Documentation · Team-AI · by Andraste Labs

Quick Start Guide

Three steps to your first project. Sign in, paste your Anthropic API key, describe what you want.

1

Create an Account

Sign up for a free account and get access to your dashboard with all 8 agents.

2

Initialize Your Project

Create a new project and define your requirements using natural language.

3

Let the Agents Work

The eight specialists pick up tasks from the Project Manager's plan, write code, and hand off to the Integration Validator for build + test before opening the pull request.

Code Example

# Create a new project — one sentence is enough
curl -X POST https://team-ai.andrastelabs.com/api/projects \
  -H "Authorization: Bearer $JWT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Weather Dashboard",
    "description": "React frontend + Node.js backend using OpenWeatherMap API"
  }'

# Response — project starts in 'planning'; confirm to kick off agents
{
  "success": true,
  "project_id": "295f665a-a171-4cdd-851d-6dcf18bf3c6c",
  "status": "planning",
  "github_repo_url": "https://github.com/...",
  "contract_preview": { ... }
}

REST API · JSON See full API reference →

Core Concepts

Understand the fundamental concepts behind Team-AI's architecture.

Multi-Agent System

Eight specialist agents work collaboratively: Project Manager, UX Designer, Backend, Frontend, Mobile, QA Engineer, DevOps, and Integration Validator.

Extended Thinking

The Project Manager and QA agents use extended thinking (up to 8K reasoning tokens) to plan architecture and reason through edge cases before writing code.

Live Build Validation

The Integration Validator clones the freshly built repo, runs the actual build and tests, and triggers up to three automatic rework rounds if something fails.

Engineering details

The specific mechanisms — what makes the pipeline trustworthy when it shouldn't be.

Per-user, per-agent Claude clients

Each of the eight agents owns its own Claude client with its own httpx connection pool, tagged with agent_id. Eight parallel requests, eight separate pools — no head-of-line blocking between the PM and the Validator.

src/utils/llm_client.py · LLMClient owns its AsyncAnthropic + httpx pool

Prompt-injection sanitiser

User-supplied text is sanitised and wrapped in boundary markers before reaching the model. Injection attempts log a prompt_safety.injection_signal rather than getting free agency over the agent's instructions.

src/utils/prompt_safety.py · sanitize_user_input + wrap_user_input

Structured rework, root-cause first

When the Validator finds failures, they're typed (deterministic / hybrid / llm-judge), sorted with deterministic at the top, then dispatched per stack to the right agent — Frontend failures go to the Frontend agent, not one mega-prompt for Claude to untangle.

src/agents/integration_validator.py · _build_validation_report

GitHub-first, ownership-first

Start from an existing repo (/api/projects/from-github), push follow-up runs back to the same branch (/sync), download artifacts as a ZIP, or have the PM generate a PDR (product definition record) for the project. The output is always yours, on your GitHub.

src/api/main.py · /github/create, /sync, /from-github, /download, /generate-pdr

Additional Resources

Explore more documentation and guides

API Reference

Complete API documentation

Support

Get help from our team

Developer Documentation