Docs · Quickstart · Examples

caw (Coding Agent Wrapper) wraps multiple coding-agent CLIs — Claude Code, Codex, and opencode — behind a single Agent / Session API. Swap providers without changing your code, attach MCP tool servers, capture structured trajectories, and manage credentials for Docker containers. caw aims at the common cases with a small, ergonomic API — if you need fine-grained control over agent behavior, reach for the underlying agent SDKs; caw isn't trying to replace them.

caw is on PyPI as coding-agent-wrapper. The recommended path is uv:

uv add coding-agent-wrapper            # use as a library in a uv-managed project
uv tool install coding-agent-wrapper   # install just the CLI (caw, caw-traj) globally

Plain pip works too:

pip install coding-agent-wrapper

Requires Python 3.10+. You also need at least one provider CLI installed and authenticated (claude, codex, or opencode) — run caw doctor to see what caw can find.

For local development:

uv sync --extra dev

from caw import Agent

agent = Agent()  # defaults to claude_code
traj = agent.completion("Explain what this repository does")

print(traj.result)
print(f"{traj.usage.total_tokens} tokens, ${traj.usage.cost_usd:.4f}")

agent = Agent(model="claude-opus-4-7", reasoning="high")
agent.set_system_prompt("You are a security reviewer.")

with agent.start_session() as session:
    print(session.send("Review src/auth.py for vulnerabilities").result)
    print(session.send("Now check src/api.py").result)
# session.end() runs on exit and returns the full Trajectory

The most common way is the CAW_PROVIDER env var — set it once and every Agent() picks it up:

export CAW_PROVIDER=codex

from caw import Agent
agent = Agent()                        # uses whatever CAW_PROVIDER says

Or hand caw a fallback order and let it pick whichever is installed & healthy at runtime:

agent = Agent(provider=["claude", "codex", "opencode"])
traj = agent.completion("Reply with a one-line hello.")
print(f"[{traj.agent}] {traj.result}")  # whichever provider handled it

Decorate a Python function with @tool and pass it in — caw stands up a tool server for you:

from caw import Agent, tool

@tool(description="Add two numbers")
def add(a: int, b: int) -> int:
    return a + b

agent = Agent(stateless_tools=[add])
print(agent.completion("What is 17 plus 25? Use the tool.").result)

For stateful tools (shared state across calls in a session), subclass ToolKit — see the guide.

Every call returns a structured Trajectory. Save it and browse it later:

with Agent(data_dir="caw_data").start_session(traj_path="run.json") as session:
    session.send("List the Python files here and count them.")

Two viewers ship with caw:

caw-traj run.json          # compact, step-indexed terminal view
caw viewer                 # web UI, open run.json by path in the browser

• Providers — one API across Claude Code, Codex, and opencode.
• Auto-provider fallback — pick the first installed/healthy provider, fall back transparently.
• Models & tiers — portable ModelTier selection instead of hard-coded model strings.
• Sessions & resume — multi-turn conversations that resume across processes via a resume_handle.
• MCP tools, ToolKit, and subagents — give the agent tools, declaratively or as child agents.
• Provider health — raw availability/credential signals, with an optional live probe (caw doctor).
• Trajectory viewer — browse saved trajectories in the browser or terminal.
• Docker credentials — bind-mount agent OAuth credentials into containers without touching host files (caw auth).

Every interaction yields a structured Trajectory with turns, content blocks, token usage, and cost — persisted to JSONL when you pass data_dir=.

Full docs, guides, and API reference: https://zzjas.github.io/caw/. Machine-readable llms.txt / llms-full.txt are published alongside, since caw's users are often agents.

Runnable examples live in examples/.

Apache-2.0