Craft Agents

How it Works (Video)

To understand what Craft Agents does and how it works watch this video.

Click Here (or on the image above) to watch the video on YouTube →

Why Craft Agents was built

Craft Agents is a tool we built so that we (at craft.do) can work effectively with agents. It enables intuitive multitasking, no-fluff connection to any API or Service, sharing sessions, and a more document (vs code) centric workflow - in a beautiful and fluid UI.

It uses the Claude Agent SDK and the Pi SDK side by side—building on what we found great and improving areas where we've desired improvements.

It's built with Agent Native software principles in mind, and is highly customisable out of the box. One of the first of its kind.

Craft Agents is open source under the Apache 2.0 license - so you are free to remix, change anything. And that's actually possible. We ourselves are building Craft Agents with Craft Agents only - no code editors - so really, any customisation is just a prompt away.

We built Craft Agents because we wanted a better, more opinionated (and preferably non-CLI way) of working with the most powerful agents in the world. We'll continue to improve it, based on our experiences and intuition.

Things that are hard to believe "just work"

How do I connect to Linear, Gmail, Slack...? Tell the agent "add Linear as a source." It finds public APIs and MCP servers, reads their docs, sets up credentials, and configures everything. No config files, no setup wizards.

Check out how I just connected to Slack →

I already have my MCP config JSON. Paste it. The agent handles the rest.

What about local MCPs? Fully supported. Stdio-based MCP servers run as local subprocesses on your machine. Point it at an npx command, a Python script, or any local binary. It just works.

Can it handle custom APIs? Yes. Paste an OpenAPI spec, some endpoint URLs, screenshots of docs, whatever you have. It figures it out and guides you through the rest.

APIs too? Not just MCPs? Craft Agents connects to anything. We have it hooked up to a direct Postgres DB behind a jumpbox. Skills + Sources = magic.

How do I import my Claude Code skills and MCPs? Tell the agent you want to import your skills from Claude Code. It handles the migration.

Here I imported all my skills in one go →

How do I create a new skill? Describe what the skill should do, give it context. The agent takes care of the rest.

Do I need to restart after changes? No. Everything is instant. Mention new skills or sources with @, even mid-conversation.

So I can just ask it anything? Yes. That's the core idea behind agent-native software. You describe what you want, and it figures out how. That's a good use of tokens.

Installation

One-Line Install (Recommended)

macOS / Linux:

curl -fsSL https://agents.craft.do/install-app.sh | bash

Windows (PowerShell):

irm https://agents.craft.do/install-app.ps1 | iex

Build from Source

git clone https://github.com/lukilabs/craft-agents-oss.git
cd craft-agents-oss
bun install
bun run electron:start

Features

• Multi-Session Inbox: Desktop app with session management, status workflow, and flagging
• Claude Code Experience: Streaming responses, tool visualization, real-time updates
• Multiple LLM Connections: Add multiple AI providers and set per-workspace defaults
• Multi-Provider Support: Run sessions with Google AI Studio, ChatGPT Plus, GitHub Copilot, or OpenAI API keys alongside Anthropic
• Craft MCP Integration: Access to 32+ Craft document tools (blocks, collections, search, tasks)
• Sources: Connect to MCP servers, REST APIs (Google, Slack, Microsoft), and local filesystems
• Permission Modes: Three-level system (Explore, Ask to Edit, Auto) with customizable rules
• Background Tasks: Run long-running operations with progress tracking
• Dynamic Status System: Customizable session workflow states (Todo, In Progress, Done, etc.)
• Theme System: Cascading themes at app and workspace levels
• Multi-File Diff: VS Code-style window for viewing all file changes in a turn
• Skills: Specialized agent instructions stored per-workspace
• File Attachments: Drag-drop images, PDFs, Office documents with auto-conversion
• Automations: Event-driven automation — create agent sessions on label changes, schedules, tool use, and more

Quick Start

1. Launch the app after installation
2. Choose API Connection: Use Anthropic (API key or Claude Max), Google AI Studio, ChatGPT Plus (Codex OAuth), or GitHub Copilot OAuth
3. Create a workspace: Set up a workspace to organize your sessions
4. Connect sources (optional): Add MCP servers, REST APIs, or local filesystems
5. Start chatting: Create sessions and interact with Claude

Desktop App Features

Session Management

• Inbox/Archive: Sessions organized by workflow status
• Flagging: Mark important sessions for quick access
• Status Workflow: Todo → In Progress → Needs Review → Done
• Session Naming: AI-generated titles or manual naming
• Session Persistence: Full conversation history saved to disk

Sources

Connect external data sources to your workspace:

Type	Examples
MCP Servers	Craft, Linear, GitHub, Notion, custom servers
REST APIs	Google (Gmail, Calendar, Drive, YouTube, Search Console), Slack, Microsoft
Local Files	Filesystem, Obsidian vaults, Git repos

Permission Modes

Mode	Display	Behavior
safe	Explore	Read-only, blocks all write operations
ask	Ask to Edit	Prompts for approval (default)
allow-all	Auto	Auto-approves all commands

Use SHIFT+TAB to cycle through modes in the chat interface.

Keyboard Shortcuts

Shortcut	Action
Cmd+N	New chat
Cmd+1/2/3	Focus sidebar/list/chat
Cmd+/	Keyboard shortcuts dialog
SHIFT+TAB	Cycle permission modes
Enter	Send message
Shift+Enter	New line

Remote Server (Headless)

Craft Agents can run as a headless server on a remote machine (e.g., a Linux VPS), with the desktop app connecting as a thin client. This lets you keep long-running sessions alive, access them from multiple machines, and run compute-heavy tasks on a powerful server.

Quick Start

From the monorepo root:

# Generate a token and start the server
CRAFT_SERVER_TOKEN=$(openssl rand -hex 32) bun run packages/server/src/index.ts

The server prints the connection details on startup:

CRAFT_SERVER_URL=ws://203.0.113.5:9100
CRAFT_SERVER_TOKEN=<generated-token>

Copy these values and use them to connect the desktop app.

Connecting the Desktop App

Launch the Electron app in thin-client mode by passing the server URL and token:

CRAFT_SERVER_URL=wss://203.0.113.5:9100 CRAFT_SERVER_TOKEN=<token> bun run electron:start

In thin-client mode, the desktop app renders the UI but all session logic, tool execution, and LLM calls run on the remote server.

Environment Variables

Variable	Required	Default	Description
CRAFT_SERVER_TOKEN	Yes	—	Bearer token for client authentication
CRAFT_RPC_HOST	No	127.0.0.1	Bind address (0.0.0.0 for remote access)
CRAFT_RPC_PORT	No	9100	Bind port
CRAFT_RPC_TLS_CERT	No	—	Path to PEM certificate file (enables wss://)
CRAFT_RPC_TLS_KEY	No	—	Path to PEM private key file (required with cert)
CRAFT_RPC_TLS_CA	No	—	Path to PEM CA chain file (optional, for client cert verification)
CRAFT_DEBUG	No	false	Enable debug logging

TLS (Recommended for Remote Access)

When exposing the server over the network, TLS encrypts the WebSocket connection (wss:// instead of ws://).

Generate a self-signed certificate (development/testing):

./scripts/generate-dev-cert.sh
# Creates certs/cert.pem and certs/key.pem (valid 365 days)

Start the server with TLS:

CRAFT_SERVER_TOKEN=<token> \
CRAFT_RPC_HOST=0.0.0.0 \
CRAFT_RPC_TLS_CERT=certs/cert.pem \
CRAFT_RPC_TLS_KEY=certs/key.pem \
bun run packages/server/src/index.ts

The server will print CRAFT_SERVER_URL=wss://<your-public-ip>:9100.

For production, use certificates from a trusted CA (e.g., Let's Encrypt) or place the server behind a reverse proxy (nginx, Caddy) that terminates TLS.

Docker

docker run -d \
  -p 9100:9100 \
  -e CRAFT_SERVER_TOKEN=<token> \
  -e CRAFT_RPC_HOST=0.0.0.0 \
  -v craft-data:/root/.craft-agent \
  craft-agents-server

To enable TLS in Docker, mount your certificates and set the env vars:

docker run -d \
  -p 9100:9100 \
  -e CRAFT_SERVER_TOKEN=<token> \
  -e CRAFT_RPC_HOST=0.0.0.0 \
  -e CRAFT_RPC_TLS_CERT=/certs/cert.pem \
  -e CRAFT_RPC_TLS_KEY=/certs/key.pem \
  -v ./certs:/certs:ro \
  -v craft-data:/root/.craft-agent \
  craft-agents-server

CLI Client

A terminal client that connects to a running Craft Agent server over WebSocket (ws:// or wss://). Use it for scripting, CI/CD pipelines, server validation, or when you prefer the command line.

Installation

# From the monorepo (requires Bun)
bun run apps/cli/src/index.ts --help

# Or add to your PATH
alias craft-cli="bun run $(pwd)/apps/cli/src/index.ts"

Connection

The CLI reads connection details from flags or environment variables:

# Via environment (set once)
export CRAFT_SERVER_URL=ws://127.0.0.1:9100
export CRAFT_SERVER_TOKEN=<your-token>

# Or via flags
craft-cli --url ws://127.0.0.1:9100 --token <token> ping

For TLS connections (wss://), use --tls-ca <path> for self-signed certificates.

Commands

Command	Description
ping	Verify connectivity (clientId + latency)
health	Check credential store health
versions	Show server runtime versions
workspaces	List workspaces
sessions	List sessions in workspace
connections	List LLM connections
sources	List configured sources
session create	Create a session (--name, --mode)
session messages <id>	Print session message history
session delete <id>	Delete a session
send <id> <message>	Send message and stream AI response
cancel <id>	Cancel in-progress processing
invoke <channel> [args]	Raw RPC call with JSON args
listen <channel>	Subscribe to push events (Ctrl+C to stop)
run <prompt>	Self-contained: spawn server, run prompt, stream response, exit
--validate-server	21-step integration test (auto-spawns server if no --url)

Run Command Flags

Flag	Default	Description
--workspace-dir <path>	—	Register a workspace directory before running
--source <slug>	—	Enable a source (repeatable)
--output-format <fmt>	text	Output format: text or stream-json
--mode <mode>	allow-all	Permission mode for the session
--no-cleanup	false	Skip session deletion on exit
--server-entry <path>	—	Custom server entry point
--provider <name>	anthropic	LLM provider (anthropic, openai, google, openrouter, groq, mistral, xai, etc.)
--model <id>	(provider default)	Model ID (e.g., claude-sonnet-4-5-20250929, gpt-4o, gemini-2.0-flash)
--api-key <key>	—	API key (or $LLM_API_KEY, or provider-specific env var)
--base-url <url>	—	Custom API endpoint for proxies or self-hosted models

The run command is fully self-contained — it spawns a headless server, creates a session, sends the prompt, streams the response, and exits. No separate server setup needed. An API key is resolved from --api-key, $LLM_API_KEY, or a provider-specific env var (e.g., $ANTHROPIC_API_KEY, $OPENAI_API_KEY).

Examples

# Quick connectivity check
craft-cli ping

# List sessions (human-readable)
craft-cli sessions

# Send a message and stream the AI response
craft-cli send abc-123 "What files are in the current directory?"

# Pipe input
echo "Summarize this" | craft-cli send abc-123

# JSON output for scripting
craft-cli --json workspaces | jq '.[].name'

# Self-contained run (spawns its own server)
craft-cli run "Summarize the README"
craft-cli run --workspace-dir ./my-project --source github "List open PRs"

# Multi-provider support
craft-cli run --provider openai --model gpt-4o "Summarize this repo"
GOOGLE_API_KEY=... craft-cli run --provider google --model gemini-2.0-flash "Hello"
craft-cli run --provider anthropic --base-url https://openrouter.ai/api/v1 --api-key $OR_KEY "Hello"

# Validate the server (auto-spawns if no --url)
craft-cli --validate-server
craft-cli --validate-server --url ws://127.0.0.1:9100 --token <token>

Architecture

craft-agent/
├── apps/
│   ├── cli/                   # Terminal client (CLI)
│   └── electron/              # Desktop GUI (primary)
│       └── src/
│           ├── main/          # Electron main process
│           ├── preload/       # Context bridge
│           └── renderer/      # React UI (Vite + shadcn)
└── packages/
    ├── core/                  # Shared types
    └── shared/                # Business logic
        └── src/
            ├── agent/         # CraftAgent, permissions
            ├── auth/          # OAuth, tokens
            ├── config/        # Storage, preferences, themes
            ├── credentials/   # AES-256-GCM encrypted storage
            ├── sessions/      # Session persistence
            ├── sources/       # MCP, API, local sources
            └── statuses/      # Dynamic status system

Development

# Hot reload development
bun run electron:dev

# Build and run
bun run electron:start

# Type checking
bun run typecheck:all

# Debug logging (writes to ~/Library/Logs/@craft-agent/electron/)
# Logs are automatically enabled in development

Environment Variables

OAuth integrations (Slack, Microsoft) require credentials baked into the build. Create a .env file:

MICROSOFT_OAUTH_CLIENT_ID=your-client-id
SLACK_OAUTH_CLIENT_ID=your-slack-client-id
SLACK_OAUTH_CLIENT_SECRET=your-slack-client-secret

Note: Google OAuth credentials are NOT baked into the build. Users provide their own credentials via source configuration. See the Google OAuth Setup section below.

Google OAuth Setup (Gmail, Calendar, Drive, YouTube, Search Console)

Google integrations require you to create your own OAuth credentials. This is a one-time setup.

1. Create a Google Cloud Project

1. Go to Google Cloud Console
2. Create a new project (or select an existing one)
3. Note your Project ID

2. Enable Required APIs

Go to APIs & Services → Library and enable the APIs you need:

• Gmail API - for email integration
• Google Calendar API - for calendar integration
• Google Drive API - for file storage integration

3. Configure OAuth Consent Screen

1. Go to APIs & Services → OAuth consent screen
2. Select External user type (unless you have Google Workspace)
3. Fill in required fields:
   • App name: e.g., "My Craft Agent"
   • User support email: your email
   • Developer contact: your email
4. Add scopes (optional - can leave default)
5. Add yourself as a test user (required for External apps in testing mode)
6. Complete the wizard

4. Create OAuth Credentials

1. Go to APIs & Services → Credentials
2. Click Create Credentials → OAuth Client ID
3. Application type: Desktop app
4. Name: e.g., "Craft Agent Desktop"
5. Click Create
6. Note the Client ID and Client Secret

5. Configure in Craft Agent

When setting up a Google source (Gmail, Calendar, Drive, YouTube, Search Console, etc.), add these fields to your source's config.json:

{
  "api": {
    "googleService": "gmail",
    "googleOAuthClientId": "your-client-id.apps.googleusercontent.com",
    "googleOAuthClientSecret": "your-client-secret"
  }
}

Or simply tell the agent you want to connect Gmail/Calendar/Drive - it will guide you through entering your credentials.

Security Notes

• Your OAuth credentials are stored encrypted alongside other source credentials
• Never commit credentials to version control
• For production use, consider getting your OAuth consent screen verified by Google

Supported LLM Providers

Craft Agents supports multiple ways to connect to LLM providers:

Direct Connections

Provider	Auth	Notes
Anthropic	API key or Claude Max/Pro OAuth	Direct Claude connection via the Claude Agent SDK
Google AI Studio	API key	Gemini models with native Google Search grounding built in
ChatGPT Plus / Pro	Codex OAuth	Sign in with your ChatGPT subscription — uses OpenAI's Codex models
GitHub Copilot	OAuth (device code)	One-click authentication with your Copilot subscription

Third-Party & Self-Hosted Providers

Additional providers are supported through the Claude / Anthropic API Key connection by choosing a custom endpoint:

Provider	Endpoint	Notes
OpenRouter	https://openrouter.ai/api	Access Claude, GPT, Llama, Gemini, and hundreds of other models through a single API key. Use provider/model-name format (e.g. anthropic/claude-opus-4.7).
Vercel AI Gateway	https://ai-gateway.vercel.sh	Route requests through Vercel's AI Gateway with built-in observability and caching.
Ollama	http://localhost:11434	Run open-source models locally. No API key required.
Custom	Any URL	Any OpenAI-compatible or Anthropic-compatible endpoint.

Architecture

Craft Agents uses two agent backends:

• Claude — powered by the Claude Agent SDK, which natively supports custom base URLs and provider routing. Anthropic API key, Claude Max/Pro OAuth, and all third-party endpoints use this backend.
• Pi — powered by the Pi SDK, which handles Google AI Studio, ChatGPT Plus (Codex OAuth), GitHub Copilot OAuth, and OpenAI API key connections. Pi connections route through their own provider infrastructure.

Configuration

Configuration is stored at ~/.craft-agent/:

~/.craft-agent/
├── config.json              # Main config (workspaces, LLM connections)
├── credentials.enc          # Encrypted credentials (AES-256-GCM)
├── preferences.json         # User preferences
├── theme.json               # App-level theme
└── workspaces/
    └── {id}/
        ├── config.json      # Workspace settings
        ├── theme.json       # Workspace theme override
        ├── automations.json  # Event-driven automations
        ├── sessions/        # Session data (JSONL)
        ├── sources/         # Connected sources
        ├── skills/          # Custom skills
        └── statuses/        # Status configuration

Automations

Automations let you automate workflows by triggering actions when events happen — labels change, sessions start, tools run, or on a cron schedule.

Just ask the agent:

• "Set up a daily standup briefing every weekday at 9am"
• "Notify me when a session is labelled urgent"
• "Track permission mode changes and summarise them"
• "Every Friday at 5pm, summarise this week's completed tasks"

Or configure manually in ~/.craft-agent/workspaces/{id}/automations.json:

{
  "version": 2,
  "automations": {
    "SchedulerTick": [
      {
        "cron": "0 9 * * 1-5",
        "timezone": "America/New_York",
        "labels": ["Scheduled"],
        "actions": [
          { "type": "prompt", "prompt": "Check @github for new issues assigned to me" }
        ]
      }
    ],
    "LabelAdd": [
      {
        "matcher": "^urgent$",
        "actions": [
          { "type": "prompt", "prompt": "An urgent label was added. Triage the session and summarise what needs attention." }
        ]
      }
    ]
  }
}

Prompt actions create a new agent session with a prompt. They support @mentions for sources and skills, and environment variables like $CRAFT_LABEL and $CRAFT_SESSION_ID are expanded automatically.

Supported events: LabelAdd, LabelRemove, PermissionModeChange, FlagChange, SessionStatusChange, SchedulerTick, PreToolUse, PostToolUse, SessionStart, SessionEnd, and more.

See the Automations documentation for the full reference.

Advanced Features

Large Response Handling

Tool responses exceeding ~60KB are automatically summarized using Claude Haiku with intent-aware context. The _intent field is injected into MCP tool schemas to preserve summarization focus.

Deep Linking

External apps can navigate using craftagents:// URLs:

craftagents://allSessions                      # All sessions view
craftagents://allSessions/session/session123   # Specific session
craftagents://settings                         # Settings
craftagents://sources/source/github            # Source info
craftagents://action/new-chat                  # Create new session

Tech Stack

Layer	Technology
Runtime	Bun
AI	@anthropic-ai/claude-agent-sdk
AI (Pi)	Pi SDK agent server
Desktop	Electron + React
UI	shadcn/ui + Tailwind CSS v4
Build	esbuild (main) + Vite (renderer)
Credentials	AES-256-GCM encrypted file storage

Troubleshooting

Debug Mode

To launch the packaged app with verbose logging enabled, use -- --debug (note the double dash separator):

macOS:

/Applications/Craft\ Agents.app/Contents/MacOS/Craft\ Agents -- --debug

Windows (PowerShell):

& "$env:LOCALAPPDATA\Programs\@craft-agentelectron\Craft Agents.exe" -- --debug

Linux:

./craft-agents -- --debug

Logs are written to:

• macOS: ~/Library/Logs/@craft-agent/electron/main.log
• Windows: %APPDATA%\@craft-agent\electron\logs\main.log
• Linux: ~/.config/@craft-agent/electron/logs/main.log

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Third-Party Licenses

This project uses the Claude Agent SDK, which is subject to Anthropic's Commercial Terms of Service.

Trademark

"Craft" and "Craft Agents" are trademarks of Craft Docs Ltd. See TRADEMARK.md for usage guidelines.

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Security

Local MCP Server Isolation

When spawning local MCP servers (stdio transport), sensitive environment variables are filtered out to prevent credential leakage to subprocesses. Blocked variables include:

• ANTHROPIC_API_KEY, CLAUDE_CODE_OAUTH_TOKEN (app auth)
• AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN
• GITHUB_TOKEN, GH_TOKEN, OPENAI_API_KEY, GOOGLE_API_KEY, STRIPE_SECRET_KEY, NPM_TOKEN

To explicitly pass an env var to a specific MCP server, use the env field in the source config.

To report security vulnerabilities, please see SECURITY.md.