Skip to main content

Ecosystem Compatibility

graph-tool-call keeps search and graph construction provider-neutral. Use an in-process adapter when the host is Python and MCP when the host runtime, language, or cloud platform should remain independent.

Support Matrix

EnvironmentRecommended SurfaceStatusValidation
Python applicationsToolGraph, GraphToolkitnativerepository tests
OpenAI Responses APIpatch_openainativerepository tests with Responses-shaped clients
OpenAI Chat Completionspatch_openaicompatibilityrepository tests
Anthropic Messagespatch_anthropicnativerepository tests
LangChain v1create_tool_selection_middlewarenativereal LangChain dependency test
Legacy LangGraph ReActcreate_agentcompatibilityrepository tests
OpenAI Agents SDKremote or stdio MCP serverprotocol integrationMCP contract tests; run an app smoke test
PydanticAIMCPToolset pointed at the remote MCP serverprotocol integrationrun an app smoke test
Google ADKMcpToolset pointed at the remote MCP serverprotocol integrationrun an app smoke test
AWS AgentCore Gatewayprivate remote MCP targetdeployment recipevalidate IAM/OAuth in the owning AWS account
Microsoft Foundry Toolboxprivate remote MCP endpointdeployment recipevalidate Entra/project connection in the owning tenant
JavaScript, Java, GoMCP clientprotocol integrationclient-specific smoke test required

“Protocol integration” means the graph-tool-call endpoint follows MCP, while the external framework remains responsible for connection lifecycle, approval, and execution policy. It does not mean every provider version is tested in this repository.

Decision Guide

  • Use patch_openai or patch_anthropic when an existing SDK client already owns execution and only the visible function catalog needs reduction.
  • Use LangChain v1 middleware when tool availability changes with runtime context, permissions, or conversation state.
  • Use MCP for OpenAI Agents, PydanticAI, Google ADK, desktop clients, and non-Python applications.
  • Use an AWS/Azure gateway in front of the MCP service for managed identity, OAuth, private networking, and policy enforcement.

Compatibility Policy

Core imports do not require framework packages. Optional integrations are lazy and must fail with an installation command rather than breaking import graph_tool_call. Release candidates run the core Python matrix and focused integration tests. Cloud recipes require an account-owned smoke test because credentials and network policy do not belong in this repository.

Known Boundaries

  • Product authentication and tenant policy stay in the host or managed gateway.
  • MCP clients may differ in dynamic tool-list caching; the proxy also keeps call_backend_tool as a fallback.
  • Direct TypeScript/Java SDKs are not shipped yet. Those runtimes use MCP.
  • OpenAPI callbacks and webhooks are metadata-only; GraphQL subscriptions need an application transport adapter.