ADHD executive function enforcement for Claude Desktop.
An MCP server that transforms task management from a social contract
into a stateful boundary Claude cannot cross without explicit tool invocation.

Python 3.11+ · MIT License · 96 tests · 14 tools

ADHD developers don't lack ideas — they lack friction between ideas. A single "while we're at it" can derail an hour of focused work into a rabbit hole that felt productive but shipped nothing.

Task Anchor adds that friction mechanically. It creates a task lock that Claude must respect, runs drift detection on every message, and forces explicit validation before a task can be marked done. Ideas that surface mid-task are parked, not lost — they go into a safe queue you can review later.

The system also tracks emotional state across sessions. If you leave frustrated and come back the next morning, it knows, and it offers re-entry strategies instead of just dumping you back into the same wall.

You say something → drift_detect scores it → drift? → park the idea, redirect
                                             clear? → continue working
                                                      ↓
                                              scope_validate_edit → in scope? → proceed
                                                                   out of scope? → block + offer options
                                                      ↓
                                              task_complete → evidence matches exit condition? → release lock
                                                             doesn't match? → reject, keep lock

Claude is required to follow these rules (enforced via .claude/CLAUDE.md):

1. No code without a lock. Claude refuses all coding help until task_lock_create defines what you're building, the exit condition, and the file scope.
2. Drift detection on every message. Every user input is scored against 26 weighted signal phrases (e.g., "while we're at it" = 5 points, "actually" = 2 points). Score ≥ 4 triggers an automatic park-and-redirect.
3. Scope enforcement before edits. Before modifying any file, Claude checks it against the locked scope. Out-of-scope edits are blocked.
4. Session checkpoints before signing off. Emotional state, next micro-action, and blocker notes are captured so the next session can resume intelligently.

Python 3.11 or later. No other system dependencies.

cd mcp-server
pip install -e .

Or install just the runtime dependency without the package:

pip install mcp

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "task-anchor": {
      "command": "python",
      "args": ["-m", "task_anchor.server"],
      "cwd": "/absolute/path/to/TaskAnchor/mcp-server"
    }
  }
}

If installed via pip:

{
  "mcpServers": {
    "task-anchor": {
      "command": "task-anchor"
    }
  }
}

Restart Claude Desktop after saving.

All user-facing messages route through a configurable tone layer. Same enforcement logic, different voice.

Strict — the original enforcement language.

⚓ DRIFT DETECTED (Score: 5/10)
ACTION REQUIRED: Call parked_add to capture this idea...
BINARY CHOICE:
[1] Park this idea and continue current task
[2] Mark current complete and switch (requires validation)

Supportive (default) — warm coaching voice, acknowledges effort, preserves agency.

⚓ New thread detected (score: 5/10)
That sounds like a separate idea — and it might be a good one.
Let me save it so you don't lose it.
What would you like to do?
[1] Park this idea — I'll save it, and we keep going
[2] This IS more important — let's switch (finish current first)

Minimal — facts only, shortest possible output.

Drift (score 5/10): "while we're at it let's..."
[1] Park  [2] Switch

Switch at any time with set_tone.

When you're in hyperfocus and the drift detection is getting in the way, activate flow mode:

flow_mode_activate(duration_minutes=45)

Drift detection is suspended. Scope enforcement stays active (safety net, not a cage). The mode auto-expires after the set duration and sends a gentle check-in. End early with flow_mode_deactivate.

Maximum duration is 120 minutes — even hyperfocus benefits from periodic check-ins.

mcp-server/
├── pyproject.toml
└── task_anchor/
    ├── config.py           — path resolution (immune to cwd, env-overridable)
    ├── models.py           — TaskLock dataclass, drift signal weights + thresholds
    ├── storage.py          — atomic file I/O, cross-platform locking (fcntl/msvcrt)
    ├── drift.py            — scoring engine, completion validation, history logging
    ├── flow.py             — flow mode activate/deactivate/auto-expire
    ├── tone.py             — tone persistence + message resolver
    ├── messages.py         — message registry (aggregator)
    ├── messages_core.py    — templates: lock, drift, parked, scope
    ├── messages_session.py — templates: completion, session, flow, celebration
    ├── helpers.py          — shared utilities (git branch, load lock, session log)
    ├── streak.py           — daily streak tracking, completion celebration
    ├── tools.py            — MCP tool schema definitions (14 tools)
    ├── handlers.py         — tool handler coroutines
    └── server.py           — MCP wiring, route table, entry point

All state lives in .claude/skills/task-anchor/ inside the repo root. Override with TASK_ANCHOR_DIR.

Why MCP, not a prompt injection? Prompt injections are social contracts — Claude can be talked out of them. MCP tools are stateful boundaries. Claude literally cannot mark a task complete without calling task_complete, which validates evidence against the exit condition before releasing the lock.

Why word-boundary regex instead of substring matching for drift detection? "Rewrite" appearing inside "overwrite" is not a drift signal. "Instead" appearing inside "instantiate" is not a drift signal. Whole-word boundary matching prevents false positives from normal technical language.

Why naive stemming instead of a real NLP library? Zero additional dependencies. The stemmer handles the 90% case (plurals, -ing, -ed, -tion, -ly) that matters for matching exit conditions like "test passes" against evidence like "tests passed." A full NLP stack would add weight for marginal accuracy gain.

Why three tones? ADHD is not one experience. Some people respond to external structure ("VIOLATION: Cannot proceed"). Others find that language triggering, especially those with rejection sensitive dysphoria. A configurable tone means the same enforcement logic works for different brains.

cd mcp-server
pip install -e ".[test]"
pytest tests/ -v

96 tests across 4 test files covering drift scoring, completion validation, model serialisation, scope validation, session filtering, tone switching, flow mode, and all 14 handler coroutines.

MIT