vibecop

Claude's auto mode but for any agent -- under your control, and configurable.

A second AI watches your coding agent so you don't have to.

You've turned on permission checks in Claude Code or Gemini CLI. You've been approving bash, read_file, and write_file for the last two hours without reading them. You are the rubber stamp. You are the weak link.

vibecop fixes this. It's a small Go daemon that sits between your coding agent and its permission prompts. When the agent wants to run something, vibecop asks a separate, independent AI — one that has never seen your conversation and can't be manipulated by whatever's in your context window — whether it looks okay. Routine stuff passes silently. Weird stuff gets escalated to you with context. You only see the interesting ones.

No Electron. No GUI. Runs in the background. Optional TUI if you want to watch the decisions fly by.

---

How it works

coding agent  →  PreToolUse hook  →  vibecop daemon  →  LLM
                                           ↓
                                    approve: silent ✓
                                    escalate: your terminal gets a prompt with context
                                    deny: same, but louder

The hook is a one-liner installed into your harness config. The daemon runs in the background. The LLM can be a local Ollama model (near-zero cost, ~13s latency on M4 Pro) or a cloud endpoint like Haiku (~1.7s). If the daemon crashes or times out, the hook fails open — your agent is never blocked by vibecop's own problems.

---

Modes

Baseline (default) — No setup required. Catches obviously bad patterns: recursive deletes outside the project, writing to ~/.ssh, shell startup file modifications, suspicious outbound network calls. Everything else passes.

Guardian — Run vibecop init once per project. A coding agent analyzes your codebase and generates a project-specific system prompt. Now vibecop knows that swift build is routine here but curl http://... probably isn't. Lower escalation rates, more trustworthy verdicts.

---

Supported harnesses

• Claude Code (Anthropic)
• Gemini CLI (Google)
• Codex CLI (OpenAI)
• Copilot CLI (GitHub)

---

Recommended models

Speed matters — vibecop is in your agent's critical path. Measured on M4 Pro 48GB:

Cloud (recommended):

Model	Latency	Notes
claude-haiku-4-5	~1.7s	Recommended default. Fast enough to beat you to the prompt. Anthropic API.
gemini-2.5-flash	TBD	Natural choice if you're running Gemini CLI. Google AI API.

Local (Ollama) — slower than cloud on typical hardware:

RAM	Model	Latency	Notes
16 GB	mistral-nemo:latest	—	Untested
32 GB	qwen3:14b	~12.9s	CoT disabled automatically. Noticeably slow.
64 GB+	qwen3:32b or gemma3:27b	—	Untested

---

Building from source

Requires Go 1.25+ (see go.mod for the current toolchain version).

go build -o ./vibecop ./cmd/vibecop

No CGO, external libraries, or build tags needed. The binary is statically linked by default.

To install directly into your $GOPATH:

go install ./cmd/vibecop

---

Quickstart

First run — just run any command. There's nothing to install first:

vibecop status       # no config? vibecop runs setup automatically

The setup wizard walks you through provider, model, timeout, and API key. After that it offers to install hooks and test the connection. All in one session, zero docs required.

When you pick the OpenAI-compatible provider, the wizard shows a numbered list of known-good chat-completions endpoints (OpenAI, Google Gemini, Groq, Together AI, OpenRouter, DeepSeek, Mistral, xAI, Cerebras, Fireworks, Anthropic compat shim, Perplexity Sonar, plus local Ollama / LM Studio / vLLM defaults). Pick one to fill in the verbatim URL vibecop will POST to. "Custom URL" is the last option for anything not listed (e.g. self-hosted gateways or Azure OpenAI, which needs per-deployment templating).

If you prefer to do it step by step:

vibecop setup              # interactive wizard
vibecop test               # verify the LLM endpoint works
vibecop start              # boot the daemon (foreground; Ctrl+C to stop)
vibecop tui                # (another terminal) watch verdicts in real-time

Then for full integration with coding agents:

vibecop install --all      # wire hooks into Claude Code, Codex, Gemini CLI, Copilot CLI
vibecop init --harness claude   # generate Guardian prompt for this project

By default the installed hook calls vibecop hook and resolves the binary through $PATH. Pass --vibecop-path to either vibecop setup or vibecop install to point the hook at a specific binary instead — handy when testing a local build without overwriting the system install:

go build -o ./vibecop ./cmd/vibecop
vibecop install --all --vibecop-path ./vibecop

The path is resolved to absolute, so the hook works regardless of the agent's working directory. Re-running with a different --vibecop-path updates the existing entry in place rather than appending a duplicate.

Config

~/.vibecop/config.toml — created by vibecop setup, or write it yourself:

Haiku (recommended):

[daemon]
enabled         = true
timeout_ms      = 5000
activity_window = 10
audit_enabled   = false

[model]
endpoint   = "https://api.anthropic.com/v1/messages"
api_format = "anthropic"
model      = "claude-haiku-4-5"
api_key    = "sk-ant-..."

Gemini Flash (for Gemini CLI users):

[model]
endpoint   = "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions"
api_format = "openai"
model      = "gemini-2.5-flash"
api_key    = "AIza..."

Local Ollama (no API key needed):

[model]
endpoint   = "http://localhost:11434/v1/chat/completions"
api_format = "openai"
model      = "qwen3:14b"
api_key    = ""

vibecop setup auto-detects Ollama models if Ollama is running locally.

audit_enabled controls whether every verdict is written to a permanent daily JSONL file at ~/.vibecop/projects/<project-hash>/audit/YYYY-MM-DD.jsonl. Each record includes tool name, input, verdict, reason, and latency. Default is false — the daemon keeps only an ephemeral in-memory window of recent verdicts (see activity_window) and discards the rest.

---

TUI

vibecop tui

Attaches to a running daemon and shows:

• Live verdict feed (Approved / Escalated / Denied) with tool name and reason
• Rolling average latency (last 50 requests)
• Daemon status and active mode (Baseline / Guardian)

---

Commands

vibecop setup              Interactive first-time setup wizard
vibecop start              Start the background daemon
vibecop stop               Stop it
vibecop status             Show daemon status and config
vibecop tui                Attach the live TUI
vibecop init               Generate Guardian prompt for current project
vibecop install            Install hooks into harness config
vibecop uninstall          Remove hooks
vibecop test               Test connection to the configured LLM endpoint
vibecop refine             Regenerate the Guardian prompt (uses recent activity as context)

---

Background

This project grew out of experimentation with the AI second-opinion concept — the idea that a second, context-free model evaluating tool-use requests is more trustworthy than a human who's been clicking Approve for two hours. vibecop is the standalone, no-GUI version: a single static binary that runs anywhere Go runs, with no Electron and no dependencies on any particular coding harness's UI.

See docs/spec.md for the full design specification.