Fable
claude-code/opus	Latest Opus
claude-code/sonnet	Latest Sonnet
claude-code/haiku	Latest Haiku (fastest)
claude-code/<model id>	A specific version, e.g. claude-code/claude-opus-5-5
codex	Codex with your default model
codex/<model>	A specific Codex model, e.g. codex/gpt-5.5 (type /model in Codex to see yours)
opencode	OpenCode with the model from its own configuration
opencode/<provider>/<model>	A specific OpenCode model, e.g. opencode/anthropic/claude-sonnet-4
claude-code-agent, codex-agent, opencode-agent	Agent mode (see below). Works with a model too: claude-code-agent/opus

Which models you can use depends on your plan. Every other model name (gpt-4o, claude-sonnet-4-5, ...) still goes to the real API with your real key, as before.

---

Examples

OpenAI SDK

from openai import OpenAI

OpenAI().chat.completions.create(model="claude-code", messages=[{"role": "user", "content": "Hi"}])

LiteLLM

import litellm

litellm.completion(model="claude-code", messages=[{"role": "user", "content": "Hi"}])

LangChain / LangGraph

from langchain_openai import ChatOpenAI

ChatOpenAI(model="claude-code").invoke("Hi")

Anthropic SDK

from anthropic import Anthropic

Anthropic().messages.create(model="claude-code", max_tokens=1024, messages=[{"role": "user", "content": "Hi"}])

OpenAI Agents SDK

from agents import Agent, Runner

print(Runner.run_sync(Agent(name="Helper", model="claude-code"), "Hi").final_output)

Streaming, tool calling, structured output (JSON / Pydantic), images and async all work. Runnable versions are in examples/.

---

Thinking

Set it the way your library normally does. If you don't set it, the CLI's default is used.

Library	How
OpenAI SDK, LiteLLM	reasoning_effort="high"
OpenAI Responses	reasoning={"effort": "high"}
OpenAI Agents SDK	model_settings=ModelSettings(reasoning=Reasoning(effort="high"))
Anthropic SDK, LiteLLM	thinking={"type": "enabled", "budget_tokens": 10000} or {"type": "disabled"}

Levels: none, minimal, low, medium, high, xhigh, max. Each is passed to the CLI as its nearest equivalent; none turns Claude's thinking off. OpenCode has no effort setting, so any level other than none turns its thinking summaries on; none hides them.

The thinking comes back where your library expects it:

Library	Where
OpenAI SDK, LiteLLM	reply.choices[0].message.reasoning_content
Anthropic SDK	thinking blocks (when you turned thinking on)
OpenAI Responses / Agents SDK	a reasoning item (when you set an effort)

It's a short summary, not the full thoughts, and the model decides when to think. Easy questions often come back without any.

---

What's ignored

Anything the CLIs don't support is ignored, so your code still runs: max_tokens, temperature, top_p, stop, n, seed, logprobs.

Not available at all: embeddings, audio, image generation.

---

Agent mode (optional)

Normally the CLI acts as a plain AI model: no tools, run from an empty folder.

With claude-code-agent, codex-agent or opencode-agent, it works like a coding agent inside a folder, using its own tools to read files, search and run commands. It's read-only unless you allow more:

export CLIBRIDGE_WORKDIR=/path/to/project       # folder to work in (default: current folder)
export CLIBRIDGE_PERMISSION_MODE=acceptEdits    # let Claude edit files
export CLIBRIDGE_SANDBOX=workspace-write        # let Codex edit files
export CLIBRIDGE_OPENCODE_AUTO=1                # let OpenCode edit files and run commands

---

Not using Python? Run the server

For JavaScript, curl, or code on another machine:

clibridge serve

export OPENAI_BASE_URL=http://localhost:8787/v1      # OpenAI-style clients
export ANTHROPIC_BASE_URL=http://localhost:8787      # Anthropic-style clients
export OPENAI_API_KEY=anything                       # any value works

---

Handy commands

clibridge ask "What is a monad?"              # quick test
clibridge ask -m codex "Write a regex for emails"
clibridge ask -m opencode/anthropic/claude-sonnet-4 "Explain this error"
clibridge ask -e high -t "Is 391 prime?"      # think hard and show the thinking
clibridge doctor --ping                       # are the CLIs installed and logged in?

Good to know

• Each call starts a CLI process, which adds a few seconds.
• Every call counts against your plan's limits, and your plan's terms still apply.
• ANTHROPIC_API_KEY and OPENAI_API_KEY are kept away from the CLIs, so they always use your existing login.
• OpenCode's run command is always an agent. A plain opencode call runs in an empty temporary folder with project instructions and skills switched off and is told not to use tools; use opencode-agent to work in a folder.
• OpenCode's JSON output does not report token usage for the last step, so usage can be zero or partial for opencode calls.
• At most 4 calls run at once (CLIBRIDGE_MAX_CONCURRENCY). Calls stop after 600 seconds (CLIBRIDGE_TIMEOUT).
• CLIBRIDGE_AUTOLOAD=0 turns the automatic mode off.

Troubleshooting

You see	Fix
Missing credentials / api_key must be set	clibridge isn't installed in this Python. Run clibridge doctor with the same Python
Could not find the claude CLI	Install it, or set CLIBRIDGE_CLAUDE_BIN=/path/to/claude
Could not find the opencode CLI	Install it, or set CLIBRIDGE_OPENCODE_BIN=/path/to/opencode
Invalid API key · Please run /login	Run claude and log in
codex: ... usage limit	You hit your plan's limit; wait for it to reset
opencode: Model unavailable	Run opencode models, then use a real opencode/<provider>/<model>

How it works

When Python starts, clibridge adds a tiny hook (about 0.3 ms). When your code calls a claude-code, codex or opencode model, the request is turned into a prompt, run through claude -p, codex exec or opencode run, and the answer is turned back into the format your library expects.

Running the tests

The tests use fake claude / codex / opencode scripts, so they need no login and use no quota.

git clone https://github.com/rumbleroii/clibridge.git
cd clibridge
pip install -e ".[dev]"
pytest