Under the Hood

No magic.
Just good engineering.

Every query enriched with codebase context. Every task routed to the right specialist. Every autonomous run recoverable. Here is exactly how it works.

1
Step 1

One command.
Full engineering stack.

npx create-merlin-brain installs everything into your Claude Code configuration. No global binaries. No Docker. No cloud dependency for the free tier.

The MCP server starts instantly in every terminal session. Agents, commands, and hooks load on demand — zero startup overhead.

34 agent system prompts
43 slash command definitions
20 event-driven hooks
MCP server with 12 tools
npx create-merlin-brain
Installing Merlin v3.12.0...
~/.claude/
agents/// 34 specialist agents
product-spec.md
implementation-dev.md
...32 more
commands/// 43 slash commands
merlin:route.md
merlin:plan-phase.md
...41 more
hooks/// 20 automated hooks
security-scanner.js
check-file-size.js
...18 more
merlin/// MCP server + config
server.js
settings.json
Done. MCP server configured. Restart Claude Code to activate.
merlin_get_context
query: "refactor auth module to JWT"
Sights Response
files: src/auth/middleware.ts, src/auth/session.ts, src/models/user.ts
decisions: "Use httpOnly cookies, not localStorage"
conventions: "All middleware in src/auth/, tests co-located"
similar_code: "src/auth/oauth.ts already uses token refresh pattern"
Context injected. Claude sees 4 files + 3 decisions + 2 conventions
2
Step 2

Every query,
fully enriched.

Without Merlin, Claude sees only your message. With Sights, every request is enriched with the relevant code, past decisions, and project conventions — before Claude writes a single line.

The knowledge graph is built from your codebase using OpenAI embeddings and pgvector. Queries return results in milliseconds, not seconds.

You ask
"Refactor the auth module to use JWT"
Hook intercepts
pre-edit-sights-check fires before any edit
Sights API responds
Relevant files, architecture decisions, conventions, similar code
Claude receives
Your request + full codebase context in one prompt
3
Step 3

Right agent.
Fresh context. Every time.

When you describe a task, Merlin scores it against 34 local agents and 1,000+ cloud-indexed agents. The best match gets spawned as a completely fresh process with its own 200K context window.

This is not function-calling or prompt-switching. Each agent is a separate Claude instance with a custom system prompt optimized for its domain. It starts clean, works autonomously, and returns structured results.

200K
context per agent
34
local specialists
1000+
cloud-indexed
<2s
routing latency
merlin_smart_route
task: "Build a responsive dashboard with charts"
Scoring 34 local + 1,000+ cloud agents...
merlin-frontendlocal94% A++
ui-builderlocal87% A+
react-specialistcloud82% A
implementation-devlocal71% B+
Selected: merlin-frontend
Spawning fresh process with 200K context...
Agent working autonomously.
merlin-loop build — auth-system
00:00
Plan loaded9 tasks in 3 waves
00:02
Task 1/9Set up database schema
00:08
Task 2/9Create user model & migrations
00:15
Task 3/9Add auth middleware
00:21
CheckpointState saved to .merlin-session/
00:22
Task 4/9Build login/register endpoints
00:31
Task 5/9Add session management — CRASH
00:31
RecoveryRestored from checkpoint. Retrying...
00:32
Task 5/9Add session management
00:40
Task 6/9Write auth tests
5/9 complete | 40min elapsed1 crash recovered | 8 commits
4
Step 4

Hours of work.
Zero babysitting.

The Merlin Loop takes a plan with N tasks and executes them sequentially — each in a fresh Claude process with full 200K context. No degradation. No hallucinations from context overflow.

If a process crashes, the loop recovers from the last checkpoint and retries. Five consecutive failures trigger the circuit breaker — it stops safely instead of creating garbage.

Fresh 200K context per task
No context window degradation over long runs
Checkpoint persistence
State saved after each task to .merlin-session/
Crash recovery
Process dies? Restored from last checkpoint automatically
Circuit breaker
5 consecutive failures = safe stop, not infinite retry
Atomic commits
Each task commits independently — clean git history

See for yourself.

30 seconds to install. Instant context on your first query.

Free forever. No credit card. No signup.