Team Mode

Status: Phase 1 implemented — multi-user auth, role-based permissions, audit logging. Phase 2 (per-user identity anchors) in progress.

Team mode extends the-brain's remote mode to support multiple developers and autonomous agents sharing a single brain server. Each user gets their own token, identity, and permissions — allowing the brain to learn per-developer preferences while building shared team patterns.

Why Team Mode?

Single-user remote mode works: one developer, one daemon. But when a team of 5 developers all work on the same codebase, you want the brain to know:

Who made a correction (Oskar prefers tabs, Anna prefers spaces)
What is a team convention vs personal preference (use Redux = team, use xdist = Oskar)
When patterns emerge across the team (both Oskar and Anna independently fixed the same bug pattern)

Architecture

Setup

1. Server (Linux)

# Initialize in team mode
the-brain init --remote --team

# Creates:
#   - Default admin user
#   - config.json with server.mode = "team"
#   - users.db alongside brain.db

# Start daemon
the-brain daemon start

2. Add Team Members

# Add a developer to the cpv project
the-brain user add --name anna --project cpv --role contributor

# Output:
#   ✓ User "anna" created
#   API Token: mb_a1b2c3d4e5f6...

# Generate additional tokens (e.g., for CI)
the-brain user token --name anna --label "CI Server"

3. Client Setup (macOS)

# Each developer sets their own token
export THE_BRAIN_REMOTE_URL="http://<server-ip>:9420"
export THE_BRAIN_AUTH_TOKEN="mb_a1b2c3d4e5f6..."

# Start push agent
the-brain agent

Permission Model

Roles are assigned per project. A user can be an admin on one project and a contributor on another.

Role	Read	Push interactions	Consolidate	Train	Manage users
admin	✅	✅	✅	✅	✅
contributor	✅	✅	❌	❌	❌
observer	✅	❌	❌	❌	❌

How Context Injection Works

When Oskar opens a session on project CPV, the brain injects context in this order:

Identity anchor — Oskar's stable self-vector (preferences, coding style)
Personal memories — Oskar's own corrections and preferences on CPV
Team memories — Shared patterns detected across the team on CPV
Global overrides — Cross-project patterns (e.g., "always use TypeScript")

This means Anna's preference for tabs doesn't leak into Oskar's sessions, but the team convention "use Redux for state management" is injected for both.

CLI Reference

# User management (admin only)
the-brain user add --name <name> --project <project> [--role admin|contributor|observer]
the-brain user list [--project <project>]
the-brain user remove --name <name>
the-brain user set-role --name <name> --project <project> --role <role>

# Token management
the-brain user token --name <name> [--label "My MacBook"]
the-brain user token --revoke <token-id>

# Audit
the-brain audit [--userId <id>] [--project <project>] [--limit 50]

API Endpoints

All endpoints require admin authentication.

POST   /api/users                          # Create user
GET    /api/users                          # List users
DELETE /api/users/:id                      # Remove user
POST   /api/users/:id/tokens               # Generate new token
GET    /api/users/:id/tokens               # List user's tokens
DELETE /api/users/:id/tokens/:tid           # Revoke token
GET    /api/audit-log?userId=&project=&limit=  # Query audit trail

Future

SSO/OAuth — Google, GitHub OAuth as authentication providers
Rate limiting — Per-user request quotas
Team dashboard — Web UI showing team activity, active users, consolidation status
Shared training — Team-wide LoRA adapters composed with per-user adapters
LDAP integration — Enterprise directory support