Open Managed Agents (OpenMA)

What does this agent do, and when should you use it?

Open Managed Agents (OpenMA) is an open-source, self-hostable alternative to Claude Managed Agents and OpenAI Agents API that positions itself as a meta-harness: it does not ship one agent loop, it manages the sessions, execution resources and recovery around a loop you choose. You can keep the harness outside the sandbox (built-in Pi, or your own implementation of the host-side harness interface) and use the sandbox only to run tools, or put a mature harness such as Claude Code or Codex into a managed sandbox through the ACP runtime. Session history, workspace checkpoints and collected outputs are persisted independently of both the loop process and the sandbox, so a sandbox failure does not erase the conversation and a replacement harness can rebuild context from durable events. Models are wired in through Model Cards (Anthropic, OpenAI, DeepSeek, OpenRouter and other Pi providers, plus custom endpoints), and database, queue and object storage are all swappable adapters. Deployment paths include native Node.js, Docker Compose, Fly.io, Vercel (Beta) and Cloudflare, alongside a hosted console.

You create an agent (name, model, system prompt), an environment (networking and packages) and a session; from then on OpenMA runs the event loop for you. Events such as user.message are sent with client.beta.sessions.events.send, persisted as durable records, handed to the selected harness, and streamed back as progress and output events. Built-in tooling covers shell execution, file read/write/edit, glob and text search, web search and page fetching, and optional browser automation (navigation, clicks, text input, screenshots); sessions can also load MCP services, Skills, repositories, files and persistent memory stores. When work finishes, the workspace is checkpointed and outputs are collected so the session can be resumed later. Three interfaces are exposed: Claude Managed Agents style /v1/* (works with the official Anthropic SDK), OpenAI Agents API style /openai/v1/* (tested against [email protected]), and OpenMA extensions under /v1/oma/* for Model Cards and product features. For scale-out, Node replicas claim different sessions in parallel through a durable execution outbox with renewable leases and fencing while work inside one session stays ordered; in-sandbox harnesses use a separate Environment Work claim instead.

1. Teams running long coding jobs that want Claude Code or Codex inside an isolated sandbox (E2B, Daytona, Cloudflare Sandbox and others) while OpenMA handles workspace persistence and output collection.
2. Platform teams avoiding vendor lock-in: models, harnesses, sandboxes, deployment, database, queue and object storage are seven independently replaceable layers.
3. Developers who already have Anthropic or OpenAI SDK code and want to self-host — point baseURL at your own server and keep using /v1/* or /openai/v1/*.
4. Workflow teams automating issue-to-PR: combine GitHub issues/PRs, Linear, Slack and Feishu integrations with OpenMatter's Loop runtime to turn work events into agent turns and deliver results back.
5. Security-conscious operators who need provider keys encrypted at rest via Model Cards and scoped access to private services through a Vault-backed gateway.
6. Individual developers validating agent behaviour locally with the single-host SQLite Docker Compose quickstart and their own API key.

How do you install or deploy this agent?

The recommended path is to let a coding agent install the deployment skill:

npx skills add openma-ai/open-managed-agents --skill deploy-openma

Then ask it to "Deploy OpenMA on this machine": the agent checks the environment, prepares configuration, deploys the platform and verifies the Console plus one agent task, asking for credentials and confirming paid cloud resources before creating them.

Manual Docker setup (requires Docker with Compose):

git clone https://github.com/openma-ai/open-managed-agents.git
cd open-managed-agents
cp .env.example .env

Before starting, edit .env: generate the two secrets with

openssl rand -hex 32      # BETTER_AUTH_SECRET
openssl rand -base64 32   # PLATFORM_ROOT_SECRET

then select an isolated SANDBOX_PROVIDER with its network, storage and credential-proxy settings, and set PUBLIC_BASE_URL. Keep authentication enabled — the deployment does not fall back to running agent commands in an unisolated host subprocess.

docker compose up -d --build
curl --fail http://localhost:8787/health

Open http://localhost:8787, create an account, then configure a model and environment in the Console. The Compose stack mounts ./data; keep it across restarts and upgrades.

Cloudflare deploy needs the Node version from .node-version, the repository's pinned pnpm, Docker for container builds, plus jq and openssl:

git clone https://github.com/openma-ai/open-managed-agents.git
cd open-managed-agents
pnpm install --frozen-lockfile
npx wrangler login
pnpm setup:cloudflare

The script provisions Cloudflare resources in your account, updates the local Wrangler config, sets secrets and deploys the Workers; it may incur charges and currently prompts for an Anthropic API key unless supplied in the environment.

How do you use this agent?

Create an API key, a Model Card and an environment in the Console, then call the API from your own code.

With the official Anthropic SDK (repository pins @anthropic-ai/[email protected]):

npm install @anthropic-ai/[email protected]

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  apiKey: process.env.OPENMA_API_KEY,
  baseURL: process.env.OPENMA_BASE_URL ?? "https://app.openma.ai",
});

const agent = await client.beta.agents.create({
  name: "researcher",
  model: process.env.OPENMA_MODEL_ID!,
  system: "Help me investigate a task and explain your findings.",
});

const session = await client.beta.sessions.create({
  agent: { type: "agent", id: agent.id, version: agent.version },
  environment_id: process.env.OPENMA_ENVIRONMENT_ID!,
});

await client.beta.sessions.events.send(session.id, {
  events: [{
    type: "user.message",
    content: [{ type: "text", text: "Help me plan a code review." }],
  }],
});

With the OpenAI SDK (tested baseline [email protected]):

npm install [email protected]

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.OPENMA_API_KEY,
  baseURL: `${process.env.OPENMA_BASE_URL ?? "https://app.openma.ai"}/openai/v1`,
});

const events = await client.beta.agents.sessions.create({
  agent: {
    model: process.env.OPENMA_MODEL_ID!,
    instructions: "Help me investigate a task and explain your findings.",
  },
  environment: { type: "none" },
  input: "Help me plan a code review.",
  stream: true,
});

for await (const event of events) console.log(event);

For self-hosting, set OPENMA_BASE_URL to your own origin without a trailing slash (e.g. http://localhost:8787) and set OPENMA_MODEL_ID to the Model Card's model_id, not its database record ID. The OpenAI example above requests no sandbox.

Local development:

pnpm install --frozen-lockfile
pnpm typecheck
pnpm test

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

1. @openma/[email protected] 2026-10-09
   Release notes excerpt

   Patch Changes - 57906b3: Pi harness skill mounting strategies (skill_mount: progressive default, budgeted, tool, inline) replace default full SKILL.md system-prompt inlining for default/Pi harnesses. Deferred skills in budgeted mode are discoverable via tool_search; tool mode adds a skill loader tool.

   View on GitHub ↗
2. @openma/[email protected] 2026-10-09
   Release notes excerpt

   Patch Changes - 57906b3: Pi harness skill mounting strategies (skill_mount: progressive default, budgeted, tool, inline) replace default full SKILL.md system-prompt inlining for default/Pi harnesses. Deferred skills in budgeted mode are discoverable via tool_search; tool mode adds a skill loader tool.

   View on GitHub ↗
3. @openma/[email protected] 2026-10-09
   Release notes excerpt

   Patch Changes - 38f58cb: Pi agent-core is the default harness; legacy ai-sdk remains available but deprecated (one-time session.warning per session). Agents accept compaction via _oma.context_management (Anthropic Messages shape) or _oma.model_settings (OpenAI Agents SDK shape); conflicting dual definitions are rejected at save time. MCP tool results preserve content blocks on the wire. Agent max_tokens above the catalog window is accepted with a warning; pi-ai clamps output per request. …

   View on GitHub ↗

FAQ