Writ

Give your agent a writ.

Writ turns plain business-language descriptions into portable AI agent specifications — through a friendly interview (for humans) or an MCP server (for other agents). No technical knowledge required.

What it does

Three ways to use it:

1. Interactive TUI — a non-technical user runs writ, answers 5–8 questions, and gets compiled specs.
2. CLI — writ compile, writ bundle, writ resolve operate on saved specs.
3. MCP server — writ mcp-serve exposes the whole pipeline to another agent. Agent-to-agent spec authoring.

Writ produces, from any of those three entry points:

• AGENTS.md — human-readable spec (agentskills.io format)
• claude.json — ready for Anthropic Messages API
• openai.json — ready for OpenAI Responses API
• gemini.json — ready for Google Gemini API
• oas.yaml — Open Agent Spec v1

Quickstart

pip install writ-agents
export ANTHROPIC_API_KEY=sk-ant-...
writ

That's it — the TUI launches and guides you through the interview.

CLI

writ                                      # start an interview
writ create                                # same as `writ`
writ compile spec.json --to claude -o out.json
writ compile spec.json --to agents-md --resolve -o AGENTS.md
writ bundle  spec.json -o ./dist --resolve           # emit all 5 formats
writ resolve spec.json                                # list matched connectors
writ doctor                                           # env/config diagnostics
writ config  --set-key sk-ant-...                     # save API key
writ mcp-serve                                        # run as an MCP server
writ version

MCP server — agent-to-agent mode

Writ runs as an MCP server so other agents can build specs without a human in the loop:

writ mcp-serve              # stdio transport (default — for Claude Desktop, Cursor, etc.)
writ mcp-serve --transport http

Exposed tools:

Resources: writ://catalog (connector catalog), writ://schema (Spec JSON schema).

Example Claude Desktop config:

{
  "mcpServers": {
    "writ": {
      "command": "writ",
      "args": ["mcp-serve"],
      "env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
    }
  }
}

See examples/mcp_client.py for a scripted agent-to-agent flow.

Example output

After a 6-question interview about a support triage agent, Writ produces:

# Support Triage Bot
> I help you route customer issues to the right team fast.

**Archetype:** triage

## Purpose
Automatically classify and route incoming customer support tickets.

## System Prompt

You are a support triage assistant. When a support email arrives, classify it by urgency (high/medium/low) and department (billing, technical, general). Route high-urgency tickets to the on-call lead via Slack immediately. Never close a ticket without human review.

Architecture

 Human via TUI ─┐
 CLI commands ──┼─► interview_step (core/step.py) ──► PartialSpec accumulated
 MCP server   ──┘                                               │
                                                                ▼
                                                        Connector resolution
                                                                │
                                                                ▼
                                                    5 compiled output formats

Every caller (TUI, CLI, MCP) sits on the same interview_step primitive in core/step.py. See docs/ARCHITECTURE.md for the full design.

Requirements

• Python 3.11+
• Anthropic API key

License

MIT. See LICENSE.