Plugin Install¶

The Distillery plugin packages all fourteen knowledge-base skills for distribution and installation via the Claude Code plugin mechanism. Once installed, the skills are available in every project without copying files.

Install via Plugin Marketplace (Recommended)¶

# Add the Distillery marketplace (one-time setup)
claude plugin marketplace add norrietaylor/distillery

# Install the plugin
claude plugin install distillery

This installs the skill definitions globally and configures the MCP server to run locally via uvx distillery-mcp — a private, self-contained knowledge base on your machine. Requires Python 3.11+ and uv on your PATH.

Install uv

curl -LsSf https://astral.sh/uv/install.sh | sh — or use pip install distillery-mcp and override the plugin's default command to distillery-mcp (see Troubleshooting).

After installation, restart Claude Code and run the onboarding wizard:

/setup

This verifies MCP connectivity, detects your transport, and configures auto-poll for ambient intelligence.

Hosted demo (opt-in)

Want a zero-install evaluation? You can opt into the hosted demo at distillery-mcp.fly.dev instead — see MCP Configuration below. The demo server is for evaluation only; do not store sensitive or confidential data.

Claude Desktop

The Claude Desktop app does not support Claude Code skills or the plugin install system. Desktop users can connect the MCP server directly (all 16 tools are available) but slash commands like /distill and /recall are CLI-only features.

Manual Install (Copy Skills)¶

Copy the skills directory into any project that should have access:

mkdir -p ~/.claude/skills
cp -r /path/to/distillery/skills/* ~/.claude/skills/

Or clone and symlink:

git clone https://github.com/norrietaylor/distillery.git ~/.claude/distillery
ln -s ~/.claude/distillery/skills/distill   ~/.claude/skills/distill
ln -s ~/.claude/distillery/skills/recall    ~/.claude/skills/recall
ln -s ~/.claude/distillery/skills/pour      ~/.claude/skills/pour
ln -s ~/.claude/distillery/skills/bookmark  ~/.claude/skills/bookmark
ln -s ~/.claude/distillery/skills/minutes   ~/.claude/skills/minutes
ln -s ~/.claude/distillery/skills/classify  ~/.claude/skills/classify
ln -s ~/.claude/distillery/skills/watch     ~/.claude/skills/watch
ln -s ~/.claude/distillery/skills/radar     ~/.claude/skills/radar
ln -s ~/.claude/distillery/skills/tune      ~/.claude/skills/tune
ln -s ~/.claude/distillery/skills/digest      ~/.claude/skills/digest
ln -s ~/.claude/distillery/skills/gh-sync     ~/.claude/skills/gh-sync
ln -s ~/.claude/distillery/skills/investigate  ~/.claude/skills/investigate
ln -s ~/.claude/distillery/skills/briefing    ~/.claude/skills/briefing
ln -s ~/.claude/distillery/skills/setup     ~/.claude/skills/setup

MCP Configuration¶

The skills require the Distillery MCP server. The plugin's default is local stdio via uvx distillery-mcp. Hosted demo and self-hosted HTTP are opt-in alternatives.

Default — Local stdio with uvx¶

claude plugin install distillery registers this configuration automatically. The plugin manifest declares:

{
  "mcpServers": {
    "distillery": {
      "command": "uvx",
      "args": ["distillery-mcp"]
    }
  }
}

uvx inherits the Claude Code process environment, so set JINA_API_KEY (and any other Distillery config vars) in your shell before launching Claude Code:

export JINA_API_KEY=jina_...   # free at jina.ai
# Optional:
export DISTILLERY_CONFIG=/path/to/distillery.yaml

Without a JINA_API_KEY, Distillery falls back to the stub embedding provider (search quality degraded). See Local Setup for full configuration (embedding providers, cloud storage, etc.).

If you prefer to manage the configuration yourself in ~/.claude.json, you can shadow the plugin registration with the same stdio block. uvx inherits the Claude Code process environment, so set JINA_API_KEY (and any other Distillery config vars) in your shell before launching Claude Code rather than relying on ${VAR} interpolation here:

{
  "mcpServers": {
    "distillery": {
      "command": "uvx",
      "args": ["distillery-mcp"]
    }
  }
}

Opt-in — Hosted demo¶

Distillery operates a hosted demo at https://distillery-mcp.fly.dev/mcp for zero-install evaluation. Authentication is via GitHub OAuth (Claude Code opens a browser on first use).

Demo Server

The demo server is for evaluation only. Do not store sensitive or confidential data. For production use, stick with the default local stdio setup or deploy your own instance.

Register the demo at user scope (shadows the plugin's local default):

claude mcp add distillery --scope user --transport http --url https://distillery-mcp.fly.dev/mcp

Or add the equivalent block to ~/.claude.json:

{
  "mcpServers": {
    "distillery": {
      "url": "https://distillery-mcp.fly.dev/mcp",
      "transport": "http"
    }
  }
}

Opt-in — Self-hosted HTTP¶

Deploy your own Distillery server with GitHub OAuth. See Operator Deployment for setup and the distill_ops repo for platform-specific guides (Fly.io, Prefect Horizon).

To shadow the plugin default with your own instance, register at user scope:

claude mcp add distillery --scope user --transport http --url https://your-instance.example.com/mcp

Remote Auto-Poll Setup¶

Enable remote auto-polling so feed sources are polled automatically even when Claude Code is not running. This uses Claude Code's scheduled remote agents (triggers).

HTTP transport required

Remote auto-poll requires an HTTP-reachable MCP endpoint (Hosted Demo or Self-hosted HTTP). The default local stdio transport (uvx distillery-mcp) only runs on your machine and is not reachable by remote triggers.

Step 1 — Register the MCP Connector¶

Open https://claude.ai/settings/connectors
Click "Add connector"
Enter the Distillery MCP server URL: https://distillery-mcp.fly.dev/mcp
Name it: distillery
Click Save

Step 2 — Run `/setup`¶

/setup

The wizard detects your transport, finds the registered connector, and creates a scheduled trigger that polls all feed sources every hour.

Step 3 — Verify¶

/watch list

You should see your feed sources and a note about the active remote trigger.

Verifying the Setup¶

After saving the settings file, restart Claude Code or reload MCP servers, then check connectivity by calling the distillery_status MCP tool:

distillery_status()

You should see a JSON response with "status": "ok".

You can also run the CLI health check directly:

distillery health

Available Skills¶

Skill	Trigger Phrases	Description
`/distill`	"capture this", "save knowledge", "log learnings"	Capture session decisions and insights with duplicate detection
`/recall`	"what do we know about", "search knowledge"	Semantic search with provenance
`/pour`	"synthesize", "what's the full picture on"	Multi-pass retrieval and synthesis with citations
`/bookmark`	"bookmark", "save this link", "store this URL"	Fetch a URL, summarize, and store
`/minutes`	"meeting notes", "capture meeting", "log meeting"	Structured meeting notes with append support
`/classify`	"classify", "review queue", "triage inbox"	Entry classification and review queue triage
`/watch`	"add feed", "remove source", "show my sources"	Manage feed sources with auto-poll scheduling
`/radar`	"what's new", "show my digest", "what have I missed"	Ambient intelligence digest from feed entries
`/tune`	"adjust thresholds", "tune my feed"	Display and adjust feed relevance thresholds
`/digest`	"team digest", "weekly summary"	Team activity summary from internal entries
`/gh-sync`	"sync GitHub", "import issues"	Sync GitHub issues/PRs into the knowledge base
`/investigate`	"investigate", "deep context"	Deep context builder with relationship traversal
`/briefing`	"team briefing", "dashboard"	Team knowledge dashboard with metrics and activity
`/setup`	"setup", "configure distillery"	Onboarding wizard — connectivity, connector registration, auto-poll

Troubleshooting¶

See MCP Server Reference — Troubleshooting for common problems and solutions.