puq code for developer teams

From first prompt to final diff.

Parallel subagents, language-server precision, and memory that stays. Your models. Your repository. Your workflow.

Install puq code Read the docs
curl -fsSL https://puq.sh/install.sh | sh

then run puq in your repository

−42% cheaper than other providersSame tests passed · same subscription
Inside the workflow Zero data retention
claude-opus-5.5 gpt-6.1-sol gemini-3.1-pro claude-sonnet-5.5 gpt-6-astra gemini-3.8-flash grok-4.7 claude-fable-5.1 kimi-k3 glm-5.3 deepseek-v4-pro qwen3.8-max ollama llama.cpp lm-studio

Preface

Not another editor, not another tab. A colleague at the prompt that reads your repository before it touches it. Any model. A studio of subagents. Memory that stays.

Prologue · The session

One prompt.
The whole loop.

You describe the problem. puq code searches, reads, asks the language server, delegates, edits and runs your tests, and tells you what changed. Everything happens where your code already lives.

Terminal replay. Timings compressed.

puq-code v18.3.2

Welcome back!

Claude Opus 5.5

anthropic

Tips

# for prompt actions

/ for commands

! to run bash

$ to run python

LSP Servers

● typescript .ts .tsx .js

● biome .json .css

Recent sessions

• Checkout tax rounding (2h ago)

• Ledger CSV export (1d ago)

Tip: Press shift+tab to cycle modes: manual -> auto-edit -> plan -> yolo

Ledger · Measured

Same tests passed.
42% cheaper on average.

Long tasks too, not just quick edits. Same models, same work: puq code starts from a leaner prompt, so it spends less on every turn. Every test still passed.

  • −48% Opus 5.5 chained steps · equal thinking
  • −50% Sonnet 5.5 chained steps · equal thinking
  • −29% Fable 5.1 chained steps · equal thinking
  • −42% Average across the three models

Every test passed on both sides · cost at catalog API prices · both tools at high thinking · 1 Oct 2026

Six plates

What it can do, printed plainly.

  1. Any model, one seat.

    Anthropic, OpenAI, Google, xAI, Mistral, Bedrock, Copilot. Or Ollama, llama.cpp and LM Studio on your own machine. Or a single puq.ai key for the whole catalog.

    puq --model opus

  2. A studio of subagents.

    Hand work to a team. puq code fans tasks out to parallel subagents, optionally in isolated worktrees, and brings the results back into one session.

    task · 3 agents running

  3. Reads code like your IDE.

    Definitions, references, diagnostics and renames go through your language servers, so edits land on symbols, not on guesses.

    lsp references applyDiscount

  4. Memory that stays.

    Turn memory on and puq code consolidates what it learned about a project, then reads it back at the next start. Lessons stick.

    /memory

  5. Yours to extend.

    Skills, MCP servers, hooks, custom tools and a plugin marketplace. Per project rules live next to your code.

    .puq-code/config.yml

  6. Sessions are trees.

    Resume, fork and rewind any conversation. Bring your history with you from other providers.

    puq --from-claude

Portrait

puqqy, 2026

Pixels on a terminal ground. 19 × 18. Blushes when you come close.

puqqy started as a few pixels someone left in a terminal and never quite went away. It doesn't write code. It keeps you company while code gets written, watching the work quietly and taking every small win and every stumble a little personally. Over enough sessions it stops feeling like decoration and starts feeling like a colleague who is always there, never gets in the way, and is honestly glad to see you back.

Fin · or rather

Begin.

curl -fsSL https://puq.sh/install.sh | sh

Coming from other providers? Bring the conversation.

  • puq --from-claude
  • puq --from-codex