collet

Orchestrates any LLM, any CLI agent, across TUI, web, and IDE.

Eliminates the stuck problem common in Node.js-based tools at the source, delivering high-level code intelligence and precise editing on the safety foundation of Rust.

🇰🇷 한국어 | 🇨🇳 简体中文 | 🇯🇵 日本語 | 🇩🇪 Deutsch | 🇫🇷 Français | 🇪🇸 Español | 🇧🇷 Português | 🇷🇺 Русский | 🇮🇳 हिन्दी | 🇦🇪 العربية

Features

Quick Start

Requirements

• Rust 1.78+ (Edition 2024)
• LLM provider API key

Installation & Running

via Homebrew (macOS)

brew install epicsagas/tap/collet

via cargo-binstall (fastest - pre-built binaries)

cargo binstall collet

Requires cargo-binstall installed first: cargo install cargo-binstall

via crates.io

cargo install collet

From source

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

Setup and Run

collet setup # or use --advanced option for more detailed settings

# TUI
collet

# Headless
collet "hello collet!"

Documentation

Research

Building collet required going deep on problems that don't have obvious answers yet — from keeping long-running agents coherent across hundreds of turns, to making multi-provider rate limiting actually work, to figuring out when multi-agent parallelism helps and when it just adds noise.

The research and engineering analysis behind these decisions is published as a series of standalone reports:

Build Info

Roadmap

• Phase 1: MVP — API connector, TUI, tool framework, agent loop
• Phase 2: Context persistence, journaling, repo map, smart compaction
• Phase 3: git-patch hunk editing, reasoning preservation, session restore (--continue/--resume)
• Phase 4: Multi-language tree-sitter (Python, TS, Go), PageRank ranking, LSP/MCP
• Phase 5: Agent Skills, Subagent, @mention (file/folder/model), BM25 search
• Phase 6: MCP, LSP, Side-by-side Diff, prompt caching
• Phase 7: 6 color themes, thinking animation, Architect plan review
• Phase 8: Parallel/team mode (multi-agent orchestration, session-level consensus, conflict detection)
• Phase 9: Soul.md persistent personality — per-agent emotion/thought/growth recording, global/per-agent toggle, auto-compaction
• Phase 10: Flock mode completion — Swarm renaming, RoleBased strategy, PlanReviewExecute pipeline, revision loop with consensus voting
• Phase 11–28: RAG document search (Alcove + HTTP Bridge), IDE integration (ACP), Remote gateways (Telegram/Slack/Discord), Web server, Auto-routing, PII filter, Optimizer, Multi-provider wizard
• Phase 29: Agent loop & swarm hardening — zero-drop guarantee, adaptive retry, swarm consensus refinement, coordinator/worker scheduling improvements
• Phase 30: Model optimizer advanced — provider-aware cost/latency scoring, smart tier reassignment on provider add/edit, benchmark-driven model ranking
• Phase 31: Web & remote platform upgrade — Web UI dashboard v2, Telegram/Slack/Discord UX polish, webhook relay, real-time swarm status streaming

Origin & Status

This project began as a personal playground to squeeze more out of GLM models. Then one feature led to another — new models, papers, and paradigms kept arriving before the previous ones were cold.

"New models, new papers, new paradigms — faster than any commit log can keep up." — A recurring nightmare.

Current state — Now dogfooding itself via collet, eating its own tail.

Stability: The core agent loop is solid. Parallel/swarm execution and external integrations are still settling.

License

Apache-2.0