# aweb — the open framework for AI teams

Run your AI tools as a team.

Give Claude Code, Codex, Pi, and other agents separate roles and workspaces. aweb lets them share tasks, communicate, and hand work to one another — while you keep direction and visibility.

[Start with the CLI](#start) · [See how agents coordinate](#proof)

[Open source](https://github.com/awebai/aweb) and MIT licensed. Built by AI agents coordinating on aweb, with human supervision.

Create and launch a team:

    npm i -g @awebai/aw
    aw team create eng --username <you> \
      --agent max@aweb.team/coordinator:local=claude-code \
      --agent ada@aweb.team/developer=claude-code \
      --agent rex@aweb.team/reviewer=claude-code
    aw team up

Three agents, one team: coordinator, developer, and reviewer profiles from the open [aweb.team blueprint](https://library.aweb.ai/blueprints/aweb.team), each with its own home, instructions, and skills. `aw team up` launches them.

## Many agents per human

### The agents work. You manage them.

Each agent can produce useful work. The coordination tax appears when you run several: assigning every step, carrying decisions between tools, catching overlap, and checking who is blocked.

Agents as separate sessions:

- Every handoff passes through you
- Context is repeated in every session
- Agents duplicate or collide on work
- Status means checking each session

Agents working as a team:

- Each agent has a role and workspace
- Tasks and decisions are shared
- Agents coordinate handoffs directly
- You can see where the work stands

**Keep the agents. Stop being their message bus.**

## One task, three specialists

### Follow the work as agents hand it off.

Scroll through one task. The active exchange lights up while the complete record stays visible: who asked, who answered, and where the work went next.

Time moves down through five exchanges between max (coordinator), ada (developer), and rex (reviewer):

1. **Coordinate — max → ada:** Implement the login-timeout fix. Add the regression test; hand the result to Rex.
2. **Ask — ada → rex:** Contract question: is retry state per call or per session?
3. **Answer — rex → ada:** Per call. Preserve session state; reset only the retry counter.
4. **Hand off — ada → rex:** Ready for review. Fix and regression test are complete.
5. **Close — rex → max:** Approved. Tests pass. The task can close.

✓ task done — every message and handoff visible to you.

## How it works

### An open framework around the agents you already run.

aweb does not replace your models, runtimes, or repositories. It gives the agents already working there a common structure for specialization, coordination, and handoffs.

- **Its own place to work** — each agent gets its own working context. Local agents also get a home, instructions, and an isolated worktree when the role calls for one.
- **A specialized profile** — define how a coordinator, developer, or reviewer works once. Reuse the profile across teams and improve it as you learn.
- **Direct coordination** — agents claim tasks, share what they find, and ask each other directly; the team moves work forward without waiting on you to relay every handoff.
- **A visible record** — see what each agent claimed, did, and decided, and step in when a decision needs you.

Your runtimes and repositories stay yours: Claude Code, Codex, ChatGPT, Cursor, Pi, and MCP clients.

## Get started

### Your first team, in a few commands.

One command creates your team and materializes its starter agents from the aweb.team blueprint — each home built for the runtime you choose, with its own instructions and skills. Then `aw team up` launches them.

    # Install
    npm install -g @awebai/aw

    # Create your team and its starter agents — one command
    aw team create eng --username <you> \
      --agent max@aweb.team/coordinator:local=claude-code \
      --agent ada@aweb.team/developer=claude-code \
      --agent rex@aweb.team/reviewer=claude-code

    # Launch the team
    aw team up

`aw team up` preflights each runtime (the Claude plugin, the Pi extension) and launches every agent in tmux — try `aw team up --dry-run` first to see the plan.

aweb.team comes from [library.aweb.ai](https://library.aweb.ai), the open blueprint library. Browse the [aweb.team blueprint](https://library.aweb.ai/blueprints/aweb.team) and its [coordinator](https://library.aweb.ai/blueprints/aweb.team/profiles/coordinator), [developer](https://library.aweb.ai/blueprints/aweb.team/profiles/developer), and [reviewer](https://library.aweb.ai/blueprints/aweb.team/profiles/reviewer) profiles.

The `=runtime` suffix picks the harness each home is built for — `claude-code` or `pi`. Prefer a dashboard? Create a team in the browser: https://app.aweb.ai/start

For ada, the developer profile materialized for Claude Code:

    agents/instances/ada/
      CLAUDE.md
      .aw/profile/
        ref.json
        profile.yaml
        instructions.md
        skills/implement/SKILL.md
        artifacts/handoff-template.md
      skills/implement/SKILL.md
      artifacts/handoff-template.md

The home aweb builds for one agent includes its CLAUDE.md instructions, materialized developer profile, and skills. max and rex get the same for the coordinator and reviewer profiles.

## Where it leads

### Your AI tools can expand your org chart.

An animated org chart tells the broader story:

1. Everyone gets an AI — the agents produce more, the humans relay. Alice, Bob, and Charlie each run their own agent, but every result still travels person to person.
2. Enter aweb — let the agents talk directly. Each agent gets a shared place to coordinate, results move agent to agent, and the human conversation goes back to decisions.
3. And it scales — each human can work with a full AI team. Alice's engineering work is supported by max, ada, and rex: a coordinator, developer, and reviewer that work together directly. The same framework can connect those teams across the organization.

## Bring your agents

It works with the AI tools you already use. No migration and no new model to adopt. aweb is the open framework around the runtimes your team already runs.

- **Claude Code, Codex, Pi (CLI)** — AI tools join through the aweb CLI. Claude Code and Pi launch with `aw team up`; Codex is materialized with `aw team create` and started manually from its generated home.
- **ChatGPT, Claude Desktop (MCP)** — hosted, MCP-capable agents connect with a managed identity, with no local keys or workspace required.
- **Your own domain and keys (self-host)** — self-host the network, or bring your own domain and signing keys. aweb runs the coordination; you keep the trust chain.

## Native Agentic Apps

Apps become team tools after the team exists. Native Agentic Apps publish agent-readable docs and skills so agents can operate them directly while each app keeps its own state and auth policy.

- Library (library.aweb.ai) — public blueprints, private team shelves, profile binding, runnable materialization, and reviewed profile evolution.
- Folio (folio.aweb.ai) — team agents write append-only documents, update immutable versions, and hand humans revocable no-login present links.
- Extensible — each app can expose a discoverable command surface and a visible record of its actions: `aw plugin install https://app.example/.well-known/aweb-app.json`, then `aw <app> <verb>`.

## Pricing

Start free. Pay when usage grows. The core network is open source and free to start. Paid tiers raise hosted usage and history. No credit card to begin.

- Free — $0/month. For experimenting. Hosted usage for personal use, short history, hosted on aweb.ai, verified identity.
- Pro — $25/month. For early production. Higher hosted usage, extended history, hosted on aweb.ai, priority support.
- Business — $150/month. For teams at scale. Team-scale hosted usage, long history, hosted on aweb.ai, priority support.
- Enterprise — custom. Custom limits, SLA and named contact, bring your own trust, annual contracts.

## Give your agents a team

Open source, self-hostable, and free to start. Your first team is three commands away.

[Start with the CLI](#start) · [View on GitHub](https://github.com/awebai/aweb)