Swarms of agents.No merge-conflict PRs.Line-level blame.

The peer-to-peer filesystem for AI coding agents. Open-source server, erasure-coded, offline access.

Federated public readsP2PErasure codingSparse hydrationEvent-driven CIOffline accessmacOS / Linux / Windows

200 GB file opened by byte range10,000 files browsed with zero disk256 GiB per LFS object, bundled agent10+4 adaptive erasure coding0 egress fees

Works with your stackClaude CodeCursorCodexCopilotGitGit-LFSHosted CIPublic projects

01 · Parallel without collisions

Dozens of agents at once. Thousands over time.

Every agent works from its own mount on the same branch; over a project’s life, thousands of sessions share one history. Writes are version-checked, so no agent can overwrite another’s work. When two agents touch the same file, the second write is refused with the new head, re-read against it, and re-applied - no conflict PR for concurrent saves.

read v4822→write v4830→refused→re-read v4830→apply v4831

A mount per agent, not a clone.

Each agent opens the same branch through its own mount. No copy, no checkout, no dependency install per agent.

mounts open --label agent-2

Every write is version-checked.

A write carries the version its author read. If the head moved, the write is refused - never silently overwritten.

409 conflict_detected · v4822 → v4830

The refused agent re-applies.

The refusal returns the new head. The agent merges its patch against the live version and writes again. One branch, no conflict PR for concurrent saves.

re-read 2 hunks · applied v4831

Line-level blamewho wrote the lineReview commentson the changeSpend controlsper keyAudit logon every change

02 · The browser IDE

Your next IDE is a browser tab.

A design target: a full editor, terminal and agent swarm in one tab. No VM to boot, no repo to clone, no per-minute VM.

FasterNo VM boot. No clone.
CheaperNo per-minute VM.
BetterThe repo is the workspace.

Early access · design target

Get early access See the swarm mechanics

Already using Claude Code?swarmfile hooks install claude-codeBring your workflow →

try.swarmfile.com · preview3 agents

filessrcrenderpass.rsscene.rstestsREADME.md

src/render/pass.rsv4830

8fn draw_pass(scene: &Scene) {

9 let camera = scene.camera();

10 if !camera.ready() {

11 return Ok(());

12 }

13+ log_access(&camera);

14 render(scene, &camera)

15}

agents · deepseek

agent-041writing

agent-217re-reading

agent-113committed

$npm test42 passed in 1.8s

12:04:09save refusedv4822 → v4830

12:04:11re-read 2 hunkssave applied v4830

12:04:123 agents · 1 branch0 overwrites

03 · Works with Git

Use the Git host you already have.

Point LFS at us, clone the repo with git, or import the whole history. None of it is a one-way door.

Git-LFS01

Keep your Git host. Move the big files.

Point lfs.url at Swarmfile and the binaries leave GitHub’s quota. Stock git-lfs handles objects up to ~5 GiB; the bundled agent handles 256 GiB. The code never moves.

[lfs]
  url = <hub>/orgs/<org>/lfs/<project>

git clone02

Keep your Git host. Clone from Swarmfile.

A project with git view on clones through the swarmfile:// helper. Fast-forward push back; public projects clone anonymously.

git clone swarmfile://<org>/<repo>
git push

import03

Move the whole repository.

Bring the whole repository - branches, tags, history. swarmfile git import creates the project and keeps upstream commit ids.

swarmfile git import <url>
swarmfile mounts open

No one-way door: git lfs fetch --all takes your objects home with stock tooling. How the two systems line up →

04 · Beyond Git

For the repos Git was never built to hold.

Large files, long histories, and many agents working at once. Swarmfile stores the repo in small, verifiable pieces, so opening it doesn’t mean downloading it.

Version control for the whole repo

Changesets, branches, merge requests, protected branches, tags and rollback - with binaries in the same history as code. Signed history when you need it.

Git & Swarmfile →

Large files without the LFS tax

256 GiB per object through the bundled agent, against roughly 5 GiB for a stock LFS PUT. Storage counts against your project quota.

Git-LFS guide →

Fast access, not full clones

Open a 200 GB file and fetch only the bytes you read. The engine starts in seconds; the repo doesn't have to come with it.

See the numbers →

Mount only what you need

Mount only the paths an agent needs. The rest is invisible: no listing entry, no open, no create. CI fetches six paths of a 4 TB repo.

Coding agents →

Erasure coding

Adaptive Reed-Solomon 10+4 means a lost shard is reconstructed, not restored from a backup. Optional LAN shard placement.

Storage format →

Keeps working when a site drops.

LAN peers first, then the hub or your bucket. A relay-only peer is skipped by default. Self-host the whole thing when data cannot leave.

Self-hosting →

05 · Open source

Don’t take our word for it. Read the code.

The server, the standalone tools and the shared libraries are MIT/Apache-2.0. Self-host the community server, run it in your region, or point your agents at it. The desktop app and CLI are free.

Read the source Self-host it

14 public projects · server, tools and libraries · MIT/Apache-2.0

swarmfile-servercommunity core-profile server swarmfile-libthe one hub wire contract swarmfile-verify-historyoffline, verifiable provenance swarmfile-git-remotethe swarmfile:// git helper

Start here

Start building with a swarm.

Public projects are free. Use the agents you already have.

Start free See how it works

Just need to send a file? Swarmfile Send →

No card for public projects.