For platform teams

An agent runtime that respects your stack.

Self-hosted on your k8s / VPC, OpenTelemetry-first, multi-language by default, MIT-licensed. Built for the team that operates the agents, not just the team that writes them.

What we don't do

• No SaaS dependency. Everything runs on your infra. No telemetry calls to our domain, no license server.
• No license maze. MIT, full stop. You can fork, modify, ship in commercial products, and we have no take-backs.
• No "community/enterprise edition" split. Every feature is in the open-source core.
• No platform tier for "production features". Circuit breakers, retry, DLQs, distributed clusters — all open source.

What we ship for platform teams

• OpenTelemetry tracing by default. Every node, every tool call, every state diff emits a span. Wire to Datadog / Honeycomb / Tempo / Jaeger out of the box.
• Distributed clusters. Coordinator/worker (gRPC) or P2P mesh (Raft consensus). Health probes, least-loaded routing, checkpoint hand-off.
• Structured logging. JSON or logfmt, per-graph-run correlation IDs.
• Health probes. /healthz, /readyz exposed by every worker.
• Tier-1 LLM providers. OpenAI, Anthropic, Bedrock, Gemini, Ollama — plus any OpenAI-compatible endpoint.
• SOC 2-ready audit log. Every state mutation, tool call, and HITL approval is recorded with cryptographic chain (optional, opt-in).

Operational model

# Deploy your agents as you'd deploy any service.
# Single-host: just the Python / Node / Go binary.
# Multi-host:  a thin Coordinator + N Workers, all gRPC.

# Worker (on each box):
agentmatic-worker --listen 0.0.0.0:9090

# Application code:
from agentmatic.cluster import ClusterConfig
config = ClusterConfig(
    topology="coordinator-worker",
    workers=["worker-1:9090", "worker-2:9090"],
    load_balancing="least-loaded",
)
agent = Agent.builder("prod").cluster(config).build()

Security posture

• No outbound calls from the framework itself (except the LLM and tool calls you configure).
• MCP tools run out-of-process — risky integrations are isolated.
• HITL interrupts for dangerous tools (filesystem writes, transfers, prod DB changes).
• Audit log with optional tamper-evident chain (Apache 2.0, opt-in).
• Reproducible builds — Rust core is checksummed; Python wheels are signed.