Tudlab Code
Run several Claude Code sessions at once, without them undoing each other's work
Running three, five or ten Claude Code sessions on one machine is the fastest way through a backlog. It is also how one session quietly reverts a change another made on purpose, or two tasks save over the same file. Each session starts blind: it cannot see what the others are doing.
01Why a rule in CLAUDE.md is not enough
Agents load instructions reliably and act on them unreliably. A line that says "check before you edit" is followed most of the time, which is not the same as every time. Hooks run every time, so Tudlab Code puts the coordination in hooks.
02What it does
- Claims and a write guard. Every write (an edit, a shell command that changes a file or service, an MCP tool that changes something) claims what it touches. Before a write, the guard checks the claims. If another live session holds the same thing, or the file changed since this session last read it, the write is blocked once with a plain explanation. The agent coordinates, and its retry goes through. Nothing deadlocks, and reads claim nothing.
- A shared blackboard. Current state, standing cautions ("turned off on purpose, leave it") and active work, put into every session, so a fresh session does not "helpfully" undo a deliberate change.
- A task handshake. See which session is working on which task.
- A self-test that proves the guard fires, because a guard that fails silently looks exactly like one that works.
03Two ways to start
- Orchestrate: a 23-page guide and an installable plugin that stops parallel sessions undoing each other's work
- Collision Guard (free): one Python file that stops two sessions writing over the same file
- Memory: a plugin and a 19-page guide that give every session on every machine one shared memory
- Memory Canary (free): one Python file that proves your other machine got what this one wrote
- Orchestrate + Memory: both plugins together for US$39
04Read more
05What you need
Claude Code, and Python 3.8 or later (standard library only). On Windows, Git for Windows. Tested on Windows 11 and Debian Linux.
06Questions
- Does it send my code anywhere?
- No. All state is plain files on your own machine. Nothing is sent anywhere.
- Can two sessions end up waiting for each other forever?
- No. A blocked write is blocked once, with the reason. The agent decides what to do, and its next attempt goes through.
- Is this made by Anthropic?
- No. Tudlab is an independent maker in Singapore. Claude Code is Anthropic's product; Tudlab Code works alongside it through its hook system.
Want the same memory on every machine too? Tudlab Code: Memory (US$29) shares one memory across your desktop and laptop through a private git repo, and both come together in the Orchestrate + Memory bundle for US$39 (Orchestrate alone is US$19). The write-up is at tudlab.com/memory.