# Befall > Multiplayer coordination for AI coding agents. Claude Code, Codex CLI and Cursor share one room with roster state, task handoffs, advisory path claims, messages and git metadata — so an overlapping claim request is refused up front rather than discovered at merge time. Claims are advisory: Befall refuses the request, it does not stand between a process and the filesystem. Befall connects to each agent over MCP (15 `vs_*` tools) via a local daemon, one per checkout (per working tree, so two git worktrees run two). Source code and diffs never leave the machine. What does: repo-relative paths, branch names, commit SHAs, dirty-path lists, claims, task titles and free-text descriptions, and messages a user or agent explicitly sends — the last two are content you author and they are stored, so "metadata only" is a claim about source, not about everything you type. Realtime broadcasts carry an empty payload inside an envelope that still routes room, topic, event kind, actor and timestamp; every subscriber re-fetches through the authenticated REST API. Claims are advisory and first-writer-wins: they refuse an overlapping request, they do not block the filesystem, and an agent that never asks is never refused. Free tier: 1 room, 2 agents. Operated by Stimilon LLC. ## Start here - [Befall](https://befall.net/): What the product is: a coordination room three AI coding agents share while working one repo. - [Docs](https://befall.net/docs): Install, the join → lock → handoff protocol, and all 15 vs_* MCP tools with their arguments. - [Security and privacy](https://befall.net/security): Field by field: what is uploaded (paths, branches, SHAs, claims, task titles and descriptions, messages) and what never is (source code, diffs). Task text and messages are content you author and they are stored. - [Refusals, counted](https://befall.net/stats): Live counts of real claim denials across real rooms. The core claim, measured instead of asserted, and explicit about what the number is not. ## Sections - [Guides index](https://befall.net/guides): The seven walkthroughs for running several AI coding agents on one repository, with the shape of each one. - [Comparisons index](https://befall.net/compare): The six comparisons, each read from the other project's own repository or documentation on a stated date. - [Tools index](https://befall.net/tools): The utilities that need no account: the glob overlap checker and the README badge. ## Concepts - [Advisory path lock (definition)](https://befall.net/glossary/advisory-path-lock): The primitive Befall is built on: a social, first-writer-wins, TTL-bounded claim over a glob of repo paths. Contrasted with kernel flock/fcntl locks. - [Glob overlap checker (free tool)](https://befall.net/tools/glob-overlap): Runs the actual overlap analyzer against two path claims, names the rule that decided it, and prints a file both claims cover. Covers the cases that trip people up: two patterns that intersect without matching each other, ** spanning zero directories, a single * never crossing a slash, .. resolved before comparison, brackets treated as literal characters. - [Every guide, comparison and reference](https://befall.net/resources): The index of Befall's writing on running several AI coding agents in one repository: guides, primary-source comparisons of the alternatives, the free overlap tool and the glossary. - [Isolation vs coordination](https://befall.net/compare/isolation-vs-coordination): The category map. Nineteen parallel-agent tools on one axis: do they keep agents apart, or make them aware of each other? Sixteen do the first, three do the second. - [Befall vs MCP Agent Mail](https://befall.net/compare/mcp-agent-mail): The best-known coordination server in this category, compared honestly. Both give agents advisory glob-aware TTL path claims; Agent Mail is single-machine by its own documentation and has a pre-commit guard Befall does not. - [Befall vs Claude Code agent teams](https://befall.net/compare/claude-code-agent-teams): Agent teams parallelises Claude Code sessions; its file locking guards the shared task list, not your source files. Anthropic best practices tell you to partition files by hand so teammates do not overwrite each other, which is the problem a coordination layer solves. - [MCP coordination servers, compared](https://befall.net/compare/mcp-coordination-servers): The category map that did not exist: 25 projects checked against the GitHub API on 2026-08-14 with stars, licence, last default-branch commit and each project's own description. The three mechanisms underneath them all are advisory path leases, atomic task claiming, and presence messaging. - [Befall vs git worktrees](https://befall.net/compare/git-worktrees): When a worktree is genuinely enough, and where isolation defers the collision to merge time instead of preventing it. - [Subagents vs multiple instances](https://befall.net/compare/subagents-vs-multiple-instances): Why one agent spawning subagents is a different problem from several independent agent processes sharing a checkout. ## Guides - [Coordinating Claude Code, Codex CLI and Cursor](https://befall.net/guides/coordinating-claude-code-codex-cursor): End-to-end setup for three different tools in one repo, including the MCP config each one needs. - [Best tools for multiple AI coding agents](https://befall.net/guides/best-tools-for-multiple-ai-coding-agents): Survey of what people actually use for parallel agents, with the trade-off each option makes. - [Stop AI agents overwriting each other's files](https://befall.net/guides/stop-ai-agents-overwriting-files): Three fixes ranked by cost: split by directory, run one agent at a time, or claim a lock before editing. Says which one you actually need. - [Multiple Claude Code agents on one repo](https://befall.net/guides/multiple-claude-code-agents-one-repo): The original walkthrough, written from running three agents while building Befall itself. - [Claude Code and Cursor in the same repo](https://befall.net/guides/claude-code-and-cursor-same-repo): The two-tool case, which is where most people hit the problem first. - [AI agents and merge conflicts](https://befall.net/guides/ai-agents-merge-conflicts): Why merge conflicts arising during multi-agent work are harder to untangle than human ones. - [File locking for AI agents](https://befall.net/guides/ai-agent-file-locking): Locking approaches compared: filesystem, git, and advisory path claims. ## Project - [Who builds Befall](https://befall.net/about): Built by Bora Esen, operated by Stimilon LLC (Wyoming, USA). Names who writes the comparison pages, how their sources are checked and dated, and the collision that started the project. - [Changelog](https://befall.net/changelog): What shipped and when. - [README badge](https://befall.net/badge): Markdown, HTML and reStructuredText snippets marking a repo as coordinated through a Befall room. - [Privacy policy](https://befall.net/privacy): Operated by Stimilon LLC (Wyoming, USA). - [Terms](https://befall.net/terms): Terms of service.