Slipbox MCP Server

Give your AI assistant an active role in managing your knowledge. Slipbox is an MCP server that turns any MCP-compatible agent into a Zettelkasten partner -- creating atomic notes, forming semantic links, detecting emergent clusters, and synthesizing insights from your existing knowledge.

Your ideas in, structured knowledge out. The agent handles the formatting, linking, and integration.

New to the method? Start with Introduction to the Zettelkasten Method for the why behind atomic notes and linked thinking. To see how Slipbox primes your agent with that method, read the server instructions it ships automatically on connect.

Built and tested with Claude. Works with any MCP client (Claude Desktop, Claude Code, OpenCode, Copilot, or anything that speaks MCP).

Plain files, zero lock-in. Notes are markdown with YAML frontmatter -- readable in Obsidian, Foam, Logseq, or any editor. The SQLite database is an index, not the source of truth. Delete it and rebuild from files anytime.

• 19 MCP tools for notes, links, search, graph analysis, and cluster management

• 6 workflow prompts (plus matching skills) encoding the Zettelkasten method so you don't re-learn it every session

• BM25 full-text search across titles and content via SQLite FTS5

• Cluster detection finds emergent topic groups and scaffolds structure notes

• Seven typed links (reference, extends, refines, contradicts, questions, supports, related)

Python 3.10+ | macOS or Linux

Walkthrough

Related MCP server: vault-master-mcp

Quick Start

1. Install

pipx install slipbox-mcp
# or, with uv:
uv tool install slipbox-mcp

This puts a slipbox-mcp launcher on your PATH (in ~/.local/bin). That single command is the whole MCP server: no clone, no PYTHONPATH, no hardcoded venv Python path. Everything below uses it. To try it without installing at all, uvx slipbox-mcp runs the server in a throwaway environment.

(Working on Slipbox itself? See Development for the clone + editable-install setup.)

2. Pick a Data Directory

One variable, SLIPBOX_BASE_DIR, configures everything: notes land in <base>/data/notes and the SQLite index in <base>/data/db/zettelkasten.db. The server creates these on first run.

# Example: use any absolute path you like
/Users/yourname/.local/share/mcp/slipbox

Use a full absolute path. A leading ~ is not expanded inside MCP client config files and would create a literal ~ directory.

3. Connect to Your MCP Client

Claude Code (one command, no file editing):

claude mcp add slipbox \
  --env SLIPBOX_BASE_DIR=/Users/yourname/.local/share/mcp/slipbox \
  -- slipbox-mcp

Claude Desktop (edit the config file):

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Linux: ~/.config/claude/claude_desktop_config.json

{
  "mcpServers": {
    "slipbox": {
      "command": "slipbox-mcp",
      "env": {
        "SLIPBOX_BASE_DIR": "/Users/yourname/.local/share/mcp/slipbox"
      }
    }
  }
}

Desktop PATH caveat: the macOS Desktop app doesn't always inherit ~/.local/bin on its PATH, so the bare "slipbox-mcp" may not resolve. If the server fails to start, replace "command": "slipbox-mcp" with the absolute path printed by which slipbox-mcp (typically /Users/yourname/.local/bin/slipbox-mcp).

Other MCP clients: register slipbox-mcp as the server command with SLIPBOX_BASE_DIR in its environment. The command and env are the same everywhere.

Instead of SLIPBOX_BASE_DIR, set absolute paths individually. Optional SLIPBOX_LOG_LEVEL is one of DEBUG, INFO, WARNING, ERROR.

"env": {
  "SLIPBOX_NOTES_DIR": "/Users/yourname/.local/share/mcp/slipbox/notes",
  "SLIPBOX_DATABASE_PATH": "/Users/yourname/.local/share/mcp/slipbox/data/db/zettelkasten.db",
  "SLIPBOX_LOG_LEVEL": "INFO"
}

4. Restart and Verify

Restart your client (Claude Code reloads on next launch; quit and reopen Claude Desktop).

Ask your agent:

• "Create a test note about something"

• "Search my slipbox for test"

• "Find orphaned notes"

In Action

The hero above is the core loop. Here's the rest of what the agent does.

Proactive Maintenance

The agent reads the slipbox://maintenance-status resource at session start and surfaces clusters that need organizing.

Full-Text Search

BM25-ranked search across notes via slipbox_search_notes.

Knowledge Graph: Central Notes

slipbox_find_central_notes surfaces the structural anchors of the graph -- the notes everything else orbits.

Note Analysis

The analyze_note prompt evaluates atomicity, finds real connections in the existing graph, suggests tags, and rewrites for clarity.

Source Decomposition

The knowledge_creation prompt splits an article into atomic literature notes with proper citation and links.

Cluster Detection

slipbox_get_cluster_report finds groups of co-occurring tags that lack a structure note. Scored by size, orphan ratio, link density, and recency.

Structure Note Creation

slipbox_create_structure_from_cluster scaffolds a structure note, links all member notes, and dismisses the cluster.

Orphaned Notes

slipbox_find_orphaned_notes surfaces unintegrated knowledge -- candidates for connection or deletion.

Similar Notes

slipbox_find_similar_notes computes similarity from shared tags, common links, and content overlap.

Graph Traversal

slipbox_get_linked_notes shows typed links from a hub note, grouped by link type.

Knowledge Synthesis

The knowledge_synthesis prompt finds bridges between unconnected areas and proposes synthesis notes from your existing knowledge.

Zero Lock-In: Plain Files in Obsidian

Notes are plain markdown. Open the vault in Obsidian and everything works -- rendered content, backlinks, and the knowledge graph.

For a graph that renders the typed links in color (supports, extends, refines, ...) rather than Obsidian's untyped built-in graph, install the companion plugin Slipbox Semantic Graph -- a force-directed view with human-readable titles and color-coded semantic link types. Install it manually from the 0.1.0 release: copy main.js, manifest.json, and styles.css into <vault>/.obsidian/plugins/slipbox-graph/, then enable it in Settings → Community plugins. (Once it's accepted into the official directory, you'll also be able to install it via Settings → Community plugins → Browse → search "Slipbox Semantic Graph".) It reads the same frontmatter id and ## Links section the server writes, so no extra configuration is needed. Open the view with the Open semantic graph command (Command Palette) or the git-fork ribbon icon.

The legend across the top maps each color to a link type (extends, refines, supports, contradicts, questions, related). Focus a structure note and its constellation comes into view. Here, Contract Testing Knowledge Map with its member notes orbiting it:

Optional: Automatic Cluster Detection

Cluster analysis scans all notes and computes similarity scores. Running it daily (6am) pre-computes results so slipbox_get_cluster_report() returns instantly. Without scheduling, cluster detection runs on-demand, which is slower for large collections.

Run manually after bulk imports, major reorganization, or when you want immediate results.

Install Cluster Detection (macOS)

chmod +x scripts/install-cluster-detection.sh
./scripts/install-cluster-detection.sh

The installer detects your Python/venv path, generates the LaunchAgent plist, and loads it.

Manual Test (File Watcher)

source .venv/bin/activate
python scripts/detect_clusters.py

Output saved to ~/.local/share/mcp/slipbox/cluster-analysis.json.

Uninstall Cluster Detection

./scripts/install-cluster-detection.sh --uninstall

Optional: macOS File Watcher for Auto-Indexing

The MCP server maintains a database index for fast searching. Editing notes in Obsidian (or any editor) makes the database stale until you run slipbox_rebuild_index.

The file watcher runs as a background daemon, monitoring your notes directory and automatically rebuilding the index when .md files change.

Use it if you frequently edit notes in Obsidian while also using Claude.

Install File Watcher (macOS)

chmod +x scripts/install-file-watcher.sh
./scripts/install-file-watcher.sh

The installer detects your Python/venv path, installs watchdog if needed, and loads the LaunchAgent. Starts on login and restarts if it crashes.

Manual Test

source .venv/bin/activate
python scripts/watch_notes.py

Edit a note file. You should see "rebuilding index..." in the watcher output.

Check Status

launchctl list | grep slipbox.watcher

# View logs

tail -f ~/.local/share/mcp/slipbox/watcher.log

Uninstall File Watcher

./scripts/install-file-watcher.sh --uninstall

Recommended System Prompt

Slipbox ships a baseline automatically: every client receives the server instructions on connect, covering how to use the tools well -- note types, link semantics, quality standards, and core workflows like search-before-create. You don't add those yourself.

docs/SYSTEM_PROMPT.md is the opt-in layer on top: the autonomy and initiative directives a server shouldn't assert on its own. Add it to your agent's preferences or system prompt to enable:

• Automatic knowledge capture during conversations

• Cluster emergence detection at conversation start

Tools Reference

Core Note Operations

Linking

Search & Discovery

Cluster Analysis

Maintenance

Prompts Reference

MCP prompts are reusable workflow templates that encode the Zettelkasten method so you don't re-explain it every session.

How to Invoke: Slash Commands and Skills

Each workflow ships two ways:

• MCP prompts: served by the running server.

• Skills: standalone bundles (skills/<name>/) that run the same workflow and add natural-language triggering.

Five of the six skills are generated from the same PROMPT_* templates the server uses (src/slipbox_mcp/server/descriptions.py), and CI fails if the committed skills/ drift from those templates. The sixth, cluster-maintenance, is authored directly in scripts/build_skills.py because its MCP prompt is a runtime-rendered status message rather than a reusable workflow.

Slash commands are the reliable path. Claude Code surfaces MCP prompts as /mcp__<server>__<prompt>; type /mcp__slipbox-mcp__ for the picker:

/mcp__slipbox-mcp__knowledge_creation
/mcp__slipbox-mcp__knowledge_exploration
/mcp__slipbox-mcp__knowledge_synthesis
/mcp__slipbox-mcp__knowledge_creation_batch
/mcp__slipbox-mcp__analyze_note
/mcp__slipbox-mcp__cluster_maintenance

(Installed skills also expose their own slash commands by directory name, e.g. /slipbox-analyze-note.)

Natural language works once the matching skill is installed. Just describe what you want:

Analyze this note for my slipbox: [paste note]

Add this to my slipbox: [paste article]

Synthesize my notes on attention and memory.

Prose triggering depends on the skill being installed and your phrasing matching its description; fall back to the slash command if it doesn't fire. Asking the model to "use the analyze_note prompt" by name does not work. The model can't invoke an MCP prompt by name. Use a slash command, or let a skill trigger from natural language.

Installing Skills

Claude Code discovers skills from .claude/skills/ (per project) or ~/.claude/skills/ (global), not from a bare top-level skills/. Symlink or copy the ones you want into a discovery path (e.g. for this project):

mkdir -p .claude/skills
ln -s ../../skills/slipbox-analyze-note .claude/skills/slipbox-analyze-note
# ...or copy the directories, or symlink all six

Claude Desktop needs each skill as a .skill bundle. Build them, then upload:

python scripts/build_skills.py     # writes dist/*.skill

Go to Settings → Skills → Upload skill and select the bundles from dist/ you want. Each installs as both a slash command and a natural-language trigger.

After editing a prompt template in descriptions.py, re-run the build to regenerate the skills.

Link Types

Note Types

Structure vs. Hub: A structure note organizes a cluster of permanent notes around a single topic. It is a curated map one level above the notes themselves. A hub note operates one level higher still: it links to structure notes (and occasionally key permanent notes) across an entire knowledge domain. Where a structure note answers "what do I know about X?", a hub note answers "how is my knowledge of this whole domain organized?" Most Zettelkastens need only a handful of hub notes.

File Format

Notes are stored as Markdown files with YAML frontmatter:

---
id: "20251217T172432480464000"
title: "Poetry Revision Principles"
type: structure
tags:
  - poetry
  - revision
  - craft
created: "2025-12-17T17:24:32"
updated: "2025-12-17T17:24:32"
---

# Poetry Revision Principles

Content here...

## Links

- reference [[20250728T125429845760000]] Member of structure

You can edit these files directly in any text editor or Obsidian. Run slipbox_rebuild_index after external edits.

Upgrading

After pulling new versions, restart Claude Desktop. If the release notes mention database changes, run slipbox_rebuild_index once to bring your existing database up to date.

Upgrading to FTS5 search (any version after the FTS5 release): The full-text search index is created automatically when the server starts against a new database. For existing databases, the FTS5 table will be created on first startup but will be empty until you run:

slipbox_rebuild_index

This populates the BM25 index from your existing notes. Search results will not be relevance-ranked until this is done.

Troubleshooting

Server not loading in Claude Desktop

1. Confirm the launcher resolves: which slipbox-mcp should print a path (typically ~/.local/bin/slipbox-mcp).

2. If it resolves in your terminal but Desktop still can't start it, the GUI app isn't seeing ~/.local/bin on its PATH. Replace "command": "slipbox-mcp" with the absolute path from step 1.

3. Check Claude Desktop logs for errors.

slipbox-mcp: command not found

The console script wasn't installed or isn't on PATH. Reinstall with pipx install --editable . --force, then verify with which slipbox-mcp. If pipx's bin directory is missing from PATH, run pipx ensurepath and restart your shell.

Notes directory points to ~/... literally

If your notes directory ends up at ./~/... relative to CWD, you used ~ in the JSON config. Claude Desktop does not expand ~. Replace it with the full absolute path.

Search returns no results

1. The FTS5 index may not be populated. Run slipbox_rebuild_index once to index existing notes.

2. If you recently edited notes outside Claude, the index may be stale. Run slipbox_rebuild_index.

slipbox_list_notes_by_date returns empty results

If start_date is later than end_date, no notes match and an empty result is returned. This is expected behavior, not an error.

Database out of sync

If notes were edited outside the MCP server:

slipbox_rebuild_index

Cluster detection not running

launchctl list | grep slipbox.cluster-detection
# Should show: - 0 com.slipbox.cluster-detection

# Check logs

cat /tmp/slipbox-clusters.log

# Reinstall if needed

./scripts/install-cluster-detection.sh --uninstall
./scripts/install-cluster-detection.sh

File watcher not running

launchctl list | grep slipbox.watcher
# Should show: - 0 com.slipbox.watcher

# Check logs

cat ~/.local/share/mcp/slipbox/watcher.log

# Reinstall if needed

./scripts/install-file-watcher.sh --uninstall
./scripts/install-file-watcher.sh

Upgrading from ZETTELKASTEN_* environment variables

If you previously used ZETTELKASTEN_NOTES_DIR, ZETTELKASTEN_DATABASE_PATH, or other ZETTELKASTEN_* variables, they are no longer read. Rename them to their SLIPBOX_* equivalents:

The server logs a warning if old names are detected, but does not migrate them automatically.

Cluster report path is not configurable

The cluster analysis report always writes to ~/.local/share/mcp/slipbox/cluster-analysis.json, regardless of SLIPBOX_BASE_DIR or SLIPBOX_NOTES_DIR. If you use non-default paths, the cluster report will still be in the default location.

Install scripts are macOS-only

The scripts/install-cluster-detection.sh and scripts/install-file-watcher.sh scripts use launchctl and ~/Library/LaunchAgents/, which only exist on macOS. On Linux, you'll need to create equivalent systemd units or cron jobs manually. See the manual test commands in the relevant README sections to verify the underlying Python scripts work on your platform.

Default paths are relative to the working directory

If SLIPBOX_NOTES_DIR and SLIPBOX_DATABASE_PATH are not set, the server defaults to data/notes and data/db/zettelkasten.db relative to the current working directory. When running via Claude Desktop, the CWD may not be what you expect. Always set absolute paths in claude_desktop_config.json to avoid this.

Development

Setup

git clone https://github.com/jamesfishwick/slipbox-mcp.git
cd slipbox-mcp
uv venv && uv pip install -e ".[dev]"

Testing

The project has three tiers of tests:

# Default: runs unit + contract tests (CI runs this)

pytest

# Run everything except LLM evals

pytest tests/ evals/tool_contracts/

# Run LLM evals (requires claude CLI authenticated)

pytest evals/llm/ -v

# Run LLM evals with a specific model

EVAL_MODEL=sonnet pytest evals/llm/ -v

# Lint

ruff check src/ evals/

Unit tests cover internal logic -- services, repository, models, parsing.

Tool contract tests verify the MCP tool output format that the LLM sees -- parseable structure, chaining (create -> search -> get), and helpful error messages. These are deterministic and don't call any LLM.

LLM evals send prompts to an LLM via the claude CLI with the MCP server connected, then grade results by inspecting the database state (notes created, links made, tags applied). They test whether the LLM actually uses the tools correctly given the tool descriptions.

CI/CD

Branch protection: Direct pushes to main are blocked. All changes go through PRs.

The LLM eval suite is expensive (~$3-5, ~10 min) and self-hosted, so it never runs automatically. A path-based trigger can't distinguish a real prompt change from a cosmetic reformat. Run it deliberately when you change prompt or tool-description semantics:

• Add the run-llm-evals label to the PR. It runs, and re-runs on each push while the label is present.

• Or trigger it manually from the Actions tab (workflow_dispatch).

• Or run it locally without the runner: pytest evals/llm/ -v.

Without a label or manual dispatch, the job is skipped (no runner allocated, no cost).

Customizing the eval setup

If you don't want a self-hosted runner: remove .github/workflows/llm-evals.yml and run pytest evals/llm/ -v locally before merging prompt changes.

If you want LLM evals on every PR automatically: add a pull_request trigger with the relevant paths: filter and drop the label gate in the job's if:. But expect incidental triggers from formatting-only edits.

To change the default eval model: Set EVAL_MODEL in your environment or in the workflow file. Default is haiku for speed/cost.

To set up a self-hosted runner:

# Get a registration token

gh api repos/OWNER/REPO/actions/runners/registration-token -X POST -q '.token'

# Download and configure

mkdir -p ~/.github-runners/slipbox-mcp && cd ~/.github-runners/slipbox-mcp
curl -sL -o actions-runner.tar.gz https://github.com/actions/runner/releases/latest/download/actions-runner-osx-arm64-2.325.0.tar.gz
tar xzf actions-runner.tar.gz
./config.sh --url https://github.com/OWNER/REPO --token <TOKEN> --unattended
nohup ./run.sh &

Releasing to PyPI

Releases are automated. The Release workflow (.github/workflows/release.yml) runs release-please on every push to main and publishes via PyPI Trusted Publishing (OIDC, so no API token is stored in repo secrets).

The flow (you never hand-edit a version or push a tag):

1. Land changes on main with Conventional Commit messages (feat: → minor bump, fix: → patch, feat!:/BREAKING CHANGE: → major). The repo's commit hooks already enforce this shape.

2. release-please keeps a standing "release PR" open, accumulating the next version bump (in src/slipbox_mcp/__init__.py) and the CHANGELOG.md entries derived from those commits.

3. When you're ready to ship, merge the release PR. That tags the release (v<version>) and, in the same workflow run, builds the sdist + wheel, runs twine check, and publishes to PyPI.

So cutting a release is one click: merge the bot's PR. Nothing else.

Commit types decide the version. So type accurately. The bump is computed mechanically from the Conventional Commit prefixes since the last release, not from the size of the change. Reserve feat:/fix: for changes to the shipped package; use the non-releasing types for everything else:

A batch of only non-releasing commits produces no release PR at all. The squash-merge title is the commit release-please reads, so the PR title's prefix is what counts. Label it for what the package gains, not for the effort spent.

One-time setup (already done for this repo, documented for forks):

1. On PyPI, register a pending trusted publisher for project slipbox-mcp: Owner: jamesfishwick · Repository: slipbox-mcp · Workflow: release.yml · Environment: release. All four must match exactly.

2. In GitHub, create an environment named release (Settings → Environments). If you restrict its deployment refs, add a tag rule v* (a branch rule of the same name will not match the tag).

The version is defined once, in src/slipbox_mcp/__init__.py (release-please bumps it; the # x-release-please-version marker tells it which line). pyproject.toml (dynamic = ["version"]) and the server's server_version both read from it, so there is nothing to keep in sync; the tag release-please cuts always matches the package version by construction.

To rehearse a build without publishing, run it by hand: python -m build && twine check dist/* (and twine upload --repository testpypi dist/* with a TestPyPI token to dry-run the upload).

Shared prompt constants

All tool descriptions and prompt templates live in src/slipbox_mcp/server/descriptions.py. Both the MCP server and the eval tests import from this single source of truth. If you change a prompt, the evals test whether the LLM still behaves correctly with the new wording.

Debug logging

SLIPBOX_LOG_LEVEL=DEBUG python -c "from slipbox_mcp.main import main; main()"

CLI Tool

The slipbox command provides terminal access for mechanical operations:

slipbox status          # Overview of notes, tags, orphans, pending clusters
slipbox search <query>  # Find notes by text
slipbox clusters        # Show pending structure note candidates
slipbox orphans         # List unconnected notes
slipbox rebuild         # Rebuild index (add --clusters to refresh cluster analysis)
slipbox export <id>     # Export note markdown to stdout
slipbox tags            # List all tags with usage counts

Install: pipx install --editable . (adds slipbox to your PATH)

Experimental: Slipbox as agent memory

An untested hypothesis, not a recommended setup. Everything above helps an agent manage your knowledge. This inverts it: the agent uses a slipbox as its own persistent memory across sessions, in place of native memory or a rules file.

The model has no memory between sessions, so the slipbox is the only channel one session leaves for the next. It writes briefings for a cold successor (a failure and why, a recurring constraint, a correction, a hard-won fact), tags them agent-memory, and searches that tag before acting. The bet is that a connected memory beats a flat rules file, because you retrieve it by traversal.

Three things to know first: namespace isolation is a tag convention, not enforced, so run it against a separate slipbox instance; "memory" is a misnomer, since nothing persists but the notes themselves; and the growth discipline is the unproven part, so expect sprawl on the first run. Full write-up and caveats: Slipbox as Agent Self-Memory.

Documentation

Contributing

See CONTRIBUTING.md for setup instructions, coding standards, and how to submit changes.

Roadmap

See ROADMAP.md for planned features and future direction.

Sponsor

If slipbox-mcp is useful to you, consider sponsoring the project.

License

MIT