Open source · MIT · v0.1.0
Agentmatic takes your LangGraph code, your LangChain tools, and your existing workflows — and runs them on a compiled Rust engine. Change one import. Keep everything else. Get built-in circuit breakers, retry, dead-letter queues, and self-hosted distributed execution — MIT-licensed, no platform fees.
pip install agentmaticfrom agentmatic import Agent, OpenAI, tool
@tool
def calculator(expression: str) -> str:
"""Evaluate a math expression."""
return str(eval(expression))
agent = Agent.build(
"math-helper",
llm=OpenAI(),
tools=[calculator],
)
result = agent.invoke("What is 42 * 17?")Same API surface in every language. Same Rust engine underneath.
Why agentmatic
LangGraph nailed the abstraction. We compiled it. Same paradigm, native execution, production primitives in the box.
Same StateGraph API. Same TypedDict state. Same conditional edges, checkpointers, interrupt_before/after, time travel. Change one import.
Learn more →Compiled Rust engine with 70–80× faster channel throughput and 50–80% lower memory. Your agents respond before they would have started.
Learn more →Circuit breakers, exponential-backoff retry, and dead-letter queues — first-class primitives. No platform tier, no SaaS.
Learn more →Self-hosted coordinator/worker or peer-to-peer over gRPC. Scale to millions of runs without leaving your VPC.
Learn more →Memory, SQLite, Postgres, Redis, S3 backends. Time-travel debugging, branch-and-replay, human approval interrupts — all built in.
Learn more →First-class Model Context Protocol client (stdio + SSE). Bridge any LangChain tool with from_langchain_tools(). 1000s of integrations.
Learn more →Python (PyO3), TypeScript (napi-rs), Rust, Go (CGO), Java (JNI). Same API surface in every language. Polyglot teams ship together.
No platform fees. No usage limits. No per-node pricing. Your data stays in your environment. Forever.
Learn more →Performance
The Pregel runtime is rewritten in Rust and compiled to native code on every platform. Channels are SPSC ring buffers. The scheduler is lock-free. State diffs are zero-copy. The result: agents that respond before LangGraph would have started.
Numbers from the internal benchmark suite (graphs ranging 100 nodes → 10k nodes, mixed ReAct + Supervisor workloads). Higher is better for the first two; lower is better for the last two.
Architecture
A single Rust core powers PyO3 (Python), napi-rs (TypeScript), CGO (Go), and JNI (Java)
bindings. Native Rust is just cargo add agentmatic. Same API surface,
same engine, same correctness guarantees in every language.
Agentmatic vs the rest
| Dimension | Agentmatic | LangGraph | CrewAI |
|---|---|---|---|
| Compatibility | Drop-in for LangGraph | — | Different paradigm |
| LangChain tools | Bridge wrappers included | Native | Via LangChain |
| Engine | Rust (compiled) | Python | Python |
| Graph traversal | 10–15× faster | Baseline | — |
| Circuit breakers | Built-in | — | — |
| Retry + DLQ | Built-in | Platform (paid) | — |
| Distributed cluster | Built-in, self-hosted | Platform (paid) | — |
| Language SDKs | Python, TS, Go, Java, Rust | Python, JS | Python |
| Checkpointing | Memory · SQLite · Postgres · Redis · S3 | Memory · SQLite · Postgres | — |
| Visual studio | Built-in | LangGraph Studio | — |
| License | MIT | MIT | Apache-2.0 |
Full comparisons with AutoGen, Vercel AI SDK, OpenAI Agents SDK and others on the /compare hub.
From the blog
End-to-end walkthrough: build an AI code reviewer using Agentmatic's MCP client to wrap the filesystem + git MCP servers, a ReAct loop, and a custom GitHub-comment tool. Production-ready.
Read article →Three different bets on how multi-agent systems should be built — role-based crews, conversation orchestration, or explicit state graphs. The honest tradeoffs.
Read article →How checkpointing actually works in Agentmatic, why every superstep is a save point, how time-travel debugging changes development, and the HITL patterns that scale to production.
Read article →FAQ
An open-source AI agent framework. Drop-in compatible with LangGraph, runs on a Rust engine. Build agents in Python, TypeScript, Go, Java, or Rust — same API in every language.
No. It's a from-scratch Rust implementation with the same API surface. Think of it as LangGraph recompiled as native code — plus circuit breakers, distributed execution, and multi-language SDKs.
Yes. from_langchain_tools() wraps any LangChain BaseTool as an Agentmatic tool. as_langchain_runnable() wraps Agentmatic agents as LangChain Runnables — so you keep LangSmith tracing, LCEL composition, and LangServe deployment.
10–15× faster graph traversal, 70–80× faster channel throughput, and 50–80% less memory in the internal benchmark suite. Production users report 8–12× p95 latency drops on multi-agent workloads.
MIT license. No platform fees, no per-node pricing, no usage limits. Everything is self-hosted. Optional paid support is available — see /pricing.
OpenAI, Anthropic Claude, Google Gemini, AWS Bedrock, and Ollama out of the box. Any OpenAI-compatible endpoint works (Groq, Together, vLLM, local models).
Different paradigm. CrewAI and AutoGen are role-based delegation. Agentmatic is graph-based state machines — deterministic, parallel, checkpointable. See /compare for head-to-heads.
MIT licensed. Drop-in for LangGraph. Native SDKs in 5 languages. Battle-tested resilience primitives in the box.