Claude Code#

What Is This?#

The claude-code provider lets you use Agentic Programming without any API key. It uses your Claude subscription's OAuth token to connect directly to api.anthropic.com — the token is resolved from the Claude Code CLI's login credentials (~/.claude/.credentials.json) and re-read on every call, so the CLI's token refreshes take effect automatically.

If you have claude installed and logged in, you're ready to go.

Prerequisites#

1. Install the Claude Code CLI:

   npm install -g @anthropic-ai/claude-code
   

2. Log in:

   claude auth login
   

3. Verify it works:

   claude -p "Hello, world!"
   

That's all the setup needed. No API keys, no environment variables.

Basic Usage#

from openprogram import agentic_function
from openprogram.providers.registry import create_runtime

# No API key needed — uses your Claude Code subscription
runtime = create_runtime(provider="claude-code", model="haiku")

@agentic_function
def explain(concept):
    """Explain a concept clearly and concisely."""
    return runtime.exec(content=[
        {"type": "text", "text": f"Explain '{concept}' in 2-3 sentences. Be clear and concise."},
    ])

result = explain(concept="gradient descent")
print(result)

Configuration Options#

runtime = create_runtime(
    provider="claude-code",
    model="haiku",        # model name or family alias (see the table below)
    api_key=None,         # usually omitted; when omitted, resolved from the credential pool on every call
    max_retries=2,        # retries for transient API-layer failures
)

Model Names#

model accepts a family alias or a full model id. Aliases expand to the current default version:

More specific ids (such as claude-opus-4-5-20251101) are passed through as is and validated by the Anthropic API.

How It Works#

Under the hood, the claude-code runtime:

1. Resolves your Claude subscription's OAuth token (sk-ant-oat prefix), or a plain Anthropic API key, from the credential pool
2. Connects directly to api.anthropic.com over the standard Anthropic Messages protocol, using Bearer auth plus the Claude Code identity header for subscription tokens
3. Writes the reply back into the session DAG

Your Python code
    → @agentic_function decorator (records a DAG node)
        → runtime.exec() (builds the prompt from the DAG)
            → api.anthropic.com (direct subscription OAuth connection)
            ← response text
        ← reply written back as a DAG node
    ← return value

No subprocess is involved — it is the standard Anthropic protocol, just with credentials from your subscription.

Limitations#

• Requires valid credentials. Construction verifies that a Claude subscription OAuth token or an Anthropic API key exists in the credential pool, and raises ValueError otherwise.
• Subscription tokens expire (roughly every 8 hours). The runtime re-resolves them on every call, so refreshes on the Claude Code CLI side take effect automatically; if you haven't used claude in a long time, just log in again.

Complete Example#

"""
Claude Code integration demo — no API key needed.
Demonstrates a multi-step agentic workflow.
"""
from openprogram import agentic_function
from openprogram.providers.registry import create_runtime

runtime = create_runtime(provider="claude-code", model="haiku")


@agentic_function
def brainstorm(topic):
    """Generate 3 creative ideas about a topic."""
    return runtime.exec(content=[
        {"type": "text", "text": f"Generate exactly 3 creative ideas about: {topic}\nNumber them 1-3, one per line."},
    ])


@agentic_function
def evaluate(idea):
    """Rate an idea's feasibility on a scale of 1-10 with brief reasoning."""
    return runtime.exec(content=[
        {"type": "text", "text": f"Rate this idea's feasibility (1-10) and explain in one sentence:\n{idea}"},
    ])


@agentic_function
def ideate(topic):
    """Brainstorm ideas and evaluate each one."""
    ideas_text = brainstorm(topic=topic)
    print(f"Ideas:\n{ideas_text}\n")

    lines = [l.strip() for l in ideas_text.split("\n") if l.strip() and l.strip()[0].isdigit()]
    for line in lines[:3]:
        rating = evaluate(idea=line)
        print(f"  {rating}\n")

    return runtime.exec(content=[
        {"type": "text", "text": "Pick the best idea from the evaluation above and explain why in 2 sentences."},
    ])


if __name__ == "__main__":
    result = ideate(topic="improving developer productivity with AI")
    print(f"\nBest idea:\n{result}")

Troubleshooting#