All posts
Daniel Bílek4 min read

Jean2: What It Is, Why It Exists, and What's the Point

Not another coding agent. Or, actually, it can be that too. Jean2 is an open-source agent platform with no baked-in behavior, no vendor lock-in, and tools in any language. Here's the full picture.

announcementai-agentsopen-source

I use AI agents for roughly 30 million tokens a day. Coding, research, ops, automation. If it can be done with an LLM, I've probably tried to do it. I've run through Cursor, OpenCode, Aider and everything in between. I daily-drive GLM, Minimax, GPT models, and whatever's new on OpenRouter.

At some point I realized the tools were great at what they did, but I couldn't make them do what I wanted.

What is Jean2?

Jean2 is an open-source AI agent platform. Apache 2.0.

One server runs on your machine as a daemon. You connect to it from any device (desktop, mobile, browser) and it stays running even when you close the client. 80–100 MB memory footprint. Multiple projects, multiple sessions, all alive.

The core idea: the agent has no baked-in behavior. No hidden system prompts. No fixed toolset. No steering you didn't ask for.

Your system message and your AGENTS.md files are the only things that shape the agent. That means you're not building "your coding agent", you're building your [anything] agent. Same binary, different configuration.

Why does it exist?

I kept running into the same four walls.

Baked-in steering

Coding agents come with pre-written system prompts that steer behavior. Fine for an out-of-the-box experience. Not fine when you want full control. You can write your own agent personality, but your words get appended to whatever the system already baked in. You're never truly in control.

Rigid tooling

Built-in tools can't be removed, modified, or even tweaked. You can bring MCPs or Skills, sure, but the core tooling is locked. These systems aren't designed around the idea that you'd want to bring all your own tools.

Session management

I wanted to handle multiple projects, loose conversations, and long-running tasks in one place. The existing TUI options weren't comfortable for that. (There are better GUI options now, but at the time, there wasn't much.)

Multi-device continuity

I just wanted to draft a spec on my phone, hand it to the agent for execution, and review it on my laptop. Seamlessly. That didn't exist.

How it works

Jean2 is a single binary. One command to install, one command to start.

curl -fsSL https://jean2.ai/install.sh | bash
jean2 init
jean2 start

The server runs on localhost:8742 with REST and WebSocket. Connect any client: desktop app, mobile app, or browser (npx @jean2/client). Works over Tailscale, VPN, or your local network.

Any LLM

OpenAI, Zhipu.ai, Minimax, Google, OpenRouter, Codex. Switch providers and models. Use a cheap model for routine work, a premium model for hard problems. Your keys, your choice.

Tools in any language

A tool is just a directory with a manifest and a script. JSON goes in via stdin, JSON comes out via stdout. That's the entire API. Write tools in Bun, Node, Python, Bash, Go, Rust. Whatever you want. Drop them in ~/.jean2/tools/ and the agent picks them up. No build step, no registration.

You control the tool definition and its description. If a model isn't using a tool right, you just change the description.

MCP & Skills

Connect any MCP server, local or remote, with OAuth support. Define Skills as discoverable instruction sets (SKILL.md files). Both are automatically available to your agent sessions.

Subagent orchestration

Agents can spawn subagents for complex tasks. Isolated sessions that inherit workspace context. Interrupts cascade automatically. Set depth limits per configuration.

Make it yours

System prompts, tools, skills. All files on disk. Edit them, add them, remove them. Download preset bundles for coding, research, or ops. Or start from nothing and build exactly what you need.

What's the point?

Most agent tools make a trade: they give you a polished experience out of the box, and in exchange, they make decisions for you. What the agent optimizes for. What tools it has. How it behaves. What provider you use.

Jean2 makes the opposite trade. You start with a blank slate. No baked-in prompts, no locked-in tools, no vendor requirements. The experience is as polished as you make it.

That's not for everyone. If you want something that works great in five minutes with zero configuration, this probably isn't it. Not yet.

But if you've ever been frustrated because your coding agent won't stop doing something you didn't ask it to do, or because you can't swap a tool out, or because you want to use your agent for more than code. That's the point.

Where things stand

Jean2 is in active development. I've been coding with it daily. The server is stable, the tooling is growing, and the architecture is solid.

Got questions, ideas for tools, or tried it out and hit a wall? I'm at @danielbilekq0.