Agent frameworks

The SDKs turn a provider's actions into LLM tools in the shape your framework expects, then run the model's tool-calls back through the governed action API — so policy, approvals, and audit apply to every call the model makes.

Coverage

The OpenAI/Anthropic/Gemini converters are plain shape transforms with no extra dependencies; the framework integrations (LangChain, CrewAI, Vercel AI) load their package lazily, so you only need it installed if you use that adapter.

Get tools and convert them

Fetch a provider's actions as framework-neutral tools, then convert to your model's format:

// Node
const tools = await client.tools.get('github');           // provider__action tools
const openaiTools = client.tools.toOpenAI(tools);         // or toAnthropic / toGemini

# Python
tools = client.tools.get("github")
openai_tools = client.tools.to_openai(tools)               # or to_anthropic / to_gemini / to_langchain / to_crewai

Run the model's tool-calls — governed

Hand the model's tool-calls back to Nolixan. Each one is executed through /actions/execute (policy + audit applied) and returned as role: "tool" messages. A failure — including a policy denial — comes back as an { error } payload instead of throwing, mirroring MCP's isError semantics, so the model can recover:

const toolMessages = await client.tools.handleOpenAIToolCalls(response.choices[0].message.tool_calls);
// append toolMessages to the conversation and call the model again

Tool names are provider__action (see Actions & tools). To call an action yourself instead of via the model, use execute.