Overview

MANTYX is an agent operating system: it owns the LLM loop, the workspace tool catalog, memory, skills, and persisted observability. The SDKs let you drive that runtime from your own process — define ephemeral agents inline, trigger persisted MANTYX agents by id, and seamlessly mix remote workspace tools with local tools that run in your process and shuttle results back over the agent loop.

What you can do with the SDKs

• Run an ephemeral agent — describe a system prompt, model, and tool list on the call site. MANTYX runs the loop and streams results back.
• Trigger a persisted MANTYX agent (agentId) — reuse an agent that already lives in your workspace (with its system prompt, model, memory, skills, and tool list) and optionally merge in extra local tools for that single run.
• Maintain conversational sessions — multi-turn agent runs whose history persists on the server, with optional per-turn tool refresh.
• Mix remote and local tools — server-resolved (mantyx, mantyx_plugin, a2a, mcp) or client-resolved (local, a2a_local, mcp_local). Connect public peers, internal services, on-device MCP servers, and your own functions in the same agent.
• Delegate across agents — call out to other Agent2Agent peers, whether MANTYX can reach them or only your SDK can.
• Connect MCP servers — expose every tool of an MCP server (remote Streamable HTTP or local stdio) to the loop in one go.
• Stream tokens — assistant deltas, thinking deltas, server tool results, local tool calls, and the terminal result event over SSE.
• Pick a model — choose a workspace BYOK provider, a specific vendor model, or a platform-hosted offering via a unified modelId string.
• Tune thinking effort — set reasoningLevel per run ("off" | "low" | "medium" | "high" or 0–100) and MANTYX maps it onto each provider’s native dial.
• Constrain replies to JSON — pass an outputSchema (a JSON Schema) and the model’s final message is guaranteed to be parseable JSON; each SDK ships a parseRunOutput helper that decodes it into your own typed value.
• Tag for observability — attach a flat metadata KV (e.g. { customer: "acme", env: "prod" }) to runs and sessions so your team can filter the dashboard by them.

Three first-party SDKs

	TypeScript	Go	Python
Package	@mantyx/sdk	github.com/mantyx-io/mantyx-sdk/go	mantyx-sdk
Install	npm install @mantyx/sdk (bundles zod + @modelcontextprotocol/sdk)	go get github.com/mantyx-io/mantyx-sdk/go	pip install mantyx-sdk
Min runtime	Node.js 18.17+	Go 1.24+	Python 3.10+
Local tool params	Zod schema	tagged Go struct	Pydantic v2 model

All three speak the same wire protocol (see Wire protocol) and expose the same conceptual surface — runAgent, streamAgent, createSession, resumeSession, endSession, listModels, cancelRun — adapted to language conventions.

Next steps

• Authentication — how to generate an API key.
• Quickstart — your first run, in any of the three SDKs.
• Wire protocol — the HTTP + SSE spec, the source of truth for third-party clients.