BossConsole

🔬 The operator's console for AI agents - and the scientists who direct them

The first open-source, multi-platform harness for AI agents - an evolvable toolbox for agents and the humans they work with.

🧪 Built for enterprises, science, and research.

⚡ Natively multi-threaded - built on the JVM, not Electron - with live hot-reload and laptop-to-supercomputer scale.

Bring your own agent - Claude Code, Codex, Gemini, or OpenCode - and give it a real browser, terminal, editor, secrets, and automation. Then decide exactly what each one is allowed to touch. Run it on your own machine, hand a live terminal to your phone with a QR code, and reshape any tool - by hand or by the agent itself - while the app keeps running.

⬇ Download · 📊 Compare · ⚡ Browser benchmark · 🤖 Run an agent · 🔐 Governance · 🧰 Toolbox · 🖥️ BossTerm

Just want to download BOSS? Head to BossConsole-Releases for pre-built installers.

Built with Kotlin Multiplatform and Compose Multiplatform, BOSS unifies an embedded browser, a blazing-fast shareable terminal, a code editor, an extensible Toolbox of plugins, and a governed MCP tool layer into one desktop workspace for complex, AI-assisted work.

How BOSS compares

Every other AI-agent desktop app and agentic IDE - Claude Desktop, OpenAI Codex, Google Antigravity, Cursor, and Windsurf/Devin - is closed source, and most lock you to a single vendor's model. BOSS is the first open-source, multi-platform harness in the category: Apache-2.0, cross-platform, and built to run any agent.

The deeper difference: BOSS is both the agent's home and an agent-operable app. The same MCP layer your agent uses to do work also exposes BOSS itself - its tabs, terminals, browser, editor, git, secrets, and automation - as 100+ mcp__boss__* tools. Read-only tools (list tabs, read a pane's output, tail the console, snapshot performance, inspect git) give an agent live situational awareness of the workspace; action tools let it act on what it finds. So the agent doesn't just chat about your code - it perceives the running app and drives it, carrying out multi-step tasks autonomously.

And unlike most of the field, BOSS bundles a real embedded browser - Fluck - the agent can navigate, script, and automate (browser_navigate, browser_run_js, plus record-and-replay RPA), with logins filled from the Secret Manager and never handed to the model.

It's also native, not Electron. BOSS runs on the JVM with true multithreading and hot-reloads plugins at runtime - so an agent can evolve a tool and see it live, without restarting the app - while Claude Desktop and the VS Code-fork IDEs (Antigravity, Cursor, Windsurf/Devin) run on a single-threaded JavaScript event loop.

✅ yes · △ partial/limited · ❌ no.  ¹ Only OpenAI's Codex CLI is open source (Apache-2.0); the Codex desktop app is proprietary.  ² Claude Desktop's "Computer Use" controls the whole screen rather than bundling a scriptable in-app browser.  ³ Enterprise-plan admin controls exist, but fine-grained per-tool RBAC isn't documented.  ⁴ Codex CLI is cross-platform; the desktop app is macOS/Windows.  ⁵ VS Code-fork IDEs reload extensions via a full window reload; none offers BOSS-style live plugin hot-reload + agent-driven tool evolution.  ⁶ Claude Desktop and the VS Code-fork IDEs run on Electron (Chromium + Node.js) - a single-threaded JavaScript event loop with worker/child-process offloading; BOSS runs natively on the JVM with true multithreading.  Compiled from public sources, July 2026 - these products move fast, so corrections are welcome via issue or PR.

Browser performance

We ran Speedometer 3.1 on the BOSS Fluck browser and five desktop browsers on the same M3 Max Mac. Each reported median below comes from three 10-iteration runs; higher is better.

Treat Fluck and Comet as tied within measurement precision. A follow-up experiment alternated them back-to-back three times: Fluck led each pair, but its 3.4% median margin was within the browsers' own run-to-run spread and was not statistically significant (p = 0.125). The durable signal is that Fluck and Comet both scored roughly 30% above Chrome and Atlas in these conditions.

Absolute scores were depressed by heavy co-tenancy - 800–1300% ambient CPU, including roughly four cores from a running Docker Desktop VM - so compare browsers within this experiment, not against published scores from quiet machines. Read the full benchmark report for the paired test, per-suite analysis, discarded runs, and caveats; the reproducible harness, per-run JSON, and screenshot evidence are committed alongside it.

¹ Safari used a live profile with existing tabs and extensions, so it is not comparable to the fresh-profile runs.  ² Firefox varied by 40% across its three runs; its median is the least stable result.

Contents

• How BOSS compares
• Browser performance
• Downloads
• Why BOSS
• Run any AI coding agent
• You decide what your agents can touch
• MCP - give agents real tools
• Toolbox - an app store inside the app
• BossTerm - a terminal you can share to any device
• Case study: the DNA Origami Inventor
• Design System
• Development
• CLI · CI/CD · Documentation · Open source & ecosystem

Downloads

Download links always fetch the newest release directly, via the latest-release edge function - no API key needed, and not subject to GitHub's 60-request/hour limit for unauthenticated callers. Release metadata as JSON - version, assets, sha256 checksums - is at ?app=boss. Prereleases are excluded by default; add &prerelease=true to consider them, or &channel=beta (alpha, rc) to track one channel. To browse or pin a specific version, see BossConsole-Releases.

A browser resolves the redirect and saves the versioned filename. Command-line tools do not: there is no Content-Disposition header, so `cu