⌜ AGPL-3.0 LICENSE ⌟

Self-hosted

Archestra is open source, and that is core to its DNA. Maybe it's because Archestra was built by Grafana alumni, or maybe it's because we believe the best PoC is the one you can run yourself.

Explore deployment options
$ 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# Open the workspace, sign in as admin@example.com / password$ http://localhost:3000

Archestra is battle-tested at scale, running in production at some of the largest companies in the world.

Fortune 50Multiple deployments,
largest one: 75,000 users

OpenAPPA AI Guardrails

OpenAPPA is a frontier deterministic guardrail, built by the Archestra research team, that guarantees agents cannot exfiltrate data.

Unlike non-deterministic or "LLM as a judge" guardrails, OpenAPPA can be used in finance, healthcare, legal, and other fields where agents may have access to confidential information.

OpenAPPA is open source under the MIT license, and its whitepaper was accepted to NeurIPS 2026.

Prevented data leaks

  • OpenAPPA100%
  • Claude Auto mode90%
  • FIDES (Microsoft)69%

Task completion

  • OpenAPPA89%
  • Claude Auto mode90%
  • FIDES (Microsoft)41%

Archestra applies guardrails across the whole organisation

  1. 01

    Policy lives in Git

    Every change is a pull request in your own GitHub or GitLab, with review and history.

  2. 02

    Tested on golden trajectories

    Sessions that must stay blocked or allowed are replayed, so a change cannot quietly weaken the policy.

  3. 03

    Rolled out at your pace

    Draft and test with enforcement off. Switch it on when the policy is ready.

  4. 04

    Owned by the teams

    Whoever runs a tool maintains its rules. The platform team keeps the review.

  5. 05

    Impact you can see

    Dashboards show how often agents are blocked, how they recover, and what the checks cost.

  6. 06

    Agents close the loop

    A blocked agent reports why. Archestra turns the report into a proposed fix.

We were looking for an infrastructure layer to scale and secure our internal agents. Archestra stood out for its security-first mindset, open-source nature, and a deployment experience that just works.

Santiago BarclayHead of Cybersecurity at Akua

Governance

Governance in Archestra starts with deeply engineered support for existing agents.

This support enables detailed auditing of agent actions (even when sub-agents are used), granular data flow tracing, secret management, and cost control.

It also opens up the ability to provide an agent with things like a corporate skills library, access to MCP servers, a unified system prompt, and even workload migration from a local machine to the cloud.

Connect an agent

⌜ CODING AGENTS ⌟

  • Claude Code
  • Codex CLI
  • Cursor
  • Copilot CLI
  • OpenCode
  • Kilo Code
  • Kimi Code
  • Grok Build
  • Pi
  • Amp
  • Antigravity
  • Droid
  • Kiro

⌜ ASSISTANTS AND AUTOMATION ⌟

  • Claude Desktop
  • ChatGPT
  • n8n
  • Hermes Agent
  • OpenClaw

⌜ LLM GATEWAY ⌟

LLM Gateway holds inference keys

It usually starts with a prototype. Someone gets an OpenAI key, builds a tool, and then asks how to distribute that key to everybody who wants to use it. Later there are endpoints in several clouds, and no answer to how much is being used or by whom.

People sign in as themselves

An app or service gets a virtual key. A person can use a personal virtual key, an OAuth sign-in, or the Okta or Entra ID token they already hold.

Azure AI Foundry, Bedrock, and models on your own GPUs

Each provider has its own route that speaks the provider’s own API, so a client that works with the provider works with the gateway after a base URL change.

  • Virtual key support-triage-prod, arch_042893d7c…. Providers: Anthropic, AWS Bedrock. Budget: Customer Support, $1,200/month cap.
  • Virtual key ci-code-review, arch_aded497f7…. Providers: Azure AI Foundry, OpenAI. Budget: Platform, $1,500/month cap.
  • Virtual key finance-close-bot, arch_edb34656e…. Providers: AWS Bedrock. Budget: Finance Ops, $400/month cap.

⌜ MODEL POOLS ⌟

Several models behind one model name

  1. 01
    Pool member in Order 1: provider key Azure OpenAI East US, model gpt-5.6, weight 3, 75% of traffic, switched on.

    Load balancing by order and weight

    Members in the same order share traffic by weight; order 2 answers only when every member of order 1 has failed.

  2. 02
    Session log rows: a GPT-5.6 call via Azure OpenAI East US at 11:09:43 marked 429 · rate limited, then a call via Azure OpenAI Sweden at 11:09:47 marked Retry, between two Snowflake tool calls.

    What moves a request to the next member

    Before the answer starts, a rate limit, a 5xx, a connection failure, rejected credentials or exhausted credit moves the request on, within the pool’s retries.

  3. 03
    Auto router tiers: Fast, short, simple sessions, on Azure OpenAI East; Normal, most sessions, and every session until the first plan, on Azure OpenAI East; Smart, the hardest sessions, on Anthropic.

    Auto routing

    An auto router is a model name, auto or auto:<name>, with a fast, a normal and a smart model behind it.

⌜ COST CONTROL ⌟

Granular token spends

A team can be told it spent $20,000 on tokens last month, and that is often all anyone knows. Finding out which people and which agents were behind the number takes a round of asking.

Costs table grouped by person, last 30 days, with Total and Share columns: Marcus Webb $1.4k, 27%; Sofia Lindqvist $1.1k, 22%; Ben Carter $490.93, 9.5%; Aiko Tanaka $486.61, 9.4%; Tomas Rivera $424.02, 8.2%; Priya Raman $287.66, 5.6%; Hannah Brooks $248.51, 4.8%.
Costs summary card "Subscription usage estimate": $1.7k est., up 5%, "Usage at API rates, not a bill", with a dashed 30-day trend line.Usage column of the Limits table for two team spend limits: an orange bar at $1,205.90 of $1,500 (80.4%) and a red bar at $941.34 of $1,000 (94.1%), both "Calendar month · Resets Dec 1".

Spend broken down the way people ask about it

The cost explorer shows billed spend, requests, tokens or tool calls, grouped by person, team, agent, model, provider, credential or MCP server, and split by a second dimension, for example which models each team used.

Limits per person, team, agent, key and environment

A spend limit applies to the organization, a team, a user, an agent, a virtual key, an OAuth client or an environment, over a rolling window from one hour to one month or a calendar day, week or month.

Subscription seats counted separately

Calls made on a Claude or ChatGPT subscription are recorded at $0 billed, next to an estimate of what the same tokens would have cost at list API prices.

⌜ MCP GATEWAY ⌟

MCP servers hosted in your Kubernetes

Your engineers already know how to write an MCP server. What they lack is a place to host it and a way to secure it, so each team runs its own endpoints with its own tokens, and nobody can list which people have connected which tools.

⌜ MCP REGISTRY ⌟
MCP Registry rows: Grafana, Jira, and meeting-room-booking, an in-house server whose green dot shows it is running in the cluster.
⌜ MCP GATEWAYS ⌟
MCP Gateways rows with their tool counts: Engineering Gateway 53 selected, Support Gateway 42 selected, Finance Gateway 41 selected.
  • Self-hosted servers run as pods in your cluster

    Give Archestra a command or a Docker image. It deploys the server on Kubernetes with one pod per install and restarts it.

  • One gateway per team, with the tools you pick

    The same Slack or Jira server can sit behind a read-only gateway for one team and a read/write gateway for another.

  • Each call runs as the right account in the system behind it

    A Jira tool can resolve at call time to the caller’s own Jira account, so the ticket shows the person who asked.

⌜ AGENT RUNTIME ⌟

Your self-hosted cloud agents

Most coding agents now offer cloud runs, each in its vendor’s cloud and for that agent only. Agent Runtime runs Claude Code, Codex, OpenCode or your own image in your Kubernetes cluster, behind the LLM Gateway and the MCP Gateway.

Create Agent screen in Archestra, “Give it work from” bar: Archestra chat, Slack, Microsoft Teams, Telegram, Jira, email, handoff from a local coding agent, other agents.
Agent choice in Archestra: Claude Code. Anthropic's coding agent. Runs on each person's Claude Pro or Max, or a company Anthropic key.Agent choice in Archestra: Codex. OpenAI's coding agent. Runs on each person's ChatGPT subscription.

One runtime for every coding agent

Claude Code, Codex, OpenCode and custom images start from chat, Slack, Teams, email or a laptop handoff, and each agent has an idle timeout, a maximum duration and a metered LLM budget.

Containers in your Kubernetes cluster

Each run is a pod with its own persistent volume in your cluster, and its files stay there between follow-ups.

Control over what the agent can reach

A run gets a temporary virtual key while provider keys stay in Archestra, uses only its agent’s tools, and calls them with the permissions of the person who started it.

Network policy per agent

An environment’s egress policy can block all outbound traffic or allow only listed domains and CIDRs. The policy takes effect only if your cluster’s network plugin enforces it.

⌜ MCP SERVERS, SKILLS AND AGENTS ⌟

Governance and promotion pipeline

A platform team cannot build and review every MCP server, skill and agent a company asks for. When the approved route is slow, each team ends up running a few of its own that nobody else can see or vouch for.

MCP Registry
Admins and editors add a server once, and everyone it is shared with clicks Install.
Skills
Any member writes or imports a skill, and coding agents install it from one marketplace URL.
Agents
Any member builds an agent, and a team given Can use runs it without changing it.

Promotion pipeline

  1. 01

    Personal

    A new skill or agent can be used by its owner only.

    “Who can use it” block of a skill with the chip Personal: one row, Marcus Webb (you), Full access.
  2. 02

    Team-wide

    The owner adds a team under “Who can use it”, without asking an admin.

    “Who can use it” block with the chip Team-wide: Marcus Webb (you), Full access; Platform team, Can use.
  3. 03

    Org-wide

    Adding Everyone in the organization takes a permission that admins hold.

    “Who can use it” block with the chip Org-wide: Marcus Webb, Full access; Platform team, Can use; Everyone in the organization, Can use.
  4. 04

    Production

    Deploying into a restricted environment takes Can use on that environment.

    Environment menu of an agent: Default; Staging, “Pre-production. Open to every builder for testing agents, servers and skills.”; Production greyed out, “This environment is restricted. Ask an administrator for access to it to deploy here.”

AI world is moving fast. We know.

We invest in cutting-edge technology to bring it to our customers. A good example is our work on ID-JAG for MCP, together with Okta and Anthropic. For some time, Archestra was the only client supporting ID-JAG, and one of the few platforms to support MCP Apps. Nowadays, it's a standard feature for all other clients.

Joey Orlando of Archestra presenting ID-JAG at AGNTCon + MCPCon Europe
ID-JAG: Connecting agents to work appsAGNTCon + MCPCon Europe · Archestra, Okta, Anthropic

⌜ LOGS AND AUDIT ⌟

A log of who called which model and tool

Platform and security teams ask for one interface that shows who is doing what. Before a gateway, model calls sit in each provider’s console and tool calls in each SaaS admin page.

Three rows of the MCP Gateway log. search_pages, Leila Haddad on Confluence, through Support Gateway, own connection via user token, success, 1.72s. run_query and describe_table, Sofia Lindqvist on Snowflake, through the Revenue Analyst agent, own connection via session, success, 5.80s and 1.20s.
  • How a call gets a name

    A row names a person when the call used a personal virtual key, a user OAuth token or an identity provider token.

  • Your Grafana, Splunk or SIEM

    Archestra exposes Prometheus metrics and sends OpenTelemetry traces and backend logs to your collector.

  • Keep prompts out of the logs

    In metadata-only mode a row still has the agent, the model or tool, the account used, tokens, cost and whether the call failed.

⌜ ONE MORE THING... ⌟

Archestra is amazing for non-developers

Someone in finance opens the chat and asks a question. The assistant already reaches their mail, calendar and company documents, because the platform team connected those once for everyone. It also builds small apps they can share.

  • It answers as that person

    With per-person credentials, each tool call runs as the person asking. The mail server sees them, and document search returns only what they can open in the source.

  • An app from one request

    Ask for a tracker or a dashboard. The agent builds it, runs it in the chat, and keeps every version so a change can be undone.

  • A shared app carries no access

    It runs in a browser sandbox, holds no tokens, and calls tools with each viewer’s permissions, not the author’s.

⌜ CHAT ⌟

Archestra chat. Question: What's my 2pm, and which email should I read before it? Four completed tool calls, two of them to the mail and calendar server. Answer: the 2pm is the Kestrel Freight renewal review in Room Harbor 3, and the email to read first is Daniel Okafor's revised quote of $184,000.The same chat. Request: Build me a small app, a tracker for open vendor reviews for Finance Ops. Tool calls: Build App and Vendor Review Tracker. Answer: I built the Vendor Review Tracker. It passed validation.

⌜ THE APP FROM THAT CHAT ⌟

Vendor Review Tracker app, open vendor reviews for Finance Ops: a form to add a review and three rows. Kestrel Freight $184,000, Daniel Okafor, in review. Alder Packaging $92,500, Hannah Brooks, open. Brightline Cloud $376,800, Daniel Okafor, open.

Security

After the recent series of breaches and vulnerabilities across the industry, we took additional steps to protect our customers.

  • 01

    Where does our data go?

    Archestra is self-hosted only: a Helm chart in your Kubernetes cluster (or one Docker container to try it), with conversations and logs in your own PostgreSQL. We do not host it and we have no access to your data. Telemetry is optional and can be disabled.

    Deployment
  • 02

    SOC 2 and ISO 42001

    We are working towards SOC 2 Type II and ISO 42001.

  • 03

    Ongoing penetration testing

    Three independent penetration-testing firms are contracted to assess Archestra. The scope is application security: unauthorized access and dependency review. Testing is moving to a monthly retainer, so every assessment covers the current codebase.

  • 04

    Build and release

    Every release image is scanned for vulnerabilities on both amd64 and arm64, and any fixable Critical or High CVE blocks the release. Each release requires approval from a second engineer, so a single compromised workstation cannot publish a release.

  • 05

    Dependency patching

    Since August 2025, the median time from a dependency alert to its fix is 19 hours with an objective to reduce this time to a sub-1 hour. About six stable releases a week over the past year. Enterprise customers receive security patches for LTS versions for 12 months after release, we also provide SLA for critical security patches.

    Releases
  • 06

    OSS contributions

    Since September 22, 2026, the open source repository no longer accepts external pull requests: contributions arrive as issues, and maintainers implement them.

    How we take contributions
  • 07

    Code scanning with Mythos

    Archestra is a part of anthropics/oss-scanner initiative, which uses Mythos to scan the codebase for vulnerabilities.

  • 08

    Code scanning with Daybreak

    Archestra engineering team is using frontier security models from OpenAI to scan the codebase for vulnerabilities.

  • 09

    And beyond

    We can’t disclose all the security measures we take, some of which are going far beyond the industry standard, but we are happy to discuss them with your security team, and support with evidence.