runx-py

Python SDK for runx, the governed runtime for agent skills, tools, and graphs.

runx-py is a thin Python client over the runx CLI JSON output. Install the CLI separately (@runxhq/cli on npm), then use this package from Python to search and run skills, continue runs awaiting agent input, and format host protocol results for popular agent frameworks.

See the TypeScript interop boundary for the package disposition and ownership rules.

Install

pip install runx-py

You will also need the runx CLI on your PATH:

npm install -g @runxhq/cli

Usage

from runx import RunxClient

client = RunxClient()

# Search the registry
for result in client.search_skills("web search"):
    print(result.skill_id, result.version, result.install_command.copy_text)

# Run a skill: each input is passed as -i name=value and read by the type the
# runner declares.
report = client.run_skill("runx/web-search", inputs={"query": "receipt verification"})
print(report["status"])

continued = client.continue_run("runx/web-search", run_id="run_123", answers_file="answers.json")
print(continued["status"])

Framework adapters

Bridge runx into an existing agent framework (OpenAI, Anthropic, CrewAI, LangChain, Vercel AI):

from runx import create_host_bridge, create_openai_host_adapter

bridge = create_host_bridge(run=my_host_run, continue_run=my_host_continue)
adapter = create_openai_host_adapter(bridge)
response = adapter.run("runx/web-search")

The bridge translates host protocol results, including needs_agent runs and approval gates, into framework-native tool messages. RunxClient remains a CLI client; host protocol execution is provided by the embedding runtime.

Links

• Homepage: https://runx.ai
• Documentation: https://runx.ai/docs
• Source: https://github.com/runxhq/runx
• Issues: https://github.com/runxhq/runx/issues

Releasing

See RELEASING.md for the automated tag-driven publish flow.

License

Apache-2.0