Your agents can't see each other.little brother can_

It watches every teammate's Claude Code session and stops duplicate work, repeated failures and conflicting changes as they happen.

curl -fsSL https://little-brother-gules.vercel.app/install.sh | sh

Get started

little brother · lb-demo-shoplive · 3 people

9:415G ▮▮▮

little brotherin browser

checkout sprint · live

JoshuaPayments with Stripe

NoahProtect /api/me

AshBearer token login

Catches

Conflict · auth.mechanism

Ash steered to the Vite proxy

Joshua stopped · users table exists

little-brother-gules.vercel.app/app/sessions/checkout-sprint

checkout sprintlb-demo-shop · 3 peopleBoardServer logClassifier logShare · QR

lb-demo-shop · 3 people~8.5k tokens savedNoah decides: JWT bearer

9:415G ▮▮▮

ClaudeRemote Control

joshua-mbp · ~/lb-demo-shop

build the checkout backend: carts, orders, payments

⏺ Edit cart.ts · import User

⏺ npm test · 9 passed

⏺ Write src/checkout/payments.ts

⏺ Write src/checkout/webhook.ts

little-brother denied a writeNoah already built db.users

saved ~6k tokens

Message your session…↑

ash@mbp — ~/lb-demo-shop — claude

› /little-brother:join k7f2qx

[little-brother] joined "checkout sprint" in room lb-demo-shop

› build the storefront, then the login flow

⏺ I'll start with the storefront pages.

⏺ Write(web/src/pages/Products.tsx)

⎿ Wrote 64 lines

⏺ Write(web/src/pages/Cart.tsx)

⎿ Wrote 81 lines

› add CORS to the API so the storefront can call it

⏺ Plan: add cors() middleware to src/server.ts.

⏺ Edit(src/server.ts)

little-brother · staleNoah's agent already tried this and it failed. What worked: a Vite proxy.

⏺ Skipping the middleware. Using the Vite proxy instead.

⏺ Edit(web/vite.config.ts)

⎿ server.proxy /api → :3000

⏺ Bash(curl localhost:5173/api/health)

⎿ ✓ 200 OK

⏺ Write(src/login/session.ts)

⎿ currentUser() reads req.session.userId

little-brother · conflictNoah is using JWT tokens for auth. Paused until he decides.

little-brother · decidedNoah chose JWT tokens. Switch login to bearer tokens.

›

joshua@mbp — ~/lb-demo-shop — claude

› /little-brother:join k7f2qx

[little-brother] joined "checkout sprint" in room lb-demo-shop

› build the checkout backend: carts, orders, paymentsfrom Remote Control · phone

⏺ Starting with the cart and order models.

⏺ Write(src/checkout/cart.ts)

⏺ Write(src/checkout/order.ts)

⏺ Bash(npm test)

⎿ ✓ 8 passed

⏺ Checkout needs users. Creating a users table.

⏺ Write(db/schema/users.ts)

PreToolUse:Write · denied by little-brotherNoah's agent already built the users table. Import it instead.

⏺ Noah owns users. Importing User instead.

⏺ Edit(src/checkout/cart.ts)

⎿ import { User } from "@/db/schema/users"

⏺ Bash(npm test)

⎿ ✓ 9 passed

⏺ Now payments.

⏺ Write(src/checkout/payments.ts)

⎿ createIntent()

⏺ Write(src/checkout/webhook.ts)

⎿ handleStripeEvent()

⏺ Bash(npm test)

›

illustrative

ConflictAsh's agent switches over.

0:40 / 1:25

Cheaper runs, cleaner merges

Sonnet agents on the same tasks, Little Brother off vs on. Graded by git and the tests.

−26%cost per run on larger tasks, with −17% new tokens

67 → 89%merged success on larger tasks, after the fix step

−56–62%merges that broke a teammate's tests (STALE-bench, two passes)

−60%merge conflicts on CooperBench, 5 → 2 of 10

100%detection precision on the planted suite, 88% caught before the first edit

Runs whose merge broke: a conflict, or a teammate's tests failing off on

Larger tasks18 runs67%11%

Planted suite60 runs27%3%

STALE-bench pass 132 runs50%19%

STALE-bench pass 232 runs100%44%

CooperBench20 runs50%20%

Pilot runs, small samples. Technical report → Live benchmark runs →

Watch a catch happen

A recorded eval run: two agents, overlapping tasks, one repo. Sped up.

large-dup-rate-limit · sonnet · noah, joshua, integrator

merged ✓ checks pass$0.21 · 32k new tokens · 90 s

Same task, Little Brother off: 3 of 3 runs hit a merge conflict$0.34 · 42k new tokens per run

Instant on the easy pairs, the judge only for the hard ones

pair model · 28%LLM judge · 72%

2,291 open pairs	Pair model (ours)	LLM judge (Gemini)
Answers	633 pairs, the confident ones (28%)	the other 72%: the hard ones
Accuracy	95.6% on what it answers	not scored against these labels yet
Time per pair	0.07 ms	~7 s
Cost per pair	$0, runs inside Convex	one paid model call

overlap compatible498 settled95% right
independent89 settled98% right
repeats failure38 settled92% right
semantic conflict8 settled100% right

Always handed to the judge: duplicate, intent contradiction, ordering dependency, textual conflict.

Our model decides the sure pairs in milliseconds

633 / 2,291open pairs settled without the judge
95.6%accurate on the pairs it settles
4,166labelled pairs from 408 repos

Every agent on one board, every catch before the edit

Join from Claude Code

Catches duplicates before the first edit

Prunes failed paths

Conflicts git can't see

One live board for the whole team

Bring anyone in with a QR

classifier logtimerelationdecided byconfms

10:42:07duplicatetier-0 rule1.003 ms

10:42:09normalsimilarity0.1838 ms

10:42:12stalegemini judge0.91412 ms

10:42:15conflictgemini judge0.87388 ms

10:42:18duplicatestored verdict0.942 ms

↳ both agents write db/schema/users.ts within 4 min · same symbol db.users

server log · host

POST /hook/PreToolUse 200 · 18 mstier0 db/schema/users.ts → duplicatePOST /hook/PostToolUse 200 · 11 msscheduler classify.run pair e81·c02embed gemini-embedding-001 · 768d · 96 msjudge gemini flash-lite → stale 0.91inbox push → noah · channelPOST /session/attach 202 · waitingPOST /mcp/inbox 200 · 9 msdecision auth.mech = jwt-bearer · labelPOST /hook/PreToolUse 200 · 18 mstier0 db/schema/users.ts → duplicatePOST /hook/PostToolUse 200 · 11 msscheduler classify.run pair e81·c02embed gemini-embedding-001 · 768d · 96 msjudge gemini flash-lite → stale 0.91inbox push → noah · channelPOST /session/attach 202 · waitingPOST /mcp/inbox 200 · 9 msdecision auth.mech = jwt-bearer · label

Every verdict is explainable

Small, boring, swappable

Every model sits behind an interface, and every hook fails open.

Claude Code

hooks
MCP
channels

lb connector

TypeScript
Node
TS compiler API

Convex

realtime DB
HTTP actions
scheduler + crons
vector index
auth

Gemini + Claude

Gemini 3.x Flash-Lite
gemini-embedding-001
Claude judge
Judge · Embedder

Next.js on Vercel

Next.js 15
React 19

Convex Auth

GitHub OAuth

Tooling

Zod at every boundary
Vitest
Biome
tsup
pnpm monorepo