vigil-mcp 0.2.0

Proactive memory layer for AI — the AI that reaches back out to you

🔥 Vigil MCP

The AI that reaches back out to you.

Most AI tools wait. You open them, you ask, they answer. Vigil flips that.

Connect it to Claude once — and it remembers what you said you'd do, reasons about when it's worth interrupting you, and follows up on its own.

---

The problem

Every AI interaction ends the same way: you close the tab and the context disappears. The AI has no persistence, no agency, no follow-through. It's reactive by design.

Vigil is a layer that changes that.

---

How it works

Vigil is an MCP server — a tool layer that any compatible AI can plug into. When connected:

1. Claude detects commitments automatically — "I'll come back tomorrow", "remind me to push this on Friday", "I'll do X after lunch" → saved instantly, no command needed
2. A policy engine runs in the background — every 60 seconds, it evaluates every pending commitment against 7 conditions before deciding whether to act
3. When you open a new session, Claude reads vigil://pending and opens with urgent items naturally — "Hey, you said you'd finish the README. Still on it?"
4. Desktop and push notifications fire when things are overdue and you haven't opened Claude yet
5. Nudges go to Slack or Discord if you've configured a webhook
6. V3 integrations run every 5 minutes — idle detection, adaptive learning, and weekly reports work silently in the background
7. Every decision is logged and explainable — ask "why did you message me?" and get the full trace

---

Policy engine

Vigil doesn't act on a timer. It reasons.

Every potential nudge passes through a 7-step decision chain:

1. Hypothetical intent?              → never act
2. Grace period (<30 min old)?       → too fresh, skip
3. Early phase (>48h to deadline)?   → too soon, skip
4. Cooldown window active?           → avoid spam (4h / 1h / 2h by phase)
5. Daily attention budget exhausted? → suppress unless importance ≥ 0.8
6. Soft intent?                      → reduce effective importance by 0.2
7. Phase + importance threshold      → fire or stay silent

Every decision — including suppressed ones — is written to decision_log with phase, urgency, budget used, and full reasoning text.

Temporal phases

Phase	Window	Cooldown	Behavior
early	> 48h until due	—	No nudge
execution	4–48h until due	4h	Nudge if importance ≥ 0.6
risk	< 4h until due	1h	Always nudge
overdue	Past deadline	2h	Escalate

Auto-decay

Nudged 3+ times with no response → importance decays 15% per cycle. After 6 ignored nudges → auto-archived. No noise, no spam.

---

Features

Feature	Description
🧠 Policy engine	7-step reasoning chain, not a simple timer
💬 Natural language input	"remind me to push the PR before EOD" → parsed and saved automatically
📬 Proactive notifications	Native macOS/Windows/Linux OR Web Push to phone/browser
🔔 Slack/Discord integration	Nudges delivered to your Slack or Discord channel
🤖 Slack slash commands	/vigil remind me to X creates a commitment from Slack
📱 Installable PWA	Add dashboard to phone home screen — iOS 16.4+ / all Android
☁️ Hosted mode	One-click Railway/Render deploy, auth-protected, multi-user
✏️ Edit commitments	Update deadline, importance, intent, or description after saving
🔁 Recurring commitments	Daily, weekly (specific days), monthly — auto-reschedules
📉 Auto-decay	Importance drops on ignores; auto-archives after 6 nudges
🌅 Daily digest	Optional morning summary notification
🖥️ Web dashboard	Dark-theme UI at localhost:7734 (or your hosted URL)
🔍 Full explainability	Every decision logged with reasoning — "Why did you message me?"
💤 Idle detection	Nudges escalate when no Claude session detected for 4+ hours
📊 Adaptive policy	Learns your follow-through patterns per category, boosts flagging commitments
📋 Reliability reports	Weekly Monday summary: follow-through % by category

---

Architecture

Claude Desktop
    │
    ├── Tools     (save / create_from_text / get / complete / snooze /
    │              update / explain / budget / digest / set_slack_webhook /
    │              set_discord_webhook / list_features / enable_feature /
    │              disable_feature)
    └── Resources (vigil://pending · vigil://status · vigil://history)

vigil-mcp process
    ├── server.py       — FastMCP server, tools, resources, background loop
    ├── policy.py       — Decision engine (phase → urgency → budget → action)
    ├── db.py           — SQLite layer (thread-local, multi-user, migrations)
    ├── dashboard.py    — HTTP dashboard + PWA + CRUD API (localhost:7734 / $PORT)
    ├── api.py          — FastAPI REST layer (localhost:7735) + cloud entrypoint
    ├── analyzer.py     — Signal analysis + NL commitment parsing (Claude API)
    ├── webhooks.py     — Slack/Discord outbound nudges + slash command handler
    ├── push.py         — Web Push delivery (VAPID + pywebpush)
    ├── pwa.py          — Manifest, service worker, icons
    ├── notify.py       — Shared notification delivery (push → webhooks → desktop)
    ├── features.py     — V3 feature flag registry (7 toggleable features)
    ├── logger.py       — Rotating structured log at ~/.vigil/vigil.log
    └── integrations/
        ├── __init__.py              — Fan-out runner (importlib-based, fault-isolated)
        ├── idle_detection.py        — Live: session heartbeat + idle escalation
        ├── adaptive_policy.py       — Live: per-category follow-through learning
        ├── reliability_reports.py   — Live: weekly Monday digest by category
        ├── github_integration.py    — Stub: stale PR commitment creation
        ├── calendar_sync.py         — Stub: pre-meeting commitment surfacing
        ├── accountability_partners.py — Stub: Slack DM to accountability partner
        └── shared_commitments.py    — Stub: team-visible commitments via Slack

~/.vigil/
    ├── commitments.db  — All data (commitments, decision_log, settings)
    └── vigil.log       — Structured log (grep-friendly)

---

Tools

Tool	What it does
create_from_text	Parse natural language into a commitment automatically
save_commitment	Capture a commitment with explicit fields (due date, intent, importance, recurrence)
get_commitments	List all pending with phase, state, and stats
complete_commitment	Mark done; auto-reschedules if recurring
snooze_commitment	Push deadline to a later time
update_commitment	Edit description, deadline, importance, intent, or summary
explain_nudge	Full decision trace for a commitment
set_attention_budget	Set daily nudge cap (default: 10)
set_daily_digest_time	Morning summary notification at HH:MM
set_slack_webhook	Route nudges to a Slack channel
set_discord_webhook	Route nudges to a Discord channel
list_features	Show all V3 features and their on/off state
enable_feature	Turn on a V3 feature by name
disable_feature	Turn off a V3 feature by name

Resources

Resource	Content
vigil://pending	Pending commitments grouped by phase — Claude reads this at session start
vigil://status	System health: budget used, config, stats
vigil://history	Completed commitments and follow-through rate

---

Commitment schema

{
  description:      str,
  due_by:           ISO 8601 datetime,
  ai_summary:       str,    # 2-sentence context of what was being worked on
  intent_strength:  "hard" | "soft" | "hypothetical",
  importance_score: 0.0–1.0,
  confidence_score: 0.0–1.0,
  state:            "captured" | "scheduled" | "in_progress" | "at_risk" | "overdue" | "completed" | "abandoned",
  recurrence:       "none" | "daily" | "weekly" | "monthly",
  recurrence_days:  ["monday", "thursday"],  # for weekly
  recurrence_time:  "HH:MM",
}

---

V3 Features

Vigil V3 adds three directions of capability behind toggleable feature flags. All are off by default — turn them on via Claude or the dashboard.

Enable via Claude

list_features()
enable_feature("idle_detection")
enable_feature("adaptive_policy")
enable_feature("reliability_reports")

Enable at startup (env var)

VIGIL_FEATURES=idle_detection,adaptive_policy,reliability_reports vigil

Feature registry

Feature	Direction	Status	Required env var
idle_detection	Context Awareness	Live	—
adaptive_policy	Adaptive	Live	—
reliability_reports	Adaptive	Live	—
github_integration	Context Awareness	Stub	GITHUB_TOKEN
calendar_sync	Context Awareness	Stub	GOOGLE_CALENDAR_TOKEN
accountability_partners	Multi-Person / Team	Stub	VIGIL_SLACK_BOT_TOKEN
shared_commitments	Multi-Person / Team	Stub	VIGIL_SLACK_BOT_TOKEN

Idle Detection

Tracks the last time Claude read vigil://pending (i.e. the last time you opened a session). If no session is detected for 4+ hours and you have overdue or at-risk commitments, Vigil fires a push/Slack/desktop nudge.

• Cooldown: 2 hours between idle nudges
• Threshold: VIGIL_IDLE_HOURS env var (default: 4)
• Resumes normal nudge schedule once you open a new session

Adaptive Policy

Runs once per day. Looks at your completed and abandoned commitments grouped by context category. Categories with ≥ 3 resolved commitments and < 50% follow-through are marked low-reliability.

Pending commitments in those categories get their importance_score boosted 15% (capped at 0.95) so they are more likely to clear the policy engine's threshold and generate a nudge. Every boost is written to decision_log — explain_nudge shows you exactly why.

Reliability Reports

Every Monday at 09:00 (configurable via VIGIL_REPORT_HOUR), Vigil sends a summary of the past week's follow-through, broken down by category:

Past 7 days: 7/10 done (70%)
3 still pending
By category:
  ✓ work: 4/5 (80%)
  ✓ personal: 2/3 (67%)
  ⚠ learning: 1/3 (33%)

Categories marked ⚠ are candidates for Adaptive Policy boosting. Report is deduped by ISO week — it won't re-fire if the process restarts.

---

Web dashboard

A local dashboard runs at http://localhost:7734 whenever Vigil is active.

• Natural language bar at the top — type "remind me to push the PR before EOD" and press Enter
• Stats: follow-through %, completed, pending, overdue, nudges today
• Commitments grouped by phase with colour-coded left borders
• "Why?" expandable button — full policy decision trace per commitment
• Integrations panel — paste a Slack or Discord webhook URL and save
• V3 Features panel — toggle switches for all 7 features, amber warning when env var is missing
• Auto-refreshes every 30 seconds
• Installable as a PWA (mobile)

---

Installation

Requirements: Python 3.10+, macOS / Windows / Linux

git clone https://github.com/prodbysilky/vigil-mcp
cd vigil-mcp
python3 -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

Option A — Claude Desktop (MCP) — simplest

pip install -e .

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "vigil": {
      "command": "/absolute/path/to/vigil-mcp/.venv/bin/vigil-mcp"
    }
  }
}

Or with uvx (no venv needed):

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

Restart Claude Desktop (Cmd+Q, then reopen). Vigil starts automatically.

Option B — Standalone / ChatGPT / any HTTP client

pip install -e ".[api]"   # adds FastAPI + uvicorn
vigil                     # starts dashboard + REST API, no Claude needed

The REST API runs at http://localhost:7735. Interactive docs at /docs. The dashboard runs at http://localhost:7734.

Option C — Hosted mode (public URL, mobile push, auth-protected)

Deploy once, visit from any device, install the PWA to your home screen, get push notifications on your phone. Works on iOS 16.4+ and all Android.

1. Generate VAPID keys (one-time):

pip install -e ".[hosted]"
python -c "from vigil_mcp.push import generate_vapid_keys; \
           priv, pub = generate_vapid_keys(); \
           print('VIGIL_VAPID_PRIVATE_KEY=' + repr(priv)); \
           print('VIGIL_VAPID_PUBLIC_KEY=' + pub)"

2. Deploy to Railway (or Render / Fly — all read $PORT):

• New project → Deploy from GitHub repo
• Set environment variables (see table below)
• Railway auto-detects Procfile + railway.json; healthcheck hits /manifest.json

3. Set a secret token (so only you can access the dashboard):

VIGIL_SECRET_TOKEN=your-secret-here

Your dashboard URL becomes: https://your-app.up.railway.app/?token=your-secret-here

4. Install on phone:

• Open https://your-app.up.railway.app/?token=your-secret-here in Safari (iOS) or Chrome (Android)
• "Add to Home Screen"
• Launch from home screen, tap "Enable notifications"

---

Environment variables

Variable	Required	Purpose
VIGIL_SECRET_TOKEN	Recommended for hosted	Bearer token for dashboard + API auth
ANTHROPIC_API_KEY	For NL parsing + signal analysis	Powers create_from_text and auto-commitment creation
VIGIL_VAPID_PRIVATE_KEY	For Web Push	Signs push notifications
VIGIL_VAPID_PUBLIC_KEY	For Web Push	Sent to browsers for subscription
VIGIL_VAPID_SUBJECT	For Web Push	mailto:[e-mail] (defaults to mailto:[e-mail])
VIGIL_SLACK_WEBHOOK	For Slack nudges	Incoming webhook URL (alternative to setting via Claude/dashboard)
VIGIL_DISCORD_WEBHOOK	For Discord nudges	Webhook URL (alternative to setting via Claude/dashboard)
VIGIL_SLACK_SIGNING_SECRET	For Slack slash commands	Verifies /vigil commands come from Slack
VIGIL_FEATURES	Optional	Comma-separated list of V3 features to auto-enable on startup
VIGIL_IDLE_HOURS	Optional	Hours of inactivity before idle nudge fires (default: 4)
VIGIL_REPORT_HOUR	Optional	Hour (0–23) for Monday reliability report (default: 9)
TIMEZONE	Optional	Local timezone for recurring scheduling (default: Europe/Budapest)
PORT	Set by Railway/Render	Cloud port — auto-detected, do not set manually
GITHUB_TOKEN	For github_integration feature	Personal access token with repo scope
GOOGLE_CALENDAR_TOKEN	For calendar_sync feature	OAuth token for Google Calendar API
VIGIL_SLACK_BOT_TOKEN	For accountability/shared features	Bot token with chat:write scope

---

Slack integration

Vigil sends nudges to Slack and accepts /vigil slash commands.

Outbound nudges (via webhook)

1. Go to api.slack.com/apps → Create New App → From Scratch
2. Incoming Webhooks → Enable → Add to Workspace → pick a channel → copy the URL
3. Set in Claude: set_slack_webhook("https://hooks.slack.com/services/...")
   Or paste it in the dashboard Integrations panel.

Inbound slash commands (/vigil remind me to X)

1. Slash Commands → Create New Command:
   • Command: /vigil
   • Request URL: https://your-app.up.railway.app/api/slack/command
   • Short Description: Create a Vigil commitment
2. Basic Information → Signing Secret → copy it → set VIGIL_SLACK_SIGNING_SECRET env var

Usage: /vigil remind me to push the PR before EOD

---

Discord integration

1. In Discord: Channel Settings → Integrations → Webhooks → New Webhook → copy URL
2. Set in Claude: set_discord_webhook("https://discord.com/api/webhooks/...")
   Or paste it in the dashboard Integrations panel.

Vigil sends rich embeds (amber colour) to the channel whenever a nudge fires.

---

Example flows

Natural language capture

You:   "remind me to push this PR before EOD"
Vigil: parses → description="Push PR", due_by="today 18:00", importance=0.7
       → saved. No form filling needed.

High-importance missed deadline

You:     "I'll push this to GitHub before end of day"
Vigil:   saves commitment, importance=0.7, due=17:00

[17:00 — no action yet]
Policy:  phase=overdue, urgency=0.72, budget=3/10
         → NOTIFY: desktop + Slack: "You said you'd push to GitHub"

[next Claude session]
Claude:  "Hey — your GitHub push is overdue. Want to do it now?"

Idle detection escalation

You:     "I'll review those PRs this afternoon"
Vigil:   saves commitment, due=17:00

[No Claude session opened since 13:00 — 4 hours idle]
Idle:    phase=overdue, no session detected
         → PUSH + Slack: "4h idle — PR review is overdue"

Adaptive policy in action

[End of day — adaptive_policy runs]
Policy:  "learning" category: 1/4 resolved (25% rate)
         → 3 pending "learning" commitments boosted +15% importance
         → Next policy cycle: importance 0.5 → 0.58, clears threshold
         → Nudge fires that would have been suppressed

Recurring commitment

You:     "Every Monday morning I want to review my job applications"
Vigil:   saves with recurrence=weekly, recurrence_days=["monday"], recurrence_time="09:00"

[Monday 09:00] → NOTIFY
You:     "Done" → complete_commitment → next Monday auto-scheduled

Weekly reliability report

[Monday 09:00]
Vigil:  "Past 7 days: 7/10 done (70%)
         ✓ work: 4/5 (80%)
         ✓ personal: 2/3 (67%)
         ⚠ learning: 1/3 (33%)"

---

REST API

The REST API runs at http://localhost:7735 when fastapi + uvicorn are installed (pip install vigil-mcp[api]). Interactive OpenAPI docs at /docs.

Method	Endpoint	Description
GET	/commitments	List all pending
POST	/commitments	Save a commitment (structured)
POST	/commitments/from-text	Create from natural language
GET	/commitments/{id}	Get one commitment
PATCH	/commitments/{id}	Edit fields
POST	/commitments/{id}/complete	Mark done
POST	/commitments/{id}/snooze	Snooze to new time
DELETE	/commitments/{id}	Archive
GET	/commitments/{id}/decisions	Policy decision trace
GET	/signals	Recent behavioral signals
POST	/signals	Ingest a signal
GET	/stats	Overall follow-through stats
GET	/settings	Current settings
PUT	/settings/budget	Set daily attention budget
PUT	/settings/digest	Set daily digest time
GET	/features	List V3 features
PUT	/features/{name}	Enable/disable a V3 feature

All endpoints accept Authorization: Bearer <VIGIL_SECRET_TOKEN> when a token is set.

---

Logs

tail -f ~/.vigil/vigil.log
grep "action=notify" ~/.vigil/vigil.log
grep "auto_abandoned" ~/.vigil/vigil.log
grep "commitment_id=3" ~/.vigil/vigil.log
grep "adaptive_boost" ~/.vigil/vigil.log
grep "idle_nudge_sent" ~/.vigil/vigil.log
grep "reliability_report_sent" ~/.vigil/vigil.log

---

Compatibility

Platform	How to connect	Notes
Claude Desktop	pip install -e . → MCP config	Plug-and-play, full tool support
Cursor / Cline / Zed	Same as Claude Desktop	Works out of the box
ChatGPT Custom GPT	Deploy hosted mode → point GPT Action at /openapi.json	Full CRUD via REST
iPhone / Android	Visit hosted URL → Add to Home Screen → enable notifications	PWA + Web Push, no app store
Slack	Incoming webhook (outbound) + slash command (inbound)	/vigil creates commitments
Discord	Incoming webhook	Nudges as rich embeds
Any web app / script	REST API at localhost:7735 or hosted URL	Standard JSON HTTP
No AI at all	vigil → open localhost:7734	Full dashboard UI — add/complete/snooze without any AI

---

Why this is different

• Not a reminder app — it reasons about when to interrupt you, not just what to remind you about
• Not a wrapper — the policy engine is the product
• Explainable — every decision is logged and queryable. Ask "why did you message me?" and get the full trace
• Self-tuning — adaptive policy adjusts importance scores based on your actual follow-through behavior
• Idle-aware — knows when you've been away too long and escalates accordingly
• Local-first — all data stays on your machine (SQLite in ~/.vigil/)
• Natural input — type plain English; Vigil figures out the structure
• Extensible — V3 feature flag system lets you add GitHub, calendar, and team features when you need them

---

Built with

• MCP Python SDK — FastMCP server
• Anthropic SDK — NL parsing via claude-haiku
• SQLite — local-first, thread-safe persistence
• Python stdlib only for webhooks — no extra dependencies for Slack/Discord delivery