crewai-openterms

CrewAI adapter for OpenTerms — wraps any callable so each invocation emits a signed ORS v0.1 receipt before the tool runs.

Depends on openterms-py>=0.1.0 for canonicalization, signing, and the HTTP client. One signing path is shared with langchain-openterms.

Install

pip install crewai-openterms

CrewAI itself is not a hard runtime dependency. CrewAI's BaseTool API has shifted repeatedly across the 0.x line (constructor signatures, the _run-vs-run hook, pydantic model versions), so this adapter wraps a plain callable — the function-tool pattern, which is stable across CrewAI versions. Plug the wrapped function into CrewAI however your project already does (Tool(name=..., func=...), the @tool decorator, or as the body of a BaseTool._run).

Quickstart

from openterms import IngestClient, generate_keypair
from openterms_crewai import OpenTermsToolConfig, wrap_tool

sk, _ = generate_keypair()
client = IngestClient(
    base_url="http://localhost:3000",
    workspace_id="00000000-0000-4000-8000-0000000000aa",
    key_id="my-key",
    private_key=sk.private_bytes_raw(),
    agent_id="crew-agent",
)
config = OpenTermsToolConfig(
    client=client,
    agent_id="crew-agent",
    terms_url="https://example.com/terms",
    terms_hash="a" * 64,
    emit_post_action=True,
)

def search(query: str) -> str:
    """Pretend to look something up."""
    return f"results for: {query}"

wrapped_search = wrap_tool(search, config=config, tool_name="search")

# Plug into CrewAI:
#   from crewai.tools import Tool
#   tool = Tool(name="search", func=wrapped_search, description="...")

Decorator form:

from openterms_crewai import openterms_tool

@openterms_tool(config, tool_name="search")
def search(query: str) -> str:
    return f"results for: {query}"

See examples/quickstart.py for a runnable copy.

Configuration

OpenTermsToolConfig carries:

• client — the openterms.IngestClient.
• agent_id — placed on every receipt.
• terms_url, terms_hash — terms binding for the receipts.
• emit_post_action — when True, emit a second receipt after the tool returns with post_state_hash = SHA-256 of the stringified result. Default False.
• strict — when True, re-raise openterms.IngestError; default False (failures logged, tool still runs).

What this adapter does

• Wraps a callable. Each invocation emits a signed ORS receipt with action_type="tool_call" and action_context = {tool_id, args}. Argument introspection is best-effort (inspect.signature first, positional fallback for builtins / C callables).
• Optionally emits a post-action receipt sharing the same receipt_id as the pre-action receipt.
• Logs and continues on ingest failures by default. The tool's return value is always passed through.

What this adapter does NOT do

• It does not depend on CrewAI at runtime. Tests run without crewai installed. The wrapper is callable-based.
• It does not run an ingest service. You must point IngestClient at a running OpenTerms API.
• It does not handle auth, host a JWKS, or manage keys. See openterms for the underlying scope.
• It does not buffer or retry. A failed emission is logged once (or re-raised if strict=True).
• It does not subclass CrewAI's BaseTool. That subclass tends to break across CrewAI minor versions. If you need a BaseTool subclass, build one in your project that calls into a wrap_tool-produced callable from inside _run.

Repository

Source and issue tracker: jstibal/openterms.

License

Apache-2.0.