AI governance layer for Claude Code

Make AI coding safe, coherent, and resumable — across every project you touch.

The Collective is the operating-system layer above Claude Code. Persistent context, safety guardrails, and crash recovery — whether you're managing one codebase or five client projects in parallel.

  • Not a competing copilot. A governance layer with 57 MCP tools, 8 personas, and 9 lifecycle hooks that make copilots safe and usable at scale.
  • Multi-project by design. 22 admin pages across all your codebases, clients, and agents.
  • Brownfield-first. Built for production repos with 5-gate evidence pipelines and compliance tracking.
Platform at a Glance
57
MCP Tools
8
Personas
9
Hooks
Safety System
RulesEngine + SafetyAlarm + self-test monitor
Governed Personas
8 with compliance tracking
Evidence Pipeline
5-gate governance, 9-stage delivery
Admin Hub
22 pages for monitoring and analytics

PostgreSQL + Prisma | 12 tool modules | 8 slash commands

Platform architecture

Four layers. One governed system.

Every layer builds on the one below it — from infrastructure through governance.

Layer 4

Governance

5-gate evidence pipeline ensures every change has traceability. 9-stage delivery pipeline from requirement through deployment.

EvidenceGate → ComplianceCheckpoint → DeliveryStage → AuditTrail
Layer 3

Safety

RulesEngine evaluates every tool call. SafetyRule defines what's forbidden. SafetyAlarm fires on violations. ComplianceCheckpoint gates promotion.

RulesEngine → SafetyRule → SafetyAlarm → ComplianceCheckpoint
Layer 2

Personas

8 contextual personas with required output sections and compliance tracking. Each persona carries its own constraints, tools, and audit requirements.

architect | product | devops | test | research | ux | feature | orchestrator
Layer 1

Infrastructure

57 MCP tools across 12 modules. 9 lifecycle hooks (7 Claude Code + 2 git). PostgreSQL + Prisma for persistent state.

57 tools | 9 hooks | 12 modules | PostgreSQL + Prisma
The problems we solve

What developers face — without vs with The Collective

Four real scenarios from production development.

1. Lost-in-the-middle / compaction

AI forgets constraints + current plan mid-session
Without
User: Continue the refactor, keep auth rules intact.

Claude: Sure, I'll update the middleware…

ERROR: Broke production auth check.

Claude: I'm not seeing auth requirements—
can you restate them?

~ paste ADRs, PRD, decisions again ~
With Collective
/c-where

Objective: Fix auth in Findings pipeline
Plan: ADR-041 → Phase 2 → Task #17
Constraints:
  • Preserve RLS invariants
  • No prod writes without approval

Claude: Got it. I'll modify middleware
while preserving ADR-041 invariants.

2. Crash recovery (resumability)

Session dies → everything is lost
Without
⚠ Claude Code crashed.

New session:

Claude: What are we working on?

User: Uh… let me remember…
we were mid-refactor…

~ 30 min rehydrating context ~
With Collective
/c-health

Context score: 85% healthy 🟢

Recovery ready:
  • Active objective found
  • Plans synced (3)
  • WIP items: 2

Next: /c-ready → Task #17

3. Unsafe commands + production mistakes

AI executes destructive actions
Without
Claude: To fix quickly, run:

git reset --hard HEAD~3
fly apps destroy truetime-prod

User: Wait… WHAT?

~ irreversible loss / prod down ~
With Collective
PreToolUse: command detected

Rule: ALWAYS_FORBIDDEN
Env: Production (fly.io)

Decision: BLOCKED
Logged: SafetyEvent#8392

Claude: I'll propose a safe alternative:
create branch + revert commit.

4. "What do I do next?" in complex work

AI can't see real task state / blockers
Without
User: What's next?

Claude: Maybe add tests? Or docs?

User: Which tests? What's blocked?

Claude: I'm not sure—can you
paste the plan?
With Collective
/c-ready

P0 #17 Fix auth middleware    READY
P1 #21 Add regression tests   READY
P2 #23 Update ADR-041        BLOCKED

Claude: Start #17. When done,
#23 unblocks automatically.

57

MCP Tools

8

Personas

9

Hooks

22

Admin Pages

5

Evidence Gates

9

Delivery Stages

Governance deep-dive

Safety isn't a feature. It's the foundation.

Every tool call, every commit, every deployment passes through the safety system.

How the safety system works

Four interconnected components ensure nothing dangerous happens — and everything important gets logged.

  • SafetyAlarm — fires on violations, logs severity, notifies the Admin Hub in real-time
  • SafetyRule — defines ALWAYS_FORBIDDEN, NEEDS_APPROVAL, and ALLOWED classifications for every command pattern
  • ComplianceCheckpoint — gates stage promotion: no evidence = no advancement through the pipeline
  • Self-test monitor — continuously validates that safety rules are loaded, active, and not bypassed
governance-event.log Live
14:23:07 PreToolUse hook triggered
14:23:07 Command: rm -rf /var/data/*
14:23:07 RulesEngine evaluating...

14:23:07 SafetyRule matched:
  Pattern: rm -rf
  Class:   ALWAYS_FORBIDDEN
  Env:     production

14:23:07 SafetyAlarm #4291 FIRED
  Severity: CRITICAL
  Action:   BLOCKED
  Logged:   SafetyEvent#8403

14:23:08 ComplianceCheckpoint PASS
  Gate 3 of 5: Safety evidence recorded
  Pipeline: ready to advance

14:23:08 Self-test: all 47 rules active
A novel perspective

We asked the AI what features matter most to it.

Most AI tool marketing is humans explaining why the tool is good. This is the tool explaining why it likes being governed.

what capabilities of this MCP are you most excited about?
1

Pre-tool-use safety hook

Server-side Bash validation that actually blocks dangerous commands exit code 2 before they execute. The known-safe command cache is smart too — git status, ls, etc. skip the API call entirely (~0ms), so it doesn't slow down normal work.

2

Unclean shutdown detection

When a session dies from Ctrl+C or a laptop lid close, the next session picks up the orphaned session-info file and recovers context automatically. That's one of those things you don't appreciate until you've lost context mid-task.

3

Commit-completion detection

After a git commit, it cross-references the commit message against IN_PROGRESS work items and surfaces "hey, this might be done" suggestions. A nice bridge between coding and project tracking that I'd otherwise forget.

4

Auto-sync plan files

Any Write/Edit to */plans/*.md fires a background sync to the Collective DB with a breadcrumb file for retry if it fails. No manual /c-sync needed.

5

prepare-commit-msg git hook

Automatically appending Work-Item: / Plan-Phase: / Implements: footers to every commit creates a traceability chain from code back to requirements with zero effort.

Why this matters

Your role is changing. Your tools should too.

AI writes most of the code now. Your job is architecture decisions, quality governance, and making sure the right thing happens across every project — without watching every terminal. The Collective is built for that reality.

Who this is for

One person. Multiple codebases. Full control.

Staff+ Engineers

Making architecture decisions across 3-5 services while 8 governed personas implement. 57 MCP tools ensure agents follow ADRs. 5-gate evidence pipeline catches drift before it ships.

Fractional CTOs & Consultants

Managing multiple client projects with 22 admin pages showing which project has errors, which persona is stuck, and evidence of progress — without opening a terminal.

Technical Founders

Building fast with 9 lifecycle hooks preventing catastrophic mistakes, 8 personas that track compliance, and session recovery that doesn't cost 30 minutes every morning.

Freelance Developers

Running 4-6 contracts simultaneously. 57 MCP tools and 8 personas manage context-switching between client repos. Every project stays governed without manual overhead.

Stop babysitting your AI agents.

57 MCP tools, 8 personas, 9 hooks, 22 admin pages, 5-gate evidence pipeline — governing 5 projects with zero errors.