BrokeCrew 🤖⚡

A lightweight, resilient, async-first Rust CLI that coordinates a specialized team of AI personas backed by OpenRouter models.

BrokeCrew makes free-tier models first-class citizens while eliminating their instability through live model catalog discovery, ranked multi-candidate fallback cascades, account-wide rate-limit budgeting, and multi-agent handoff pipelines.

---

Why BrokeCrew (vs LiteLLM or RouteLLM)?

Feature	BrokeCrew	LiteLLM Proxy / RouteLLM
Footprint & Startup	Single standalone native binary (~15MB), sub-10ms startup.	Heavy Python runtime, tens of pip dependencies, slower startup.
Hardware Fit	Designed to run smoothly on low-RAM ARM tablets, Raspberry Pis, and local dev boxes.	Heavy memory consumption unsuited for constrained devices.
Persona Fallback Walk	Declarative ranked candidates per persona (qwen:free → deepseek:free → paid fallback).	Static routing with complex configuration layers.
Live Catalog Awareness	Syncs /v1/models in background; skips dropped/offline models before wasting requests.	Often returns raw 404/400 errors when models vanish.
Account-Wide Budgeting	Tracks rolling 20 req/min and daily quotas (50 or 1,000 req/day) across all calls including router.	Typically only tracks proxy tokens, not OpenRouter account limits.
Multi-Agent Handoffs	Built-in typed pipelines (coder → reviewer) with real-time SSE streaming per persona.	Requires external orchestration frameworks (LangGraph, CrewAI).

---

Terminal Experience

BrokeCrew formats multi-agent discussions like a vibrant team group chat:

$ brokecrew chat --pipeline build-and-review "Write a Rust LRU cache"

[coder] (qwen/qwen-2.5-coder-32b-instruct:free)
Here is a thread-safe, high-performance LRU cache implementation using std::collections::HashMap
and a doubly-linked list with generational indexing...

[reviewer] (meta-llama/llama-3.3-70b-instruct:free)
Code Review:
1. Concurrency: Eviction edge case when capacity is 0 is unhandled.
2. Memory Safety: Generational indices avoid raw pointers; good choice.
3. Suggestion: Implement std::borrow::Borrow to allow lookup by &str rather than String.

---

Quickstart

1. Installation

cargo install brokecrew

Or build from source:

git clone https://github.com/brokecrew/brokecrew.git
cd brokecrew
cargo install --path .

2. Configure OpenRouter API Key

BrokeCrew securely stores your API key in your OS credential store (keyring), never in plaintext:

brokecrew config set-key

(You can also export OPENROUTER_API_KEY="sk-or-v1-..." or pass --api-key for headless servers).

3. Start Chatting

Zero-config instant routing:

brokecrew chat "build me a proxy checker site"

Stage A regex/keyword routing automatically detects "build/site" and dispatches to the coder persona in 0ms!

Run an interactive team session:

brokecrew chat

Run a multi-agent handoff pipeline:

brokecrew chat --pipeline build-and-review "Implement a token bucket rate limiter in Python"

Check model catalog and quota health:

brokecrew status

List available personas:

brokecrew personas list

---

Declarative Personas (personas.toml)

Personas are defined in ~/.config/brokecrew/personas.toml (or ./config/personas.example.toml):

[[persona]]
name = "coder"
system_prompt = "You are a precise, terse coding assistant. Output working code, minimal commentary."
trigger_words = ["build", "code", "script", "function", "website", "api"]

[[persona.candidates]]
id = "qwen/qwen-2.5-coder-32b-instruct:free"
tier = "free"
is_stealth = false

[[persona.candidates]]
id = "deepseek/deepseek-coder:free"
tier = "free"
is_stealth = false

[[persona.candidates]]
id = "google/gemini-2.0-flash-001"
tier = "paid"
is_stealth = false

---

License

Dual-licensed under either:

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.