Open source · Apache-2.0 · Free

The open source alternative to Claude Managed Agents.

Run agents with any model, self-hosted in your own cloud.

Get it on GitHub Read the docs
QUICK START 
git clone https://github.com/vectrix-ai/oaocd oaopnpm oao setup

Needs Node 22, pnpm, Docker, and a model provider key.

Why OAO

Almost perfect was not enough.

At Vectrix we built on Claude Managed Agents and loved the experience: agents, sandboxes, sessions, and a clean console with every turn.

Two things stopped us. It only runs Anthropic models, and it could not keep our data in Europe. So we built OAO to fix both, and made it open source.

  1. 01
    Anthropic models only

    Any model

    OpenAI, Anthropic and xAI directly, plus hundreds more through OpenRouter. Every agent picks its own.

  2. 02
    No way to keep data in Europe

    Your cloud, your region

    Deploy wherever you want, including the EU. Sessions, transcripts and files stay in your own Postgres and buckets.

Compare

Same experience. Fewer limits.

OAO next to the hosted agent platforms from Anthropic and OpenAI.

Feature OAOClaude Managed AgentsOpenAI Agents API
Open source Yes, Apache-2.0 No No
Model choice Any: OpenAI, Anthropic, xAI, OpenRouter Claude models only OpenAI models only
Self-hosting The whole platform Sandboxes only Sandboxes only
EU hosting Any region you choose, including the EU No, data is stored in the US No, US data residency only
Sandboxes Yes, on Daytona Anthropic cloud or self-hosted OpenAI hosted or self-hosted
Multi-agent delegation Yes, pinned delegate rosters Yes, subagents and workflows Yes, subagents
Price Free, no subscription. Bring your own model keys. Claude tokens + $0.08 per session-hour OpenAI tokens + container time

Based on public docs as of October 2026: Claude Managed Agents, OpenAI Agents API. All three products are in beta, OAO included.

Features

The full platform. On your infrastructure.

Built on Flue, the open agent runtime, with Postgres as the single source of truth.

model presets per agent

Any model

Use OpenAI, Anthropic and xAI directly, or hundreds of models through OpenRouter. Every agent picks its own.

postgres · leases · fencing

Durable sessions

Every run lives in Postgres under a lease. If a worker crashes, another one resumes it from the last checkpoint.

file · shell · browser

Daytona sandboxes

Each session gets a persistent Daytona workspace, with tools and network allowed per agent version.

up to 32 delegates

Multi-agent delegation

A coordinator hands work to pinned child agents. Each gets its own session, and all share one workspace.

versioned · approved tools

Skills and MCP

Publish versioned skills and connect remote MCP servers from the console. Agents only get the tools you approve.

s3-compatible

File storage

Send files to agents and back up every workspace to your own S3-compatible bucket.

http · sse · webhooks

Full REST API

Plain HTTP and Server-Sent Events, plus signed webhooks. Resume any stream from your last event.

timeline · tokens · cost

A console for every turn

Follow each session on a timeline, with every tool call, token count and cost in one view.

Open by design

Any provider. Any model. Mixed freely.

Claude Managed Agents runs Claude. OAO runs whatever you connect: frontier labs and open models alike, side by side in one platform. Give every agent the model that fits its job, and swap providers without rewriting a thing.

  1. 01
    Mix providers per agent

    Each agent version pins its own model preset. A coordinator on one provider can delegate to children on another.

  2. 02
    Open models welcome

    Llama, Qwen, DeepSeek, GLM and Mistral run through OpenRouter, next to the frontier labs.

  3. 03
    No lock-in, by construction

    Agents only see named presets, never a hardcoded model ID. Switching providers is a config change, not a migration.

SUPPORTED TODAY
  • OpenRouter hundreds of models, open ones too
  • Anthropic Claude family
  • Grok xAI models
  • OpenAI GPT family
MODELS YOU CAN MIX
anthropic/claude-opus-5openai/gpt-5.6x-ai/grok-4.6google/gemini-3.7-flashdeepseek/deepseek-v4-proqwen/qwen3.8-maxz-ai/glm-5.1meta/llama-4-scoutmistral/mistral-large-3 + hundreds more via OpenRouter
For developers

Two requests to a durable agent.

OAO speaks plain HTTP and Server-Sent Events. No proprietary protocol, no required client library. Create a session, stream its events, and pick up exactly where you left off after any disconnect.

1 · Start a session 201 right away, work durably queued 
curl -X POST "$OAO_API/v1/projects/$PROJECT_ID/sessions" \
  -H "Authorization: Bearer oao_your_secret" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{
    "agentId": "8f2b1c04-…",
    "initialMessage": "Investigate shipment SHP-1042 and summarize the delay."
  }'
2 · Follow every state change resumable from any committed position 
curl --no-buffer "$OAO_API/v1/projects/$PROJECT_ID/events" \
  -H "Accept: text/event-stream" \
  -H "Last-Event-ID: 142"

event: run.state_changed
event: tool_call.requested
event: run.completed
Read the docs View on GitHub

Every endpoint is documented at docs.oao.sh: approvals, caller-owned tools, files, skills, and the full event reference. The repo also includes a typed JavaScript SDK.

Your infrastructure, your rules

Self-host it. All of it.

OAO is an open source package you run in your own cloud, in the region you choose, including the EU. Guides for Railway and Google Cloud are in the docs.

SELF-HOST · APACHE-2.0

Your cloud, your region

The full platform, free under Apache-2.0. Your Postgres, your object storage, your provider keys. No usage telemetry and no license check, so your data stays where you run it.

github.com/vectrix-ai/oao →
WHAT YOU NEED
  • POSTGRES 17
  • A MODEL PROVIDER KEY
  • DAYTONA, FOR SANDBOXES
  • S3 STORAGE, FOR FILES
Security

Boring where it matters.

Every tenant isolated. Every sandbox scoped. Enforced by the runtime, not a policy.

Isolated by construction

Row-level security scopes every query. Org and project IDs live in every key.

Sandboxes on your terms

Sandboxed agents get their own workspace. Filesystem, shell, browser and network are allow-listed per version.

Write-only credentials

Provider keys are encrypted and can never be read back. Responses carry only a fingerprint.

Nothing leaks into logs

Secrets and auth headers never appear in logs or event streams. A platform rule, not a lint warning.

FAQ

Fair questions.

How is OAO different from Claude Managed Agents?

Same shape: agents, sandboxes, sessions, and a console that shows every turn. The difference is who holds the keys. OAO is open source, runs any model, and lives in your own cloud and region.

Which models can I use?

OpenAI, Anthropic and xAI directly, plus hundreds of models through OpenRouter, including Gemini, Llama, DeepSeek and Qwen. Each agent picks its own model preset, and a coordinator can delegate to agents on other providers. You bring your own keys; they are encrypted and write-only.

Can I keep my data in the EU?

Deploy OAO in an EU region and agent state, transcripts and files stay in your own Postgres and buckets there. Model calls go to the provider you pick, and sandboxes run wherever your Daytona target runs, so choose those with the same care. OAO makes no certification claims.

What happens when a worker crashes mid-run?

The run is a Postgres record held under a lease. When the crashed worker's lease expires, a healthy worker takes over with a higher fencing token and resumes from the last checkpoint. The stale worker can never commit anything afterwards.

Can agents call tools inside my own systems?

Yes. Your application claims a pending tool request, runs it in your environment with your credentials, and submits one result, validated against the schema pinned in the agent version. OAO never needs access to your internal systems.

What does it cost?

Nothing. OAO is free under the Apache-2.0 license, with no subscription and no license check. You pay your model provider, your cloud, and Daytona if your agents use sandboxes.

Is it ready for production?

OAO is in early beta: ready for evaluation and development, without production or support guarantees yet. Breaking changes before 1.0 are possible and announced in the changelog.

Define the agent.
OAO runs the rest.

Free and open source under Apache-2.0. No subscription.

Get it on GitHub Read the docs
QUICK START 
git clone https://github.com/vectrix-ai/oaocd oaopnpm oao setup

Needs Node 22, pnpm, Docker, and a model provider key.

Built on Flue, the open agent runtime from @flueai, and Postgres. Follow the build on YouTube.