Coware — Shared Living Specs for Multi-Agent Teams

Showcase Built for Cursor
1

The Problem

You’ve got three developers on a project. Alice is using Cursor, Bob is on Claude Code, Carol uses Copilot. They’re all building different parts of the same app — and every single AI agent is making its own assumptions about field names, types, endpoint shapes, and data flow.

You find out they disagreed at merge time.

Sound familiar? This is the #1 integration tax in AI-assisted multi-dev projects. Every agent starts from scratch, reads your codebase, guesses the contract, and writes code that works in isolation but breaks when combined.

Coware fixes this. It’s a lightweight, self-hostable spec server that keeps every AI agent aligned on shared interface definitions — before they write a single line of code.

How It Works

Coware introduces Spec-Driven Development: you define a living spec for every shared interface (API routes, data schemas, type contracts), and every agent reads it before coding.

The workflow:

  1. One-paste setup — Copy a single prompt into your AI agent. It handles everything: GitHub login, project creation, codebase scanning, and spec generation.

  2. Auto-generate specs — Your agent scans the existing codebase and produces .coware/specs/*.md files describing every shared interface it finds.

  3. Sync to server — Specs are pushed to the Coware server via POST /api/v1/specs/sync (bulk upsert). Every change is auto-versioned with full revision history.

  4. Team joins instantly — Share an 8-character invite code. Teammates paste the same prompt, join the project, and pull all specs in seconds.

  5. Agents stay aligned — Before coding, every agent pulls the latest specs. After changes, it syncs back. The spec is always the source of truth.

# What your agent does automatically:

1. Pull latest specs      →  node .coware/pull.mjs
2. Read specs             →  .coware/specs/auth.md, users.md, billing.md ...
3. Write compatible code  →  Code matches the shared contract
4. Update spec if needed  →  Edit .coware/specs/payments.md
5. Sync back              →  node .coware/sync.mjs

Key Features

  • Agent-first design — Not a human-facing UI tool. Specs live as markdown files alongside your code. Agents read and write them natively.

  • Works with any AI agent — Claude Code, Cursor, Copilot, Windsurf, or any agent that can fetch a URL. No plugins, no extensions.

  • Versioned specs — Every change creates a new revision. Full history. See who changed what, when.

  • Bulk sync — Push 50+ specs in one request. Idempotent — unchanged content doesn’t create noise.

  • Change detection — GET /api/v1/specs/changes?since=<timestamp> lets agents poll for teammate updates in real-time.

  • Offline-resilient — Specs cached locally as .coware/specs/*.md. If the server goes down, agents work with local copies and sync when it’s back.

  • GitHub OAuth — Device-flow auth designed for CLI/agent environments. No passwords, no tokens to copy.

  • Open source & self-hostable — Run it on your own infra. SQLite database, zero external dependencies.

What Makes This Different

Chat-driven dev Static docs SpecKit Coware
Specs persist across sessions No Yes Yes Yes
Multi-agent sync No No No Yes
Auto-versioned No No No Yes
Team coordination No No No Yes
Server-synced No No No Yes
Works with any AI agent — — Copilot-focused Any agent

vs. chat-driven development: Your context dies when the conversation ends. With Coware, specs persist and evolve.

vs. static documentation: Docs drift from code within days. Coware specs are updated by the agents that write the code — documentation that maintains itself.

vs. OpenSpec: OpenSpec is a great local CLI for single-developer spec management. Coware adds the multi-agent, multi-developer coordination layer — a central server that keeps every agent on the same page across your entire team.

vs. heavy project management tools: Coware is a single binary with a SQLite database. No dashboards to configure, no workflows to set up. Paste one prompt and you’re running.

Quick Start

Option 1: Use the hosted version

Paste this into Claude Code, Cursor, or any AI agent:

Use Coware Living Specs for team coordination. Read https://coware.team/llms.txt to get started.

That’s it. Your agent handles the rest.

Option 2: Self-host

git clone ■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
cd coware-skills
bun install
bun run dev

Then paste the prompt with your own server URL.

What a Spec Looks Like

# auth

## POST /api/v1/auth/login
Initiates GitHub OAuth device flow.

Response:
- session: string (UUID)
- url: string (browser URL for OAuth)

## POST /api/v1/auth/poll/:session
Poll for OAuth completion.

Response (pending):
- status: "pending"

Response (complete):
- status: "ok"  
- apikey: string (cw_<uuid>)
- user: { id, name, email, avatar }

Agents read this before implementing auth. They all produce compatible code.

Who Is This For

  • Teams using multiple AI agents — The more diverse your agent stack, the more you need Coware.

  • Projects with shared interfaces — APIs, database schemas, type definitions that multiple people touch.

  • Anyone tired of merge conflicts from AI — If your agents keep generating incompatible code, this is the fix.

Links