brain-init

name	brain-init
description	Bootstrap your AI Brain from connected tools and Claude's memory. Zero-input onboarding — discovers your connectors, pulls meta-knowledge, and saves it automatically.

Brain Init

Bootstrap your AI Brain by automatically discovering what tools you have connected and extracting durable meta-knowledge from them.

Prerequisites

The AI Brain connector must be available. If mcp__ai-brain__capture_thought and mcp__ai-brain__search_thoughts MCP tools are not available, stop and tell the user to connect AI Brain first.

Workflow

Step 1: Check Brain Status

Call mcp__ai-brain__get_stats to see if the brain already has content.

• If the brain has thoughts, tell the user: "Your brain already has [N] thoughts. Running brain-init will add new knowledge without duplicating what's already there. Proceeding..."
• If the brain is empty, tell the user: "Setting up your AI Brain for the first time. I'll scan your connected tools and build your knowledge base automatically."

Step 2: Discover Connectors

Enumerate available MCP tools by checking what's loaded in this session. Look for these patterns:

Connector	Tool patterns to look for	What it tells us
Email	email_search, outlook_email_search, gmail_*	Communication patterns, key contacts
Calendar	calendar_*, google_calendar_*, outlook_calendar_*	Meeting rhythm, team structure
ClickUp	clickup_*, get_task, search_tasks	Projects, responsibilities
GitHub	GitHub MCP tools or gh CLI available	Repos, collaborators
Slack	slack_*, send_message, search_messages	Team context, channels
Linear	linear_*	Projects, issue tracking
Jira	jira_*	Projects, issue tracking

Report which connectors were found: "I found connections to: [list]. I'll use these to learn about your work."

If no connectors beyond AI Brain are available, skip to Step 4 (Claude Memory) and then Step 5 (Fallback Questions).

Step 3: Pull Meta-Knowledge from Connectors

For each available connector, extract durable meta-knowledge — not transient task data.

For email/communication tools:

• Search recent emails (last 14 days) to identify the 5-10 most frequent contacts
• Note relationships: who do they report to? who reports to them? who do they collaborate with?
• Do NOT save email content — just relationship patterns

For calendar:

• List events from the last 14 days
• Identify recurring meetings: name, frequency, attendees
• Infer: team structure, work rhythm, role (e.g., "has 3 direct report 1:1s = likely a manager")
• Do NOT save individual event details — just patterns

For project management (ClickUp/Linear/Jira):

• List spaces/projects the user is active in
• Identify what they're assigned to most
• Note project names and their apparent purpose
• Do NOT save individual task details

For GitHub:

• List repos with recent activity
• Note primary languages, collaborators
• Identify PR review patterns (who reviews whose code?)

For Slack:

• List channels the user is most active in
• Note frequent conversation partners
• Do NOT save message content

Compile findings into structured notes organized by: role signals, key people, active projects, work patterns.

Step 4: Import Claude Memory

Check for existing knowledge Claude has about this user:

1. Read ~/.claude/CLAUDE.md if it exists — this contains user-stated preferences and instructions
2. Read memory files from ~/.claude/projects/*/memory/ — these contain stored memories from previous sessions
3. Draw on conversation context — what Claude already knows from prior sessions

Organize findings into: people, projects, preferences, decisions, recurring topics.

Step 5: Fallback Questions (only if no connectors found)

If no connectors beyond AI Brain were discovered in Step 2, ask these 3-4 quick questions:

1. "What's your role? (e.g., frontend engineer, product manager, founder)"
2. "What are you mainly working on right now? (1-3 projects)"
3. "Who do you work with most closely? (2-5 people and their roles)"

Use the answers as the basis for Step 6 instead of connector data.

Step 6: Synthesize and Save

Consolidate all sources into focused brain thoughts. Before saving each thought, call mcp__ai-brain__search_thoughts with the topic to check for duplicates.

Note: search_thoughts returns a compact index — {id, summary, snippet, type, topics, score}. If a candidate looks like a duplicate from its summary + snippet, call mcp__ai-brain__get_thoughts with the candidate's id to fetch full content and confirm before deciding.

Thoughts to create:

1. About me — Role, responsibilities, what I work on, communication style, tools I use. Format: "About me: [role] at [company if known]. Responsibilities: [list]. Primary tools: [list]. Communication style: [preferences from CLAUDE.md or inferred]."

2. My team — Key people, their roles, how we work together. Format: "My team: [Person] ([role]) — [relationship/how we work together]. [repeat for each key person]."

3. Active projects — Current focus areas with context. Format: "Active projects: [Project 1] — [what it is, my role in it]. [Project 2] — [description]. Priority order: [if determinable]."

4. Work patterns — Meeting rhythm, schedule patterns, preferences. Format: "Work patterns: [recurring meetings]. Typical schedule: [if determinable]. Preferences: [from CLAUDE.md or inferred]."

Save each via mcp__ai-brain__capture_thought. capture_thought returns the new thought's thoughtId — collect these so Step 7 can cite them.

Additionally: If enough signal exists to identify project priorities, create a pinned goal list via mcp__ai-brain__create_list with the top projects, then call mcp__ai-brain__update_list to pin it.

Step 7: Report

Show the user a summary of what was captured, with each item cited as thought:<id> so they can trace provenance:

• Which connectors were scanned
• What was saved — one bullet per thought, each cited as thought:<id>
• Whether a goals list was created (and its id)
• Next steps: "Try /brain-thread <topic> to trace how your thinking on a theme evolves, or /brain-context <date> to restore what was on your mind at a specific time."

End with: "Does this look right? Anything missing or incorrect?"