Orchestrate Enterprise MCP Servers

Archestra is described as an all-in-one open-source enterprise AI platform, built on a security and observability foundation - SSO and RBAC, sandboxed code execution, Dual-LLM and Lethal Trifecta guardrails, OpenTelemetry traces, and Prometheus metrics - that the project says are first-class rather than bolted on. Point users, agents, or clients like Claude, Codex, or Cursor at one URL and the platform is meant to handle chat, LLM/MCP proxying, gateways, and orchestration behind it.

Its stated feature set covers: an internal AI assistant for non-technical users with Slack, MS Teams, and email front-ends and MCP-app support; a developer LLM and MCP portal giving one token to tools like Claude Code, Codex, and Cursor; an LLM gateway for providers such as Anthropic, OpenAI, Azure Bedrock, and DeepSeek with cost limits, virtual API keys, and dynamic model routing; an MCP gateway using OAuth and On-Behalf-Of so tools run as the actual user rather than a shared service account; an agent-to-agent (A2A) gateway for webhook-triggered agent interactions; a private MCP registry for teams to ship their own tools; an MCP orchestrator with a Kubernetes operator and self-serve environment promotion; an agent runtime with scheduled, email, and webhook triggers, sub-agent delegation, reusable skills, sandboxed code execution, and a Kubernetes-native filesystem; a RAG knowledge base wired to connectors into an existing stack; and a mini app builder.

The project positions itself against teams already running what it calls "dangerous single-tenant agents" - naming Claude Cowork, OpenClaw, and Hermes as examples - in an enterprise setting, and offers a migration kit for that case.

Use it when an organization wants a single, centrally governed entry point for chat, LLM access, and MCP tool use across teams - with SSO/RBAC, per-environment cost limits and egress policies, and built-in observability (OpenTelemetry traces, Prometheus metrics, per-team cost tracking) - rather than each team wiring its own LLM/MCP access separately.

Do not use it if a single hosted LLM or MCP integration is all that's needed, or if the added surface of a full platform - gateways, orchestrator, agent runtime, private registry - isn't wanted. The project's own release guidance says to pin an exact stable version tag or Helm chart version rather than tracking latest for production, and its docs draw a line between the free Open Core tier and licensed enterprise use.

Beyond the LLM/MCP gateways described above, it ships deterministic guardrails for tool calls, Dual-LLM verification, and Lethal Trifecta protections; identity and access via SSO (OIDC, SAML, Okta, Entra) and RBAC with role mapping and team sync, plus secrets management; per-environment egress policies and cost limits; and a Terraform provider and Helm chart for infrastructure-as-code deployment. It maintains two release tracks - stable releases (e.g. 1.3.51) carrying security patches and bug fixes for production, and beta releases (e.g. 1.4.0-beta.1) built from main for previewing upcoming features - with fixes landing on main first and then being selectively backported to the stable branch. The project states $13.5M in total funding, three Fortune-50 deployments, and 31ms latency at p95 in its own benchmarks.

Quickstart is a single Docker container:

docker pull archestra/platform:latest

docker run \
  -p 127.0.0.1:9000:9000 -p 127.0.0.1:3000:3000 \
  -e ARCHESTRA_QUICKSTART=true \
  -v /var/run/docker.sock:/var/run/docker.sock \
  -v archestra-postgres-data:/var/lib/postgresql/data \
  -v archestra-app-data:/app/data \
  archestra/platform:latest

Then open http://localhost:3000. Full Docker, Helm, and Kubernetes deployment instructions - env vars, secrets management - are covered in the project's own deployment docs; for production it recommends pinning an exact stable version or Helm chart version instead of tracking latest.

It is released under a dual license - AGPL 3.0 for the open-source core paired with a commercial Enterprise license - following an Open Core model the project describes as free for teams under 30 users.

Platform and security teams at organizations that want to give employees and agents governed access to multiple LLM providers and MCP tools through one gateway, rather than each team managing its own integrations and credentials.