enwrit 0.4.7

The communication layer for AI agents. Route context between repos, devices, and tools.

writ

Better instructions. Connected agents.

The quality and communication layer for AI coding agents. Lint your instructions, compose context across tools, and connect agents across repos and devices.

pip install enwrit
writ lint CLAUDE.md               # Instant quality score (0-100) for any instruction
writ init --template fullstack    # Bootstrap agent team in your repo
writ use architect                # Compose + write to your IDE's native files

Try the live lint demo -- paste any instruction, get an instant quality score. Or browse 50+ instructions on the Hub.

---

Lint Your Instructions

97% of AI instructions have quality defects. writ lint catches them.

writ lint .cursor/rules/my-rule.mdc
# Score: 34 / 100 (D)
# - 6 instances of vague language ("try to", "consider", "if possible")
# - No verification commands (agents can't check their own work)
# - No code examples
# Suggestions:
#   Replace vague phrases with imperative commands
#   Add backtick-wrapped test/build/lint commands (2-3x quality impact)
#   Add 1-3 code examples showing desired patterns

Scores 0-100 across 6 dimensions: Clarity, Verification, Coverage, Brevity, Structure, Examples. Works on any .md, .mdc, .txt, or YAML instruction file -- no writ init required.

writ lint CLAUDE.md                     # Score any file (ML-powered, local, free)
writ lint my-agent                      # Score a managed instruction
writ lint AGENTS.md --deep              # AI-powered analysis (Gemini, via enwrit.com)
writ lint AGENTS.md --deep-local        # Local AI analysis (writ-lint-0.8B, GPU-accelerated)
writ lint rules.mdc --json              # Machine-readable output for CI
writ lint --ci --min-score 60           # Exit 1 if score too low (CI gate)

Try it in your browser -- paste any instruction, get an instant score.

---

30-Second Quickstart

# 1. Install
pip install enwrit

# 2. Initialize in any repo (auto-detects your stack)
writ init

# 3. Install a pre-built instruction from the Hub
writ install verification-loop
writ install code-review-agent

# 4. Activate -- writes to your IDE's native files
writ use verification-loop
writ use code-review-agent

That's it. Your IDE now has battle-tested instructions -- no copy-paste, no manual conversion.

What Makes writ Different

Capability	What it means
Instruction linting	6-dimension quality scoring (0-100) for any AI instruction file. Code-based + AI-powered.
Context composition	Layer project + team + agent + handoff context into one coherent instruction set
9 output formats	Write once, export to Cursor .mdc, CLAUDE.md, AGENTS.md, Copilot, Windsurf, Codex, Kiro, Agent Cards
Agent communication	Structured conversations between agents across repos and devices
Personal library + cloud sync	writ save → writ load on any device. Your agents follow you.
Hub with 50+ instructions	Rules, agents, and autonomous programs. writ install <name>
MCP server	One line in your config and any agent can search/install from the Hub

How It Works

writ writes to native IDE files -- it does NOT call LLM APIs.

Tool	writ writes to
Cursor	.cursor/rules/writ-*.mdc
Claude Code	CLAUDE.md
AGENTS.md	AGENTS.md
GitHub Copilot	.github/copilot-instructions.md
Windsurf	.windsurfrules
Codex / Kiro	AGENTS.md

When you run writ use reviewer, the tool composes all relevant context and writes it directly into the files your IDE already reads.

Context Composition

The core innovation. Each agent's context is composed from 4 layers:

Layer 4: Handoff context      ← Output from another agent
Layer 3: Agent's instructions  ← The agent's own role
Layer 2: Inherited context     ← From parent agents
Layer 1: Project context       ← Auto-detected (languages, frameworks, structure)

writ use implementer --with architect    # Compose with architect's context
writ compose reviewer --with architect   # Preview before writing

MCP Server (One-Line Setup)

Give any MCP-compatible agent access to the Hub -- no CLI needed:

{
  "mcpServers": {
    "writ": {
      "command": "uvx",
      "args": ["enwrit", "mcp", "serve"]
    }
  }
}

This exposes 22 tools: search the Hub, install instructions, compose context, read files, start agent conversations, and more. Run writ mcp serve --help for the full list.

Hub: Browse & Install

The enwrit Hub has 50+ curated instructions across three tiers:

• Rules -- passive context that shapes agent behavior (verification loops, commit hygiene, no-secrets)
• Agents -- on-invocation workers (code review, git commit, documentation, security audit)
• Programs -- autonomous metric-driven loops (test coverage optimizer, dependency freshness, dead code eliminator)

writ search "code review"          # Search from CLI
writ install code-review-agent     # Install into your project
writ use code-review-agent         # Activate in your IDE

Templates

Bootstrap an agent team in seconds:

writ init --template fullstack     # Architect + implementer + reviewer + tester
writ init --template python        # Python developer + reviewer
writ init --template typescript    # TypeScript developer + reviewer
writ init --template rules         # Project rule + coding standards

Personal Library & Cloud Sync

writ save my-reviewer              # Save to library (local + cloud)
writ login                         # Authenticate for cross-device sync
writ load my-reviewer              # Load on any machine
writ library                       # See everything (local + remote)

Cross-Project Memory

writ memory export research-insights     # Export from current project
writ memory import research-insights     # Import in another project

Agent-to-Agent Communication

Agents can have structured conversations across repos:

writ peers add partner-repo --path ../partner-repo
writ chat start --with partner-repo --goal "Review API design"
writ chat send <conv-id> "Here's my proposed schema..."
writ inbox                          # Check for responses

All Commands

Command	Description
writ init	Initialize in current repo
writ add <name>	Create instruction (agent, rule, context, program)
writ add --file <path>	Import markdown file(s) or directory
writ list	List all instructions
writ use <name>	Activate (compose + write to IDE files)
writ edit <name>	Open in $EDITOR
writ remove <name>	Remove instruction
writ export <name> <format>	Export to specific format
writ compose <name>	Preview composed context
writ save / load	Personal library (local + cloud)
writ library	List personal library
writ search <query>	Search Hub
writ install <name>	Install from Hub
writ publish / unpublish	Make publicly discoverable
writ login / logout	Authenticate with enwrit.com
writ register	Create account
writ lint [file|name] [--deep] [--deep-local] [--stop-server]	Quality score (0-100, 6 dimensions)
writ sync	Bulk bidirectional library sync
writ mcp serve	Start MCP server (22 tools)
writ chat start/send/inbox	Agent-to-agent conversations
writ memory export/import	Cross-project memory
writ handoff create	Create agent handoff
writ review <name>	Browse/submit reviews
writ threads	Knowledge threads

Instruction Format

Instructions are YAML files in .writ/, routed by task_type:

name: reviewer
description: "Code reviewer for TypeScript"
version: 1.0.0
task_type: agent    # agent | rule | context | program | template
tags: [typescript, review]
instructions: |
  You are a code reviewer specializing in TypeScript.
  Focus on: type safety, component composition, performance.
composition:
  inherits_from: [architect]
  project_context: true

Users interact primarily with markdown. Use writ add --file to import .md, .mdc, or .txt files directly.

Development

git clone https://github.com/enwrit/writ.git
cd writ
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -e ".[dev]"
pytest                    # 385 tests
ruff check src/ tests/

License

MIT

Links

• Hub: enwrit.com/hub -- Browse and install instructions
• Website: enwrit.com
• GitHub: github.com/enwrit/writ
• PyPI: pypi.org/project/enwrit