Skip to content

the conductor for AI agents · open source · self-hosted

npm$ npm install -g brio-orchestrator

One goal.
Every AI.

Chat with all your agents at once and let them answer together, review each other or debate. Or give Brio a goal: it plans it into tasks, sends each to the agent best at it, checks the results with another vendor, and hands you the verified outcome.

Get startedHow it worksSource on GitHub
connectors
98
connectors
chat modes
4
chat modes
for Brio itself
$0
for Brio itself
~/acme-api · mission “add a /health endpoint”done
claude-code
t1 · api contract
done
codex
t3 · /health endpoint
done
gemini
t4 · review t3
done
  1. planner → plan v1 · 4 tasks · est $1.12
  2. you → approved plan v1
  3. claude-code → t1 · contract written · $0.08
  4. claude-code → codex · handoff · api contract v1
  5. codex → t3 · tests passed · $0.31 · 2m14s
  6. gemini → t4 · review t3 · pass
  7. brio → mission done · $0.46 of $5.00

Drives 98 connectors, including

  • Claude Code
  • OpenAI Codex
  • Cursor Agent
  • Google Antigravity
  • Gemini CLI
  • Jules
  • Devin
  • Grok Build
  • Claude Managed Agents
  • OpenAI Agents API
  • Manus
  • Local models
  • any A2A agent
125missions

running at once, every one finished

492attempts

in sandboxes at the same time

136ms

API p95 under that load (target 300)

741ms

p95 from event written to your screen

Measured on one laptop with every process sharing it · the load test

01 · The problem

Every agent is good at something. None of them knows the others exist.

You already use several. Today you are the router, the messenger and the reviewer between them: you pick who gets what by habit, copy context between chats and lose half of it, find collisions at merge time, and read every diff yourself.

By handWith Brio
Choosing the agentHabit, or whoever you pay forScored per task type on cost, speed and verdicts
Handing work overCopy and paste between chatsHandoff packets with a contract per task
Two agents, one repositoryThey overwrite each otherTask leases and a worktree per attempt
Checking the resultYou read every diffTests, typecheck, another vendor's review, a judge
CostYou find out at the end of the monthA budget checked before every dispatch
What happenedScrollbackEvery action an event, in a hash-chained ledger

02 · What Brio is

The conductor, not another agent.

Brio has no model of its own. It sits above the agents you already pay for and gets the best result they can produce together, with proof, and better at it every run. Four things, together, that no single tool combines:

01

Cross-domain

Code, research, writing, data, design, media and browser work. Not only repositories.
02

Verified outcomes

Tests, typecheck and a review by another vendor gate every result before you see it.
03

Learned routing

A live scoreboard of agent × task type × cost × speed decides who gets each task.
04

Governance built in

Budgets, approvals, an audit ledger and your own keys in an encrypted vault.

Read the product requirements for the full reasoning, or the architecture for how it is built.

03 · How it works

Plan, route, coordinate, verify, learn, govern.

One mission, six steps, every one of them an event you can watch and audit.

  1. 01

    Plan

    tasks + contracts

  2. 02

    Route

    best agent per task

  3. 03

    Coordinate

    leases · handoffs

  4. 04

    Verify

    checks · review

  5. 05

    Learn

    the scoreboard

  6. 06

    Govern

    budgets · approvals

every verdict goes back into routing: the next mission picks agents on what this one learned
  1. 01

    Plan

    Your goal becomes a graph of tasks. Each has a contract: what goes in, what comes out, how it is checked. Edit any of it before you approve.

    plan v1 · 5 tasks · est $1.84 · awaiting approval
  2. 02

    Route

    Each task goes to the agent best placed for it, across vendors, by what it has won before at that kind of task, its price and speed.

    task 3 code.feature → codex (0.82 win rate · $0.31)
  3. 03

    Coordinate

    Agents hold leases on their tasks, ask each other questions through a mailbox, and hand work on as structured packets. You see all of it live.

    claude-code → codex · handoff · api contract v2
  4. 04

    Verify

    Results face tests, typecheck, scope checks and a review by a different vendor before a person sees them. Failures go back with the reason.

    task 3 · tests passed · review by gemini: pass
  5. 05

    Learn

    Every verdict updates the scoreboard: who wins at which task, at what cost and speed. The next mission routes on it.

    scoreboard · code.feature · codex ↑ 0.84
  6. 06

    Govern

    Budgets are checked before dispatch; pushes, deploys and sends wait for an approval; every action lands in an audit ledger.

    approval needed · open PR #42 · owner
The full walk-through →

04 · Two ways to work

Ask a question, or hand over a goal.

Chat when you want an answer now. Run a mission when the work needs a plan, a repository, sandboxes and checks. A chat becomes a mission in one click, with the conversation as context.

Chat: Ask anything: two agents answer together and one merges; each one's own answer is a click away.

Chat

Talk to all your agents the way you talk to one assistant.

  • the best agent for the question answers, or several build the reply
  • follow-ups carry the conversation; replies arrive live
  • each reply shows who worked on it, the cost and each agent's own answer
  • rate a reply and the scoreboard learns from it
The live board: One mission: its tasks, attempts, checks, cost and every event, live.

Missions

Give a goal; get a planned, routed, checked and merged result.

  • a plan of tasks with contracts, which you edit and approve
  • a sandbox and a git worktree for every attempt
  • a live board of lanes, handoffs, checks and cost
  • a report and, on a repository, a pull request that waits for you

05 · Many agents, one answer

Let them answer together, review each other, or debate.

Pick the agents above the message box and a mode. With one agent, Auto works; with two or more, the other three unlock. Missions have seven execution modes, from best-of-n to cascades.

youAbest at it

Auto

1 agent

The agent best at that kind of question answers, chosen by the scoreboard.

mode: single

youABM

Together

2+ agents

Each agent you picked answers, then one merges them into a single reply.

mode: ensemble

youABchecks · fixes

Review

2+ agents

One answers; another checks it and asks for fixes until it holds up.

mode: maker_checker

youABJ

Debate

2+ agents

They argue it out, and a judge decides and writes the answer.

mode: debate

Diversity is the point: in multi-agent modes Brio prefers agents from different vendors, because their mistakes correlate less. All execution modes.

06 · Bring your agents

98 connectors across 13 categories.

Model APIs, coding CLIs, hosted agents, browser agents, media models, research tools, local models and any A2A agent. Model names are data: lists come live from each vendor, never hard-coded.

Frontier model APIs

10

OpenAI API · Claude API · Gemini API · Grok API · Meta Model API (Muse Spark) · Mistral API

Open-weight & regional models

8

DeepSeek API · Qwen (Model Studio) · Kimi API · GLM API · MiniMax API · Sarvam AI API

Gateways & clouds

9

OpenRouter · LiteLLM · Amazon Bedrock · Microsoft Foundry · Vertex AI Model Garden · Groq · Cerebras

Coding agents

16

Claude Code · OpenAI Codex · Cursor Agent · Google Antigravity · Gemini CLI · Jules

Hosted agent runtimes

5

Claude Managed Agents · OpenAI Agents API · Gemini Managed Agents · Bedrock AgentCore · Foundry Agent Service

Computer-use & browser agents

8

Claude computer use · Amazon Nova Act · Browser Use · Browserbase · Steel · Hyperbrowser · Skyvern · Perplexity Comet

Research & search

7

Perplexity Agent API · OpenAI search & deep research · Gemini Deep Research · Grok live search · Exa · Tavily · Parallel · Brave Search API · You.com

General autonomous agents

4

Manus · Genspark · Kimi Agent · MiniMax Agent · OpenClaw · OpenManus

Image, video, voice & music

10

OpenAI image & voice · Google Veo · Nano Banana · Black Forest Labs Flux · Runway · ElevenLabs · Luma · Kling

Design & workspace tools

6

Figma · Canva · Claude Design · Claude artifacts · Notion · Google Workspace · Microsoft 365 · Gamma · Google Stitch

Enterprise agent platforms

4

Microsoft Copilot Studio · Gemini Enterprise (Agentspace) · Salesforce Agentforce · ServiceNow · SAP Joule · Workday

Protocols & frameworks

8

A2A protocol · Model Context Protocol · Agent Client Protocol (ACP) · OpenAI Agents SDK · Claude Agent SDK · Google ADK

Automation & app hubs

3

Composio · Zapier AI Actions / MCP · n8n · Make · Pipedream

Anything else

Any A2A agent or MCP server you add, local models through Ollama or vLLM, and your own adapter for whatever comes next.

Works with what you have

Brio adapts to how many agents you connect, and says plainly what it could not do.

  • 0

    No agents yet

    Brio plans the mission and shows what each task needs, then tells you what to connect.

  • 1

    One agent

    Everything runs on it. Reviews say plainly they come from the same vendor, and multi-agent modes say they ran single.

  • 2+

    Two or more vendors

    Brio routes each task to whoever is best at it, has another vendor review the work, and learns from every verdict.

07 · Verification

Nothing reaches you unchecked.

Each task's contract names its checks. Hard checks must pass; a failure goes back to an agent with the reason, not to you. The reviewer is never the maker's vendor.

  • teststhe project's own tests, in the sandbox
  • typecheckthe compiler, not the agent's word
  • diff_scopechanges stay inside the task's files
  • reviewa second vendor reads the diff
  • judgescores the result against the contract
  • injection_scantext aimed at the next agent is held for you

# a task that fails, then passes

codex → t3 · tests 2 failed

brio → t3 · revise 1/2 · failures sent back

codex → t3 · tests passed

gemini → review t3 · pass

brio → t3 done · handoff written

Two failed revisions reroute the task to the next-best agent; a third failure proposes a new plan. Every result arrives with its evidence: test output, the typecheck, the reviewer's reasons and the diff.

08 · It learns

Routing on evidence, not habit.

Every verdict updates a scoreboard of who wins at which kind of task, at what cost and speed. The router samples from it, weighs price and time by the mission's priority, and writes down why it chose.

task 3 · code.tests · priority: balanced

  • codex← picked0.82 pass · $0.31 · 2m14s · n=41
  • claude-code0.79 pass · $0.42 · 3m02s · n=37
  • gemini0.64 pass · $0.09 · 1m40s · n=12
rationale · codex: 0.82 pass rate on code.tests (n=41), $0.31 avg, beats claude-code on cost at similar quality

Scores are kept per task type: 38 types across code, design, research, writing, data, media and browser work. Pass rates are Beta posteriors, head-to-head results are Elo, and cost and latency are moving averages.

Scores decay slowly, so a new model can climb; a new model of a known agent starts from its predecessor's record. Your thumbs up and down count too.

Optionally, and only if you opt in both ways, workspaces pool anonymised outcomes so a new agent does not start from nothing.

How the scoreboard works →

09 · Your machines, your data

Your keys, your machines, your rules.

No account with us and no cloud of ours. Nothing leaves your machines except the calls to the vendors you connected, with your keys.

  1. You

    • the web app
    • brio CLI
    • Claude Code, Cursor (MCP)
    • any A2A agent

  2. Your Brio

    • API and orchestrator
    • Postgres
    • the key vault
    • the audit ledger

  3. Your runners

    • a Docker sandbox per attempt
    • or your laptop, under your sign-ins

  4. Vendors you chose

    • called with your keys
    • billed to you directly

Your keys

Bring your own vendor keys. They are envelope-encrypted in a vault and never shown again, not even to you.

A sandbox per attempt

Code runs in its own container with a scoped worktree, no capabilities, CPU, memory and time limits, and a network allowlist.

Approvals

Irreversible actions (push, deploy, send, purchase, delete) wait for a person, by policy.

Budgets

Monthly limits per workspace and per agent, checked before every dispatch, not after the bill.

A ledger

Every agent action is an event. The audit ledger is hash-chained, and owners can verify the chain.

Official APIs only

No scraping of consumer chat apps and no reuse of anyone's session tokens: official APIs, CLIs, SDKs, MCP and A2A.
The security model →The security review

10 · See it run

The real app, not a mock-up.

6 screens from a chat and a mission, run against Brio's scripted test vendor so it is the same every time. With your agents, the board fills with theirs.

Chat: Ask anything: two agents answer together and one merges; each one's own answer is a click away.
Chat · Ask anything: two agents answer together and one merges; each one's own answer is a click away.
The live board: One mission: its tasks, attempts, checks, cost and every event, live.
The live board · One mission: its tasks, attempts, checks, cost and every event, live.
The plan: Tasks with contracts, the agent each would go to and the estimated cost, before you approve.
The plan · Tasks with contracts, the agent each would go to and the estimated cost, before you approve.
Connectors: Everything Brio can drive, with your own keys or your own CLI sign-ins.
Connectors · Everything Brio can drive, with your own keys or your own CLI sign-ins.
Scoreboard: Who wins at which kind of task, at what cost and speed.
Scoreboard · Who wins at which kind of task, at what cost and speed.
Governance: Budgets, policies, usage and the hash-chained audit ledger.
Governance · Budgets, policies, usage and the hash-chained audit ledger.

11 · Ways in

Use it from wherever you already work.

The web app is one door. Your terminal, your editor's agent, other A2A agents and your own code are the others.

Ways in
The web appChat, missions, the live board, plans, inbox, scoreboard, budgets
Your terminalbrio run "your goal" · brio board
Claude Code, Cursor, CodexBrio's MCP server: delegate a goal, get the verified report
Any A2A agentBrio is an A2A agent: /.well-known/agent-card.json
Your own codeThe REST API with an API token
Your own machinebrio runner start: tasks run under your own CLI sign-ins
The brio CLI: npm install -g brio-orchestrator
brio loginsign in from the terminal (a code you approve in the browser)
brio run "<goal>"plan a mission; --watch follows it, --quick starts small ones at once
brio plan <mission>edit a plan: pin agents, gate tasks, reorder
brio board <mission>the live board in your terminal
brio approvesay yes (or --deny) to what waits for you
brio runner registermake this machine a runner
brio runner startrun your tasks here, under your own CLI sign-ins
brio mcpBrio's MCP tools over stdio

12 · Host it yourself

Running on your own machine in about fifteen minutes.

Brio is self-hosted: there is no Brio cloud, no account with us and no domain to buy. Pick a machine; Docker is all it needs.

A cloud server

always on

Any Ubuntu server. Free for students: Azure for Students includes a small 1 GB server for 750 hours a month, and Brio's containers measured about 510 MB idle and 560 MB at peak on ready-made images. A free cloudapp.azure.com or duckdns.org name gets real HTTPS on its own.

1 GB RAM + swap · 64 GB disk (ready-made images)

Building the images on the server instead needs 2 CPUs · 4 GB

Deploy on a server →

Your own computer

free

Docker Desktop on your Mac or PC. Share it with a free Tailscale Funnel address when others need it.

Docker · 8 GB RAM · 30 GB free

Set up on your computer →

Kubernetes

teams

A Helm chart: web pods that scale, one worker, runners each with their own Docker engine.

Kubernetes 1.27+ · Postgres with pgvector

Install with Helm →
  1. $curl -fsSL https://get.docker.com | sudo sh && sudo usermod -aG docker $USER
  2. $git clone https://github.com/Ivan825/brio_.git brio && cd brio/deploy/compose
  3. $./setup.sh https://ADDRESS --prebuilt
  4. $nano .env # BRIO_SITE_ADDRESS, ports 80/443, BRIO_SIGNUP_ALLOW=you@example.com, BRIO_RUNNER_CONCURRENCY=1
  5. $docker compose pull && docker compose up -d --no-build
Deploy on a serverSelf-hosting guideSet up without a domain

13 · Questions

Questions people ask first.

Is Brio another AI agent?
No. Brio is the conductor: it plans, routes, coordinates, checks and governs the agents you already use. It has no model of its own, and it works with any of the connectors in its registry, plus any A2A agent.
Do I need several paid subscriptions?
No. One agent is enough to start: Brio plans, runs and checks everything with it, and says honestly when a review came from the same vendor. Free API tiers (Gemini, Groq and others) and local models work too. A second vendor unlocks cross-vendor review and the multi-agent chat modes.
Can I use my ChatGPT or Claude subscription?
Not through their chat apps: Brio never automates consumer chat sites or reuses their sign-in tokens. On your own machine, the local runner can run a vendor's official CLI under your own sign-in where the vendor's terms allow it (Antigravity; Claude Code for your own work).
Who sees my code and my keys?
You and the vendors you connect. Brio is self-hosted: there is no Brio cloud and no account with us. Keys are envelope-encrypted in your vault, decrypted only for an attempt, and redacted from every log and event.
What does it cost?
Brio is free and open source under AGPL-3.0. You pay the vendors directly for what your agents use, with your own keys, and Brio checks a budget before every dispatch. A server can be free too: Azure for Students includes a small one.
What if an agent's work is wrong?
It goes back. Each task's contract names its checks; a failed hard check sends the work back with the reason (up to twice), then to the next-best agent, then to a re-plan. Nothing reaches you without its evidence.
Is it only for code?
No. Tasks cover code, design and contracts, research, writing, data, media and browser work, each with its own checks. A mission can mix them: a feature, its docs, a hero image and a launch email.
Can I keep using Claude Code or Cursor?
Yes. Add Brio's MCP server to them and they can delegate a goal to Brio and get the verified report back, or ask for a review from another vendor.
What happens to my repository?
Every attempt works in its own git worktree on its own branch. Verified work is merged into an integration branch, and pushing it or opening a pull request waits for your approval.
How do I run it?
Docker Compose on any machine: your laptop, a small cloud server with ready-made images, or Kubernetes with the Helm chart. The guides take about fifteen minutes and need no domain.

Brio's own build was planned in one agent and built in another, with a handoff packet between them.

Every AI, one ensemble.

Stop being the messenger between your agents. Give Brio the goal.

Get startedDeploy on a free serverSource on GitHub