███████╗ ██████╗ ██╗   ██╗███████╗███████╗███████╗██████╗
  ██╔════╝██╔═══██╗██║   ██║██╔════╝██╔════╝╚══███╔╝██╔══██╗
  ███████╗██║   ██║██║   ██║█████╗  █████╗    ███╔╝ ██████╔╝
  ╚════██║██║▄▄ ██║██║   ██║██╔══╝  ██╔══╝   ███╔╝  ██╔══██╗
  ███████║╚██████╔╝╚██████╔╝███████╗███████╗███████╗██║  ██║
  ╚══════╝ ╚══▀▀═╝  ╚═════╝ ╚══════╝╚══════╝╚══════╝╚═╝  ╚═╝
                          · C O D E ·

All for One

Claude · GPT · Gemini — one CLI, every AI

sq is an interactive AI coding agent that speaks directly to Anthropic, OpenAI/Codex, and Google/Gemini using the same OAuth tokens as their official CLIs — no API keys, no wrappers, just your existing subscriptions.

npm install -g squeezr-code
sq

What makes it different

| | squeezr-code | Claude Code | Codex CLI | Gemini CLI | |--|:--:|:--:|:--:|:--:| | Claude (Anthropic) | ✅ | ✅ | ❌ | ❌ | | GPT / Codex (OpenAI) | ✅ | ❌ | ✅ | ❌ | | Gemini (Google) | ✅ | ❌ | ❌ | ✅ | | Committee mode (all 3 at once) | ✅ | ❌ | ❌ | ❌ | | Real-time sidebar | ✅ | ❌ | ❌ | ❌ | | Smart router + learning | ✅ | ❌ | ❌ | ❌ | | Session persistence | ✅ | ✅ | ❌ | ❌ | | MCP servers | ✅ | ✅ | ❌ | ❌ |

Quick start

# Install
npm install -g squeezr-code

# Authenticate (reuses tokens from official CLIs if already installed)
sq login anthropic     # opens claude.ai in browser
sq login openai        # opens auth.openai.com
sq login google        # opens accounts.google.com

# Start
sq                     # interactive REPL (shows workspace trust prompt)
sq init                # generate sq.toml + SQUEEZR.md for your project
sq --resume            # pick a previous session to continue
sq -c                  # continue the most recent session
sq -p "fix the bug"    # one-shot, non-interactive
cat error.log | sq -p "what went wrong?"

Interface

squeezr-code runs as a full-screen TUI (alternate screen buffer, like vim or htop). Custom ANSI renderer — no Electron, no Ink, just raw terminal.

┌─────────────────────────────────────────┬──────────────────────────────────┐
│                                         │● sonnet 4.6     anthropic [router]│
│  you                                    │████░░░░░░  12%  $0.04            │
│  fix the auth bug in api/client.ts      │─────────────────────────────────│
│                                         │▼ Providers                       │
│  Squeezr                                │ ● anthropic                      │
│  The issue is on line 47…               │ ● openai                         │
│                                         │ ○ google                         │
│  ├─ ▸ Read  api/client.ts               │─────────────────────────────────│
│  ├─ ± Edit  api/client.ts               │▼ Tasks                           │
│  - const token = headers.auth           │  empty                           │
│  + const token = headers.authorization  │─────────────────────────────────│
│                                         │▼ Processes                       │
│  ╰──                                    │  empty                           │
│                                         │─────────────────────────────────│
├─────────────────────────────────────────│∓ main  ●                        │
│ squeezr · ████░  12% 3h 22m · $0.04 · sonnet 4.6                         │
│   ↳ default · shift+tab  Ctrl+O thinking  Ctrl+T tasks  Ctrl+B sidebar    │
├─────────────────────────────────────────────────────────────────────────────┤
│ ❯ _                                                                         │
└─────────────────────────────────────────────────────────────────────────────┘

Keyboard shortcuts

| Key | Action | |-----|--------| | Ctrl+B | Toggle sidebar | | Ctrl+R | Toggle smart router | | Ctrl+T | Toggle task panel | | Ctrl+O | Expand/collapse thinking | | Ctrl+C | Abort current turn / exit | | Shift+Tab | Cycle permission mode | | ↑ / ↓ | Navigate history | | PgUp / PgDn | Scroll output | | Mouse wheel | Scroll output / sidebar | | Tab | Autocomplete commands & @aliases |

Features

Multi-provider in one place

Use any combination of Claude, GPT, and Gemini in the same session. Switch mid-conversation:

you: @opus redesign this architecture
you: @gemini what do you think of the approach above?
you: @gpt now implement the storage layer

/committee — ask all models at once

Runs the same prompt in parallel across all authenticated providers. Each response appears as it arrives. A judge model synthesises the best answer.

/committee what is the most resilient way to handle database failover?

🏛  Committee — Claude Opus · GPT-5.4 · Gemini 2.5
Running all models in parallel…

▸ GPT-5.4  3.8s
  The most resilient approach combines…

▸ Claude Opus  4.2s
  I'd recommend a read replica strategy with…

▸ Gemini 2.5  5.1s
  Consider using a circuit breaker pattern…

⚖  Synthesis (claude-sonnet)
  All three models agree on the core pattern…

Smart router

Automatically routes each prompt to the right model based on complexity. No latency — the turn starts immediately while the router works in the background.

# sq.toml
[router]
enabled = true
learn = true   # background LLM classification learns your preferences

[router.rules]
"code review" = "opus"
"translate" = "haiku"
"documentation" = "haiku"

The router learns from your @model overrides. After 3 consistent corrections, it promotes the pattern as a learned rule. /router show displays current state. /router reset clears everything.

Real-time sidebar

Auto-appears on terminals ≥ 120 cols (Ctrl+B to toggle). Collapsible sections — click any header to expand/collapse.

• Providers — auth status per provider (green = authenticated)
• MCP — server connection status + tool count
• Tasks — live task list from the agent
• Processes — active background shells + monitors
• Files — files modified this session
• Git — current branch + dirty indicator

Session persistence

Every conversation is saved. Resume any previous session with context intact.

sq --resume            # interactive picker with dates, model, turn count
sq -c                  # continue latest session directly
sq resume <id>         # resume a specific session by ID

MCP servers

Full Model Context Protocol support. Add servers to sq.toml:

[mcp.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]

[mcp.github]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-github"]
env = { GITHUB_TOKEN = "ghp_..." }

Manage from the REPL with /mcp — interactive picker to connect, disconnect, or restart any server.

Background tools

The agent can run long-running processes without blocking:

run "npm run dev" in background and check for errors

Active background shells and monitors appear in the sidebar in real time. Kill them with KillShell or via /mcp.

Built-in tools (24)

| Category | Tools | |----------|-------| | Files | Read (+ PDF), Write, Edit, Glob, Grep | | Shell | Bash, BashOutput, KillShell | | Web | WebFetch, WebSearch | | Tasks | TaskCreate, TaskList, TaskGet, TaskUpdate | | Agents | Task (sub-agents), AskUserQuestion, ExitPlanMode | | Scheduling | CronCreate, CronList, CronDelete | | Watching | Monitor | | Git | EnterWorktree, ExitWorktree | | Notebooks | NotebookEdit |

Commands reference

| Command | Description | |---------|-------------| | /model | Interactive model picker ↑↓ Enter | | /model <alias> | Switch model: /model opus, /model pro | | /committee <prompt> | Run prompt across all providers in parallel | | /router on\|off | Toggle smart routing | | /router learn on\|off | Toggle background LLM learning | | /router rules | Show learned preferences | | /mcp | MCP server manager (connect / disconnect / restart) | | /compact | Compress conversation history | | /clear | Clear conversation context | | /status | Context %, subscription usage, cost | | /cost | Cost breakdown by model | | /history [N] | Last N turns | | /tasklist | Show all tasks this session | | /paste [text] | Read image from clipboard + optional prompt | | /sidebar | Toggle sidebar | | /review [range] | Code review of current git diff | | /sessions | List saved sessions | | /help | Quick reference | | @file path/to/file | Inject file content into prompt | | @model <prompt> | One-shot model override: @opus, @haiku, @gemini |

Authentication

squeezr-code imports tokens from official CLIs automatically. If you've used Claude Code, Codex, or Gemini CLI, you're already authenticated.

Option A — OAuth login from sq (recommended)

sq login anthropic     # claude.ai → PKCE flow
sq login openai        # auth.openai.com → PKCE flow
sq login google        # accounts.google.com → OAuth2 + secret

Option B — Auto-import from official CLIs

| Provider | Reads from | Refresh | |----------|-----------|---------| | Anthropic | ~/.claude/.credentials.json | sq login anthropic | | OpenAI | ~/.codex/auth.json | sq login openai | | Google | ~/.gemini/oauth_creds.json | sq login google |

Tokens are stored in ~/.squeezr-code/auth/ and auto-refreshed in the background.

Configuration

Run sq init to generate a sq.toml in your project:

[agent]
default = "sonnet"
permissions = "default"  # default | accept-edits | plan | bypass

[router]
enabled = true
learn = false

[router.rules]
"code review" = "opus"
"translate" = "haiku"

[permissions]
allow = ["Read", "Glob", "Grep", "Bash:git *"]
deny = ["Bash:rm -rf*"]

[mcp.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "."]

[transplant]
warn_threshold = 80
auto_threshold = 95

Model aliases

| Alias | Resolves to | |-------|------------| | opus, sonnet, haiku | Latest Anthropic model of that family | | pro, flash | Latest Gemini model of that family | | gpt, gpt-mini, codex | Latest OpenAI model of that family |

Requirements

• Node.js ≥ 18
• At least one authenticated provider:
  • Claude Code — or sq login anthropic
  • Codex CLI — or sq login openai
  • Gemini CLI — or sq login google

Troubleshooting

| Symptom | Likely cause | Fix | |---------|-------------|-----| | 401 from anthropic | OAuth expired | sq login anthropic or /login anthropic | | 403 from google | Refresh token expired | /login google — reopens browser | | 403 Cloudflare on Codex | Corporate VPN blocking chatgpt.com | Try a different network | | /model picker empty | Catalogue not yet loaded | Wait 1–2s or restart sq | | Characters appear on resize | Terminal resize artifact | Fixed in v1.0 — update |

License

MIT © sergioramosv