Module interop 

Wire-level interop: guard LLM tool calls from any agent framework.

Agent frameworks (Pydantic AI, LangChain, the OpenAI and Anthropic SDKs) all share the same last-mile shape: the model emits a tool call — a tool name plus JSON arguments — and the host process decides whether to run it. That decision point is exactly where Typesec belongs, and it is the one place none of the frameworks guard for you.

This module provides:

• ToolCallRequest — a framework-neutral, normalized tool call.
• ToolBinding — the declaration of how one tool maps onto the Typesec (action, resource) plane, optionally taking the resource from a tool argument.
• ToolCallGuard — evaluates normalized calls against any PolicyEngine, deny-by-default for tools without a binding.
• Dialect codecs (openai, anthropic, langchain, pydantic_ai) that parse each framework’s wire shape into ToolCallRequests and render denials back in the shape the framework expects (an error tool-result / retry part), so a blocked call flows back to the model as feedback instead of crashing the run.

model output ─▶ dialect::parse_tool_calls ─▶ ToolCallGuard::check_all
                                                  │ Allow ─▶ run the tool
                                                  └ Deny  ─▶ dialect::denial ─▶ model

The typed ProtectedTool path remains the strongest boundary (a capability is required to compile the call); this module is the runtime bridge for tools that live on the other side of a JSON wire, where Rust types cannot reach.

Modules

anthropic

Anthropic Messages dialect: tool_use content blocks in, tool_result blocks with is_error: true out.

dialects

Named dialect registry: one dispatch table shared by every binding layer (Python, WASM, the CLI) so dialect names can’t drift between them.

langchain

LangChain dialect: AIMessage.tool_calls in, error ToolMessages out. The same shapes serve LangGraph nodes, which pass LangChain messages.

mcp

Model Context Protocol dialect: JSON-RPC tools/call requests in, isError tool results out.

openai

OpenAI Chat Completions dialect: tool_calls in, role: "tool" denial messages out. Works with the OpenAI SDKs and any OpenAI-compatible server.

pydantic_ai

Pydantic AI dialect: tool-call message parts in, retry-prompt parts out. Complements the capability metadata in typesec-integrations::pydantic_ai — the metadata declares the binding, this codec enforces it on the wire.

Structs

GuardedToolCall

A tool call together with its resolved binding and verdict.

ToolBinding

Declares how one tool maps onto the Typesec (action, resource) plane.

ToolCallGuard

Evaluates normalized tool calls against a policy engine.

ToolCallRequest

One tool invocation requested by a model, normalized across frameworks.

Enums

InteropError

A framework payload could not be interpreted as tool calls.

TaintError

Why tool output could not be protected.

ToolCallVerdict

The guard’s verdict on one tool call.

Constants

TOOL_OUTPUT_KIND

Resource kind attached to tool-output resources.