Real-time global intelligence dashboard. AI-powered news aggregation, geopolitical monitoring, and infrastructure tracking in a unified situational awareness interface
MCP server for Things 3 task management
This server connects Things 3 to MCP clients so you can manage tasks in natural language. It exposes tools for todos, projects, areas, tags, lists, search, bulk moves, health checks, and usage stats, with AppleScript handling writes and structured output for reads.
Builders who want their agent to work with Things 3 for task management.
You can ask your agent to manage Things 3 instead of switching apps and retyping tasks by hand.
What it does
Todo operations
Create, read, update, delete, and search todos, including partial updates, bulk updates, and item lookup by ID.
Project and area management
List, create, and update projects and areas, including project hierarchies and project headings.
Built-in list access
Read Inbox, Today, Upcoming, Anytime, Someday, Logbook, and Trash views from Things 3.
Tag tools
List tags, create tags, add and remove tags on todos, and report tag usage for cleanup.
Bulk move tools
Move single records or many records under a new parent in one operation.
Diagnostics and status
Run health checks, queue status, server capability checks, usage recommendations, and context stats.
HTTP transport support
Run the server over HTTP instead of stdio when a client subprocess cannot get the right macOS permissions.
Client setup commands
Generate or write MCP client config for Claude Desktop and Claude Code with the `config` command.
How to get it
- 1Option B: config CLI
mcp-server-things config --client claude-desktop --write
- 2Run
claude mcp add-json things '{"command":"uvx","args":["--python-preference","only-managed","--python","3.12","mcp-server-things"]}' claude mcp add-json things '{"command":"uvx","args":["--python-preference","only-managed","--python","3.12","mcp-server-things"]}' -s user
README
Things 3 MCP Server
A Model Context Protocol (MCP) server that connects Claude and other AI assistants to Things 3 for natural language task management.
Why this server?
Writes go through AppleScript, not the Things URL scheme, which is what enables delete_todo, move_record/bulk_move_records, remove_tags, real IDs returned synchronously, and no Things auth token — the trade-off is a one-time macOS Automation permission prompt on first write. Operationally it also ships built-in doctor diagnostics, config --write client setup, context-optimized response modes for large databases, and 1300+ unit tests.
hald/things-mcp is a solid, lighter URL-scheme-based alternative — several of its ideas (Someday-project filtering, tag usage reporting, .mcpb packaging) are adopted here too. See docs/COMPARISON.md for the detailed matrix.
Prerequisites
- macOS 12+
- Things 3 installed and opened at least once
- uv (
brew install uv) - macOS will ask for Automation permission for Things 3 on the first write — that's expected (AppleScript is what enables delete/move operations other servers lack).
Install
Claude Desktop
Option A: One-click .mcpb
Download the latest .mcpb file from the releases page and double-click it to install into Claude Desktop.
The bundle launches the server via uvx, so uv must be installed and on PATH (brew install uv). The generated config pins uv's managed Python (--python-preference only-managed) so a stray Intel/Rosetta Python on your PATH can't break the install; first launch may download a managed CPython.
Option B: config CLI
mcp-server-things config --client claude-desktop --write
Safely adds/updates the things entry in your Claude Desktop config (--force overwrites an existing, different entry instead of refusing).
Option C: Manual JSON
{
"mcpServers": {
"things": {
"command": "uvx",
"args": ["--python-preference", "only-managed", "--python", "3.12", "mcp-server-things"]
}
}
}
Claude Code
claude mcp add-json things '{"command":"uvx","args":["--python-preference","only-managed","--python","3.12","mcp-server-things"]}'
claude mcp add-json things '{"command":"uvx","args":["--python-preference","only-managed","--python","3.12","mcp-server-things"]}' -s user
mcp-server-things config --client claude-code prints these exact commands.
Any MCP client
{
"command": "uvx",
"args": ["--python-preference", "only-managed", "--python", "3.12", "mcp-server-things"]
}
Verify
Run mcp-server-things doctor (or uvx mcp-server-things doctor) to confirm Things 3, permissions, and the database are all reachable.
Then ask your client "What's in my Things inbox?".
uvx mcp-server-thingsfails withBuilding cryptography==.../ maturin / Rust errors? Your default Python is an x86_64 (Intel/Rosetta) build —cryptographyno longer ships macOS x86_64 wheels. Fix: run with an arm64 interpreter, e.g.uvx -p 3.12 mcp-server-things(Homebrew or uv-managed Python), oruv python install 3.12first.mcp-server-things doctorwarns about this.
Advanced: pip, virtualenv, from source, existing installs
Upgrading from an existing install? See docs/UPGRADING.md.
Option 1: From PyPI
- Create and activate a virtual environment:
python3 -m venv venv
source venv/bin/activate # On macOS/Linux
- Install the package:
pip install mcp-server-things
Option 2: From Source (Development)
- Clone the repository:
git clone https://github.com/ebowman/mcp-server-things.git
cd mcp-server-things
- Create and activate a virtual environment:
python3 -m venv venv
source venv/bin/activate # On macOS/Linux
- Install dependencies:
pip install -r requirements.txt
- Install in development mode:
pip install -e .
Claude Desktop Configuration
config CLI
mcp-server-things config --client <claude-desktop|claude-code|generic> [--via uvx|current-python] [--write] [--force]
prints the MCP client configuration for the requested client (or, for
claude-desktop --write, safely merges it into
~/Library/Application Support/Claude/claude_desktop_config.json, backing up
the previous file first and refusing to clobber an existing, different
things entry unless --force is also passed). Run
mcp-server-things config --client claude-desktop --write or
mcp-server-things config --client claude-code instead of hand-editing JSON
or memorizing the claude mcp add-json syntax.
Shortcut for venv/pip installs: mcp-server-things config --client claude-desktop --via current-python targets the currently-running interpreter (sys.executable -m things_mcp) instead of the default uvx mcp-server-things.
For PyPI Installation
Add to your Claude Desktop configuration (~/Library/Application Support/Claude/claude_desktop_config.json):
{
"mcpServers": {
"things": {
"command": "/path/to/your/venv/bin/python",
"args": ["-m", "things_mcp"],
"env": {
"THINGS_MCP_LOG_LEVEL": "INFO",
"THINGS_MCP_APPLESCRIPT_TIMEOUT": "30"
}
}
}
}
For Source Installation
Add to your Claude Desktop configuration (~/Library/Application Support/Claude/claude_desktop_config.json):
{
"mcpServers": {
"things": {
"command": "/path/to/mcp-server-things/venv/bin/python",
"args": ["-m", "things_mcp"],
"env": {
"PYTHONPATH": "/path/to/mcp-server-things/src",
"THINGS_MCP_LOG_LEVEL": "INFO",
"THINGS_MCP_APPLESCRIPT_TIMEOUT": "30"
}
}
}
}
Notes:
- PyPI: Replace
/path/to/your/venv/bin/pythonwith your virtual environment's Python path - Source: Replace
/path/to/mcp-server-thingswith your actual installation path and include thePYTHONPATH - Use the full path to the Python executable in your virtual environment
- See Configuration section below for environment variable options
Creating tasks with natural language through Claude
📚 Documentation
- User Examples - Rich examples of how to use Things 3 with AI assistants
- Architecture Overview - Technical design and implementation details
- Troubleshooting - Common issues and solutions
Features
Core Todo Operations
- Create: Add todos with full metadata (tags, deadlines, projects, notes)
- Read: Get todos by ID, project, or built-in lists (Today, Inbox, Upcoming, etc.)
- Update: Modify existing todos with partial updates
- Delete: Remove todos safely
- Search: Find todos by title, notes, or advanced filters
Project & Area Management
- Get all projects and areas with optional task inclusion
- Create new projects with initial todos
- Update project metadata and status
- Create and rename areas, including tags (
add_area,update_area) - Organize todos within project hierarchies
Built-in List Access
- Inbox: Capture new items
- Today: Items scheduled for today
- Upcoming: Future scheduled items
- Anytime: Items without specific dates
- Someday: Items for future consideration
- Logbook: Completed items history
- Trash: Deleted items
Advanced Features
- Tag Management: Full tag support with AI creation control, plus usage reporting (
get_tag_usage) for weekly-review cleanup - Date-Range Queries: Get todos due/activating within specific timeframes
- URL Schemes: Native Things 3 URL scheme integration
- Health Monitoring: System health checks and queue status monitoring
- Error Handling: Robust error handling with configurable retries
- Logging: Structured logging with configurable levels
- Concurrency Support: Multi-client safe operation with operation queuing
- Input Validation: Configurable limits for titles, notes, and tags
- Structured Output: Every read tool returns both human-readable text and machine-readable
structured_content(via FastMCP 3.x) with a consistent{items, count, total, mode, limit, offset}shape ({item: {...}}for single-item lookups likeget_todo_by_id), so clients can consume results programmatically without re-parsing text
Requirements
- macOS: This server requires macOS (tested on macOS 12+)
- Things 3: Things 3 must be installed and accessible
- Python: Python 3.8 or higher
- Permissions: AppleScript permissions for Things 3 access
Quick Start
Once installed, Claude (or other MCP clients) can automatically discover and use all available tools. No additional setup required.
Configuration
The server uses environment variables for configuration, settable via system environment variables or a .env file (auto-loaded from the current directory, or pointed to with --env-file). The env vars that matter most:
| Variable | Default | Description |
|---|---|---|
THINGS_MCP_LOG_LEVEL | INFO | Logging level: DEBUG, INFO, WARNING, ERROR, CRITICAL |
THINGS_MCP_AI_CAN_CREATE_TAGS | false | Whether the AI can create new tags (false = existing tags only) |
THINGS_MCP_APPLESCRIPT_TIMEOUT | 30.0 | AppleScript execution timeout in seconds (1-300) |
THINGS_MCP_TRANSPORT | stdio | Transport to use: stdio or http |
THINGS_MCP_PORT | 8000 | Port to bind to when THINGS_MCP_TRANSPORT=http |
See .env.example for the full list of options, including validation limits, retry counts, the auth-token file, and THINGS_MCP_HOST.
Things URL-scheme auth token
add_checklist_items, prepend_checklist_items, and replace_checklist_items
(and any other tool built on things:///update, e.g. moving a to-do under a
heading) require a Things URL-scheme auth token. Without one configured,
these tools return success: false with an actionable error instead of
silently doing nothing - Things itself rejects un-authenticated update
requests, but open -g still exits 0, so the failure has to be caught before
the URL is ever opened. things:///add-based tools (add_todo,
add_project, including todo creation with a checklist) do not need a
token.
To configure one:
- In Things 3: Settings > General > Enable Things URLs > Manage.
- Save the token to one of (checked in this order, first match wins):
.things-authin the project rootthings-auth.txtin the project root~/.things-authin your home directory
- No restart required - the token is loaded at startup, then reloaded from disk automatically the next time an auth-gated tool is called while no token is currently loaded, so a token file added or fixed after the server starts is picked up on its next use. Once a token is successfully loaded it stays loaded for the life of the process.
An empty or whitespace-only token file is treated the same as a missing one
(the loader falls through to the next candidate path). health_check and
get_server_capabilities report the current auth_token_configured state,
and an AUTH_TOKEN_NOT_CONFIGURED error includes a checked_paths field
showing which candidate paths were checked and why each was rejected
(missing / empty / unreadable) - never the token value itself.
HTTP Transport
By default the server speaks MCP over stdio. It can optionally run an HTTP transport instead, which is the reliable fix when a client's stdio subprocess lacks Automation (TCC) access to Things 3: run the server from a Terminal that has been granted access, then point the client at the HTTP URL instead of launching it as a subprocess (see Troubleshooting for details).
| Variable | Default | Description |
|---|---|---|
THINGS_MCP_TRANSPORT | stdio | Transport to use: stdio or http |
THINGS_MCP_HOST | 127.0.0.1 | Host to bind to when THINGS_MCP_TRANSPORT=http |
THINGS_MCP_PORT | 8000 | Port to bind to when THINGS_MCP_TRANSPORT=http |
THINGS_MCP_TRANSPORT=http THINGS_MCP_PORT=8000 uvx mcp-server-things
Then add it to Claude Code as an HTTP server:
claude mcp add --transport http things http://127.0.0.1:8000/mcp
--transport, --host, and --port CLI flags are also available and take
precedence over the environment variables above.
Command Line Options
The server supports several command-line options:
# Start with debug logging
python -m things_mcp --debug
# Use a custom .env file
python -m things_mcp --env-file ~/my-config.env
# Check system health
python -m things_mcp --health-check
# Test AppleScript connectivity
python -m things_mcp --test-applescript
# Show version
python -m things_mcp --version
# Customize timeout and retry settings
python -m things_mcp --timeout 60 --retry-count 5
# Run with HTTP transport instead of stdio
python -m things_mcp --transport http --host 127.0.0.1 --port 8000
Claude Desktop Environment Variables
You can set environment variables directly in your Claude Desktop configuration:
{
"mcpServers": {
"things": {
"env": {
"THINGS_MCP_LOG_LEVEL": "DEBUG",
"THINGS_MCP_AI_CAN_CREATE_TAGS": "true",
"THINGS_MCP_APPLESCRIPT_TIMEOUT": "60"
}
}
}
}
Available MCP Tools
Todo Management
get_todos(project_uuid?, include_items?)- List todosadd_todo(title, ...)- Create new todoupdate_todo(id, ..., heading?, list_id?, list_title?)- Update existing todo;headingmoves it under a heading (requires the Things URL-scheme auth token),list_id/list_titlemove it to a different project or areabulk_update_todos(todo_ids, ...)- Update multiple todos in one operationget_todo_by_id(todo_id)- Get specific todo. A to-do/heading that is itself untrashed but whose containing project is trashed reportstrashed: trueplustrashedViaParent: true(Things marks only the trashed container, not its descendants); direct trash keepstrashed: truealone.delete_todo(todo_id)- Delete a to-do or a project (auto-detects the id type; headings/areas/tags cannot be deleted via any public Things 3 API)
Project Management
get_projects(include_items?)- List projectsadd_project(title, ..., todos?)- Create new project; a##-prefixed line intodoscreates a real heading (via the Things URL scheme'sjsonaction), with subsequent lines nesting under itupdate_project(id, ...)- Update existing project. Completing a project (completed="true") cascades to its child to-dos (Things marks them completed too), but reopening a project (completed="false") does not cascade back - child to-dos already completed stay completed. This is upstream Things behavior, not a bug here; reopen specific to-dos explicitly (e.g.bulk_update_todos) if needed.get_project_headings(project_id, mode?)- Read a project's heading structure (title, order, open-todo count per heading), in Things' own order. Read-only: headings can only be created at project-creation time (add_project's##lines) and cannot be renamed/deleted via any public Things 3 API.
Area Management
get_areas(include_items?)- List areasadd_area(title, tags?)- Create new areaupdate_area(id, title?, tags?)- Update existing area
List Access
get_today, get_upcoming, get_anytime, get_someday, and get_trash never return headings, and exclude projects by default - pass include_projects=true to also include projects that belong to that list (e.g. a project due today), matching the Things app's list views. get_inbox has no include_projects flag since the Inbox can never contain projects.
get_inbox()- Get Inbox todosget_today(include_projects?)- Get Today's todosget_upcoming(days?, include_projects?)- Get upcoming todos (with optional days filter)get_anytime(include_projects?)- Get Anytime todosget_someday(include_project_tasks?, include_projects?)- Get Someday todos. By default only returns items whose own start state is Someday; passinclude_project_tasks=trueto also include tasks that live inside Someday projects (markedinheritedSomeday: true). Today/Anytime/Upcoming always exclude tasks that belong to a Someday project, regardless of this flag.get_logbook(limit?, period?, offset?, include_canceled?)- Get completed todos; includes canceled todos by default (include_canceled=true), matching the Things app's own Logbook viewget_trash(include_projects?)- Get trashed todos
Date-Range Queries
get_due_in_days(days, include_overdue?, mode?, limit?)- Get todos due within specified days; includes already-overdue todos by default (include_overdue=true)get_activating_in_days(days, mode?, limit?)- Get todos activating within days
Search & Tags
search_todos(query, status?, offset?)- Basic search; matches incomplete todos by default (status='incomplete'), passstatus=Noneto search all statusessearch_advanced(...)- Advanced search with filters; searches all statuses by default when nostatusfilter is givenget_tags(include_items?)- List tagsget_tag_usage(only_unused?, mode?)- Per-tag open/total/area usage counts, sorted by usage, for cleanup. Caveats: tags sharing an identical title are merged into one row (uuid picks the last match), and area-only tags are counted viaarea_count/total_countbut never affectopen_count.create_tag(name)- Create a new tagget_tagged_items(tag)- Get items with specific tagadd_tags(todo_id, tags)- Add tags to a todoremove_tags(todo_id, tags)- Remove tags from a todoget_recent(period, status?, type?)- Get recently created items; returns all statuses and both to-dos and projects by default (headings are never included unlesstype='heading'is passed explicitly)
Bulk Operations
move_record(record_id, to_parent_uuid)- Move single recordbulk_move_records(record_ids, to_parent_uuid)- Move multiple records
System & Utilities
health_check()- Check server and Things 3 statusqueue_status()- Check operation queue status and statisticsget_server_capabilities()- Get server features and configurationget_usage_recommendations()- Get usage tips and best practicescontext_stats()- Get context-aware response statistics
Troubleshooting
Run mcp-server-things doctor first. It's a read-only diagnostic that checks
Things 3 installation, whether it's running, macOS Automation permission,
database readability (Full Disk Access/TCC), uv/uvx availability, the
optional auth token, and environment/version info - printing a PASS/FAIL/WARN
table with a one-line fix hint per row (exits non-zero only if something
actually needs fixing). Use mcp-server-things doctor --json for
machine-readable output, or python -m things_mcp doctor if you're running
from source.
Reads fail but writes work ("unable to open database file")
| Operation | Result |
|---|---|
Read tools (get_today, get_inbox, search_todos, ...) | Fail instantly with unable to open database file |
Write tools (add_todo, update_todo, ...) | Work normally |
| URL-scheme features needing the auth token | Also fail (the token is read from the same database) |
Cause: Under Claude Desktop, MCP servers are spawned via
/Applications/Claude.app/Contents/Helpers/disclaimer, which disclaims TCC
(privacy/permissions) responsibility for the process it launches. As a
result, the spawned server does not inherit Claude Desktop's Full Disk
Access grant, even though Claude.app itself has it. The Things 3 SQLite
database lives under the TCC-protected
~/Library/Group Containers/JLMPQHK86H.com.culturedcode.ThingsMac/ThingsData-*/Things Database.thingsdatabase/main.sqlite
(confirmed via things.database.Database().filepath - the same read path
this server uses for every read tool), so any process without Full Disk
Access gets unable to open database file immediately. Granting Full Disk
Access to Claude.app does not fix this, since the disclaimer helper is
what actually launches the server process.
Fix ladder (try in order):
- Recommended: Run the server with HTTP transport from a Terminal,
which already has disk access, instead of letting Claude Desktop spawn it
directly:
Then point Claude Code at it:THINGS_MCP_TRANSPORT=http THINGS_MCP_PORT=8000 uvx mcp-server-things
Stdio-only clients (including Claude Desktop) can bridge to the HTTP server withclaude mcp add --transport http things http://127.0.0.1:8000/mcpmcp-remote:npx mcp-remote http://127.0.0.1:8000/mcp - Grant Full Disk Access directly to the actual launched binary (e.g. the
uvxexecutable or the venvpythonin uv's cache) via System Settings -> Privacy & Security -> Full Disk Access. This works but is fragile - the grant can silently break when Homebrew/uv updates the binary, and the FDA picker sometimes greys out these paths, requiring drag-and-drop from Finder to add them. - Run
mcp-server-things doctorto confirm the fix - the "Database readable" row reports PASS once Full Disk Access (or the HTTP transport workaround) is in place.
This is the same failure mode reported upstream in
hald/things-mcp#62; we've
verified it applies here too since we read the Things database via the same
things.py library and code path.
Checklist tools return "Things URL-scheme auth token not configured"
add_checklist_items, prepend_checklist_items, and replace_checklist_items
need a Things URL-scheme auth token (see "Things URL-scheme auth token" under
Configuration above). mcp-server-things doctor reports this as a WARN on
the "Auth token file" row, naming the affected tools, until a token file is
in place.
Common Issues
Permission Denied Errors
# Grant AppleScript permissions to your terminal/IDE
# System Preferences > Security & Privacy > Privacy > Automation
# Enable access for your terminal application to control Things 3
Things 3 Not Found
# Verify Things 3 is installed and running
python -m things_mcp.main --health-check
# Check if Things 3 is in Applications folder
ls /Applications/ | grep -i things
Connection Timeouts
# Increase timeout value via environment variable
export THINGS_MCP_APPLESCRIPT_TIMEOUT=60
# Or in your .env file
THINGS_MCP_APPLESCRIPT_TIMEOUT=60
Debug Mode
# Enable debug logging
python -m things_mcp.main --debug
# Check logs
tail -f things_mcp.log
Health Diagnostics
# Comprehensive health check
python -m things_mcp.main --health-check
# Test specific components
python -m things_mcp.main --test-applescript
Boot diagnostics
If the server appears to hang before an MCP client can connect (especially on a cold start), the process writes timestamped boot-phase markers to stderr:
things-mcp boot: 2026-07-20T09:00:00.000+00:00 +0.001s process-start
things-mcp boot: 2026-07-20T09:00:00.010+00:00 +0.011s watchdog-armed (25.0s)
things-mcp boot: 2026-07-20T09:00:00.050+00:00 +0.051s things-import-start
things-mcp boot: 2026-07-20T09:00:00.120+00:00 +0.121s things-import-done
A one-shot startup watchdog also runs in the background: if boot doesn't
complete the MCP handshake within the deadline, it dumps every thread's stack
to stderr (Timeout (0:00:25)! followed by a traceback for each thread). On a
healthy, long-running server this fires exactly once, at the deadline, and is
harmless - it's stderr-only and does not affect the MCP stdio protocol (which
only uses stdout).
Relevant environment variables:
# Startup watchdog deadline in seconds. 0 (or any value <= 0) disables it.
THINGS_MCP_BOOT_WATCHDOG_SECS=25
# Timeout for lazily importing the third-party `things` package, in seconds.
# 0 (or any value <= 0) makes the import unbounded (blocking).
THINGS_MCP_THINGS_IMPORT_TIMEOUT_SECS=10
To diagnose a cold-start hang from a client's debug log: find the last
things-mcp boot: marker line - the phase named there is where boot stalled.
If a watchdog stack dump follows, its traceback shows exactly where each
thread was blocked at that moment.
Performance
- Startup Time: Less than 2 seconds
- Response Time: Less than 500ms for most operations
- Memory Usage: 15MB baseline, 50MB under concurrent load
- Concurrent Requests: Serialized write operations to prevent conflicts
- Throughput: Multiple operations per second depending on complexity
- Queue Processing: Less than 50ms latency for operation enqueuing
Security
- No network access required (local AppleScript only)
- No data stored outside of Things 3
- Minimal system permissions needed
- Secure AppleScript execution with timeouts
- Input validation on all parameters
Contributing
Contributions are welcome! Please follow these guidelines:
- Set up a virtual environment and install dependencies
- Follow existing code style and patterns
- Add tests for new features
- Submit pull requests with clear descriptions
Testing: pytest tests/unit runs entirely against mocked AppleScript/things.py
calls and needs no Things 3 install. pytest tests/integration is mostly mock-based
too; its real_things_tools/cleanup_test_todos fixtures, plus the local fixtures in
test_bulk_operations_comprehensive.py, test_search_comprehensive.py, and
test_search_performance.py (like the opt-in tests/live smoke suite) all require a
real, running Things 3 and are gated behind THINGS_MCP_LIVE_TESTS=1 - without it they
skip with a clear reason rather than writing to your live database. Run the live smoke
suite with make test-live (or
THINGS_MCP_LIVE_TESTS=1 pytest tests/live -q); it creates and cleans up its own
throwaway project and never touches pre-existing data. tests/regression is a second,
opt-in live suite (same THINGS_MCP_LIVE_TESTS=1 gate) that drives the real MCP tool
boundary end-to-end - run it with make test-regression. See
docs/TESTING.md for the full testing policy.
Documentation
- Troubleshooting Guide - Common issues and solutions
- Development Roadmap - Implementation status and missing features
License
This project is licensed under the MIT License - see the LICENSE file for details.
Support
- Issues: GitHub Issues
- Discussions: GitHub Discussions
- Email: ebowman@boboco.ie
Built for the Things 3 and MCP community.
Files in the repo
- .agents
- .beads
- .claude
- .codex
- .github
- docs
- scripts
- src
- tests
- .env.example
- .gitignore
- AGENTS.md
- CHANGELOG.md
- CLAUDE.md
- CODE_OF_CONDUCT.md
- CONTRIBUTING.md
- LICENSE
- Makefile
- manifest.json
- pyproject.toml
- pytest.ini
- README.md
- requirements.txt
- SECURITY.md
Discussion (0)
Ask about usage, or say what you built with itSign in to join the discussion.
No comments yet. Be the first to say what this is good for.
More connectors
High-performance code intelligence MCP server. Indexes codebases into a persistent knowledge graph — average repo in milliseconds. 158 languages, sub-ms queries, 99% fewer tokens. Single static binary, zero dependencies.

Universal provider proxy for OpenAI Codex & Claude Code — use any LLM (Claude, Gemini, Grok, DeepSeek, Ollama…) with Codex CLI, App, SDK, and Claude Code
Git-native persistent memory for AI coding agents. Implements Google OKF v0.2 with sub-300µs in-memory BM25 search, embedded MCP server, and progressive disclosure. Slashes token bloat by 80% with zero external databases or dependencies. Built in pure Go.
Local-first code intelligence graph for MCP and CLI. Builds a persistent map of your codebase so AI coding tools read only what matters, with benchmarked context reductions on reviews and large-repo workflows.
Stop your AI from making things up — it proposes, deterministic tools decide, every claim checked against ground truth with evidence. Grounded facts and context survive resets. Reverse engineering is the proving ground. MCP server + CLI.