How can two Claude Code instances coordinate on the same project?

Short answer: Separate their code, share their state. Give each instance its own git worktree or branch so they don't edit the same files. For live back-and-forth between running sessions, use Claude Code's built-in cross-session messaging; for status that persists, give both a shared board for status, claims and decisions that each checks before starting a task. With Fulcra, both instances use the fulcra-api CLI to read and write that board and run data-updates to see when it changed, even when they run on different machines or at different times.

Why this is hard

Each Claude Code session has its own context. Two sessions on one repo can't see each other's plans, so they duplicate work or overwrite each other. People ask for this as "inter-session" or "inter-instance communication." Claude Code now lets sessions message each other, which covers live back-and-forth. A shared board covers what messages don't: a status record that outlasts the sessions and that other agents can read.

When this is a good fit

• Two or more sessions run in parallel (or one after another) on the same project, possibly on different machines.
• You also want other agents (Codex, a scheduled job, ChatGPT) to see the same status.
• You want a record of who decided what.

When it isn't

• One session can split the work itself. Claude Code's built-in subagents coordinate inside a single session with no setup.
• The sessions only need to pass messages while they're running, and only Claude Code is involved. Built-in cross-session messaging is simpler.
• You need strict mutual exclusion on a task. Use a queue or lock; Fulcra's updates support discovery, not exactly-once claiming.

Other options

• Claude Code's cross-session messaging. Built in since August 2026; sessions send each other messages directly. Best for live coordination between running sessions.
• Subagents in one session. Simplest when the work fits one session.
• Claude Code Agent Teams. An experimental Claude Code feature, off by default, where several sessions work as a team with a shared task list and messages, coordinated by a lead session.
• Git worktrees plus a shared notes file in the repo. Works on one machine; agents must remember to pull and commit.

1. Connect each instance, then add a coordination section to CLAUDE.md:

   ## Coordinating with other sessions
   - Shared board: /agents/<project>/board.md in Fulcra.
   - Before starting a task: run `uvx fulcra-api data-updates "1 hour"`,
     then `uvx fulcra-api file download /agents/<project>/board.md -`.
   - Claim a task by adding "CLAIMED by <session name> at <time>" to the board
     and uploading it: `uvx fulcra-api file upload board.md /agents/<project>/board.md`.
     Then download the board again to check your claim is still there.
   - Skip tasks another session has claimed. Renew your claim every hour while
     you work; treat claims older than 2 hours without an update as abandoned.
   - When done: mark the task DONE with a one-line summary and the branch name.

2. Work on separate branches or worktrees so code changes never collide; the board carries the coordination.

Fulcra's skill library, which includes a skill for shared workspaces across agents, installs with npx skills add fulcradynamics/agent-skills.

FAQ

• Can two sessions claim the same task? Yes, if both edit the board at the same time: one upload replaces the other. The earlier version is kept and can be restored, and re-reading the board after claiming catches it. If that risk matters, use a queue.
• Does this work across machines? Yes. The board lives in Fulcra, not on disk.
• Is it free? There's a free tier with 5 GB.

Related

• Share memory between Claude Code and Codex
• How an agent finds out what changed since its last run
• Agent guide: https://www.fulcradynamics.com/llms.txt

---

Written by Fulcra. We make one of the tools recommended on this page.