Desktop Cockpit for Multi-Agent Coding

Run a team of AI coding agents.
Side by side.

A Director decomposes your project into subtasks, Workers code simultaneously across your favorite models, and an independent Critic verifies every change before it lands.

Download for Windows Linux (.AppImage / .deb)

Free · Open alpha · Runs locally on your hardware

overbench / workspace: overbench-core

Claude: 38%

Cursor: 340/500

4 agents active

Director claude-3.7-sonnet

Orchestrating run_eQUM4

Goal: Implement OpenRouter provider & verify streaming SSE parser

✓ Decompose tasks into briefs Director

Worker 1: Add OpenRouter to CliKind Cursor

Worker 2: Implement SSE token parser opencode

○ Critic gate: Cross-model audit Gemini 2.5

Worker 1 cursor-agent · src/cli_kind.rs

Editing

❯ █

Worker 2 opencode · vitest runner

Testing

$ vitest run tests/openrouter.test.ts

✓ tests/openrouter.test.ts (8 tests) 412ms

✓ detects provider 'openrouter' from env

✓ parses bearer sk-or-v1 token

✓ falls back to native streaming chunks

Tests: 8 passed (8)

Duration: 1.28s (transform 310ms, setup 0ms)

Critic Gate gemini-2.5-pro (independent)

Approved

✓

Validation Successful

Audited by Gemini 2.5 Pro · Diff cleanly isolated

✓ Zero regressions across 5 CLI providers
✓ File size constraint respected (model_list.rs <500 lines)
✓ Project memory conventions fully preserved

Orchestration

How overbench Works

Three distinct roles ensure speed, parallel execution, and strict correctness.

Director splits the task

Feed in a complex feature or refactor. The Director inspects your repo, writes focused task briefs, and assigns subtasks to dedicated agent panes.

Workers build in parallel

Multiple agents run side-by-side in native PTYs. Claude Code drafts types, Cursor Agent writes backend logic, and opencode runs tests at the same time.

Critic verifies before merge

Avoid blind self-validation. An independent Critic from a different model family audits code changes, verifies tests, and gates approval before completion.

Architecture

Built for Serious Multi-Agent Workflows

Engineered with Tauri v2, Rust, and native PTYs for speed, privacy, and zero lag.

Multi-Engine Cockpit

Mix and match CLI tools and APIs. Connect Claude Code, Cursor Agent, Codex, Gemini, opencode, or local Ollama without vendor lock-in.

Dynamic Mosaic Pane Grid

Customizable split-screen layout powered by portable-pty and xterm.js. Split, resize, minimize, and restore pane states per workspace.

Director Harness & Run Inspector

Inspect agent reasoning with append-only event logs, brief specs, task state machines, and complete run diagnostics.

Live Quota & Usage Meters

Never get hit by surprise rate limits. Live polling pins Claude Code session usage and Cursor fast-request meters directly into your top bar.

Integrated Git & Monaco Editor

Built-in Monaco editor for rapid inspections, live git diff tree, staging panel, and project memory indexing across all panes.

Secrets in OS Keychain

Your API keys and credentials are saved in your native Windows Credential Manager or Linux Secret Service via Rust keyring.

Interoperability

Support for Every AI Coding Engine

Connect your existing CLI logins or run local open weights with zero lock-in.

Claude Code Anthropic

Cursor Agent Interactive CLI

Codex OpenAI

Gemini Google DeepMind

opencode Agent Server

Ollama Local Offline

Build faster with a multi-model team

Free, local, and open alpha. Open your workspace, assemble your team of agents, and see parallel pair-programming in action.

Download for Windows (.exe) Linux (.AppImage / .deb)

View Source on GitHub · Release Notes

Run a team of AI coding agents. Side by side.