JARVIS-BRAIN

STOP GREPPING. START ASKING.

Federated knowledge graph for Claude Code. Measured on a 5-repo production codebase: -52% fresh input tokens (what counts against Max plan rate limits), 29% faster wall-clock, up to 15× faster on architectural queries - plus cross-repo analytics grep can't do at all.

[ SEE THE NUMBERS → ] [ GET STARTED ]
LIVE production
5 repos federated
5560 wiki pages
5 MCP tools
▓ THE PROBLEM

Your AI assistant wastes most of its context on Grep

Every time Claude Code answers a question about your codebase, it scans files with Grep and Read - eating thousands of tokens per question. Tokens it then can't use to remember your conversation.

▓ BASELINE - "What's the largest connected cluster in acme-core?" (Q45)
633 s
Claude Code grepping across 5 repos. 10+ minutes waiting. Baseline avg over 50 questions: 44s/q · 2204s total.
▓ BRAIN - same question
41 s
15× faster on architectural queries. Brain avg over 50 questions: 31s/q · 1563s total (29% overall speedup).
▓ BASELINE - Q45 "Largest connected cluster in Acme?"
> What is the largest connected community in acme-core? Provide size.

[thinking] "I need to analyze graph structure..."
→ Grep: imports across 5 repos         [long]
→ Read 14 files across Acme / Inventory / Checkout / Customer / Analytics
→ [CC attempts manual clustering logic]

633 seconds  ·  6 tool calls  ·  10 reasoning turns  ·  answer: partial
▓ BRAIN - same question
> What is the largest connected community in acme-core? Provide size.

→ brain_query(..., scope="acme/Acme")       [200ms]
→ brain_graph(group="acme", repo="Acme")    [150ms]
→ [CC reads pre-computed Louvain community_id field]
→ Answer: "Community 0 - >900 components. Hub: useCart (29 inbound)."

41 seconds  ·  7 tool calls  ·  answer: complete with evidence

The Real Wins - Tokens, Time, Both

Full breakdown across 50 questions × 2 configs = 100 Claude Opus runs on production codebase. All measured via Anthropic Agent SDK.

▓ FRESH INPUT (100% rate-limit weight)
-52%
Baseline: 4,145 tok. Brain: 2,003 tok. Counts against 5h + weekly Max limits. Brain cuts it in HALF.
▓ WALL-CLOCK TIME
-29%
Baseline: 2,204s total. Brain: 1,563s total. Saved 641s (~11 min) across 50 questions. Peak: 15× on Q45 (633s → 41s).
▓ OUTPUT (100% rate-limit weight)
+6%
Brain: 55,398 vs 52,229 baseline. Marginal increase - brain answers are more thorough (good) at slight token cost.
▓ CACHE READ (~10% weight)
tie
~7.5M both. Heavily discounted. Total billed cost tied - but cache doesn't hit rate limits.
Practical impact for Max plan users: brain cuts the tokens that actually count against your 5-hour and weekly limits by ~52%, saves 29% wall-clock time overall, and up to 15× on architectural queries. Same subscription, more queries before rate limit, faster answers - plus cross-repo analytics grep literally can't do.
▓ THE SOLUTION

Pre-compute the map. Query it instead of grepping.

Brain extracts a semantic graph from your repos once (LLM or static-analysis). Then Claude Code queries that graph instead of re-reading files every time.

[ 01 ]

Extract

Your repos are analyzed once (local via /brain-extract skill, or LLM on push webhook). Produces graph.json per repo.

[ 02 ]

Federate

Brain merges graphs from all repos in a group (e.g., core + 5 frontends). Cross-repo edges, communities, DRY detection - all pre-computed.

[ 03 ]

Serve

5 MCP tools over HTTPS: brain_query, brain_graph, brain_path, brain_explain, brain_ffcss.

[ 04 ]

Use

2-minute setup: claude mcp add brain ... and Claude Code picks up the tools automatically. Zero changes to your repos.

5
MCP TOOLS
5560
WIKI PAGES
< 3min
REGEN ON PUSH
1M
CONTEXT (OPUS)
$0
MAX PLAN EXTRA
▓ FOR WHOM

Different wins for different roles

▓ FRONTEND DEVELOPER

Writing features across 5 repos

"Acme-core probably has a hook for X - but what's it called?"

Ask Claude Code. It queries brain. Gets answer. No grep across 5 repos.

  • 10× less token burn per code-discovery question
  • CC sessions stay coherent 15× longer
  • Cross-repo view: see what core exposes without leaving your repo
  • Zero friction - brain tools appear in CC automatically
▓ ARCHITECT / TECH LEAD

Keeping a federated codebase sane

"Do Checkout and Inventory duplicate logic that should be in core?"

Insights dashboard + brain_ffcss(violations). DRY violations surfaced automatically.

  • 1395+ code communities detected via Louvain clustering
  • Top 10 god-nodes identified - refactor candidates surface themselves
  • FFCSS token federation: find cross-repo duplicates to promote
  • System-level decisions in 15 min vs 1 day of analysis
▓ CTO / DECISION MAKER

Running a dev team on AI

"How do I know what my AI costs and if it's working?"

Brain dashboard tracks cost / latency / queries / jobs. Full observability out of the box.

  • 29% wall-clock speedup measured on 50-Q benchmark
  • 15× faster on architectural / cross-repo queries
  • Alerts on cost threshold, errors, poll failures
  • Runs on own VPS: no code sent to third parties
▓ NEW TEAM MEMBER

Onboarding to 5-repo codebase

"Where do I even start? What does Acme-core do?"

Federated wiki (5560 pages) + interactive graph. Self-serve exploration.

  • 15 min to productivity instead of 2 hours of hand-holding
  • Visual graph per repo - click to drill into wiki
  • "Who owns this code": git-blame metadata on every node
  • Cross-linked docs auto-generated from code structure
▓ PROOF

50 real questions, real repo, real Claude Opus

We ran 50 code-analysis questions against ~/dev/example/Acme (production Nuxt 4 + Vue 3 codebase) in two configurations - standard Claude Code vs Claude Code + brain MCP. Tokens measured via Anthropic Agent SDK. Same model. Same questions. Same repo.

[ SEE FULL BENCHMARK REPORT → ]

Interactive tabs: Overview · Methodology · Raw Results · By Category · Token Economics · Accuracy · Implementation Guide · FAQ

▓ TECH STACK

Open, auditable, self-hostable

Brain runs on your VPS. Your code never leaves. Open source (MIT). 3h to deploy from scratch on a 2-core 4GB box.

Runtime Python 3.12 + FastAPI + ARQ
Storage Postgres + Redis + SQLite FTS5
Extraction Qwen local (free) + Gemini Flash fallback
Graph NetworkX + Louvain communities
Visualization vis-network (click-through to wiki)
Transport MCP over HTTPS (JSON-RPC 2.0)
Auth Per-dev tokens (scope + rate limit)
Observability Built-in dashboard (cost/latency/jobs)
Deploy Single docker compose
▓ GETTING STARTED

2-minute setup for any developer

▓ STEP 0: Get a dev token
Ask the admin of your brain instance (for brain.sdet.it:
DM @darco81). You'll get a one-time string like
brain_xxxxxxxxxxxxxxxxxxxxxx. Save it to your shell:

export BRAIN_DEV_TOKEN=brain_xxxxxxxxxxxxxxxxxxxxxx

Self-hosting your own brain? See
/docs/getting-started for POST /admin/tokens flow.
▓ STEP 1: Install MCP (one command)
claude mcp add brain \
    --transport http --scope user \
    https://brain.sdet.it/mcp \
    --header "Authorization: Bearer $BRAIN_DEV_TOKEN"
▓ STEP 2: Verify
claude mcp list
# → brain: https://brain.sdet.it/mcp (HTTP) - ✓ Connected
▓ STEP 3: Use in any Claude Code session
> "Show me how useCart is used across acme-core"

[CC calls brain_query + brain_explain automatically]
→ Answer in <3 seconds with file paths, neighbors, community

Full docs: GETTING_STARTED.md

▓ READY?

Try it on your team's codebase

Open-source. Self-hostable. Costs $0 if your devs are on Claude Max. Cuts pay-per-use costs dramatically.

[ SEE BENCHMARK → ] [ GITHUB ] [ TRY IT LIVE ]

READY