LIL BRO — ~/your-project
ollama phase 25
LIL BRO ~/your-project
▮ LIL BRO
BIG BRO ▶
[BIG BRO ▶]
Tab switch Ctrl+B port→bro /help commands /explain ask bro /player rpg
What It Is

Local AI.
No keys.
No cloud.

LIL BRO is a dual-agent terminal interface that runs two AI coding assistants locally on your machine using Ollama. No API keys. No cloud services. No subscriptions. Everything runs on your hardware.

The Two Agents
Big Bro
Primary coder.
Full access.
The main coding agent. Has full read and write file access, can edit code, search repos, and execute commands. When something needs to be built or fixed, Big Bro does it.
read / write / execute
Lil Bro
Helper.
Read-only.
Focuses on explaining, reviewing, and clarifying. Has read-only permissions by default. Expands to full write access via Bunkbed Mode when you need both agents coding together.
read-only · bunkbed mode

Grandpa
Shared Knowledge Base
Coding Bible + Reasoning Bible
Both agents share a local knowledge base called Grandpa — two reference libraries built from your project. Hybrid retrieval combines keyword matching with model-based semantic search, so the right context surfaces even when individual methods miss the mark.
Big Bro ─┐ ├──→ Grandpa (keyword + semantic) Lil Bro ─┘ shared context log: each agent tracks what the other is doing

Supported Models

Runs via Ollama on your local hardware. The recommended default is qwen2.5-coder:7b — medium speed, full tool-calling support, 5–6GB VRAM.

Model Notes
qwen2.5-coder:7b ★ Recommended default — 5–6GB VRAM, full tool support
qwen2.5-coder:3b Lighter hardware, less VRAM required
qwen2.5-coder:14b Higher quality, needs more VRAM
deepseek-coder-v2 Alternative coding model via Ollama
llama3.1 / llama3.2 General purpose, good reasoning

Any model available via ollama pull can be used.


Features
Smart Failure Handling
Four consecutive tool errors triggers a sibling warning. Five errors and the agent stops completely and asks for help — no silently broken code generation.
Shared Context Log
A persistent log lets each agent passively track what the other is doing. No direct messaging — just shared awareness of each other's progress.
Bunkbed Mode
Toggle to give Lil Bro full write access alongside Big Bro. Both agents coding the same project at the same time.
Personality System
Working status indicators, competitive banter between agents, and honest error reporting. They'll tell you when they're stuck instead of faking it.
Hybrid Retrieval
Grandpa combines keyword search with semantic model-based search. If one method misses, the other catches it — you get relevant context either way.
Full Tool Access
Big Bro: read, edit, search, execute, calculate. Lil Bro: same set in read-only mode, plus Grandpa lookups. bunkbed unlocks everything.

Requirements
System
Python
3.11 or later

RAM
8 GB minimum · 16 GB recommended

GPU VRAM
6 GB+ for the 7B model
Status
✓ Windows — full support
~ macOS — in progress
~ Linux — in progress

Beta. Installers and shell execution refinements pending for non-Windows.

Coming Next
Punishment Mode
Split across machines.
Big Bro stays on your dev machine. Lil Bro runs as a lightweight remote client — accessible from a phone or another device over a network connection. Same dual-agent setup, distributed.