# agentsync (remote) [Health: Active]

**Category:** ☁️ Cloud Platforms  
**Repository:** https://github.com/jarmstrong158/agentsync-remote  
**GitHub Stars:** 1  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/agentsync-remote

## Description
Self-hosted Cloudflare Worker remote transport for agentsync (streamable-HTTP).

## Claude Desktop Quick Installation
No confirmed install details yet, so no config is published. Follow the setup steps in the repository above.

## Documentation

## What agentsync (remote) does

The agentsync (remote) MCP server adds a remote peer to an existing agentsync coordination mesh. It lets claude.ai, including a phone-based session, coordinate work with local agents through one shared `claims.json` file. The remote peer uses the same coordination branch and claim rules as local agentsync, so claims from both transports appear together and are distinguished by agent ID.

The Worker does not keep coordination state itself. The shared repository contains the coordination file, while the Worker provides the MCP transport and applies updates through GitHub. No Git installation or local clone is needed for the remote peer.

## How it works

The deployed Cloudflare Worker exposes seven tools over Streamable HTTP. Requests are authenticated by placing an `AUTH_TOKEN` in the `/mcp/<token>` URL path. The Worker uses `GH_PAT` to access the configured GitHub repository through the Contents API. It can bootstrap the coordination branch and `claims.json` in a new, otherwise empty repository.

Claim updates use compare-and-swap behavior against the shared coordination data. A claim that overlaps an active peer is blocked, while completed claims no longer prevent other work. Local and remote commits are shown together in the coordination history.

## Setup and configuration

Deploy the repository to your own Cloudflare account using the provided Cloudflare deployment flow. The deployment forks the repository, creates the Worker, and can connect Workers Builds for later redeployments. Configure `REPO` with the coordination repository in `owner/name` form. `AGENT_ID`, `BRANCH`, and `CLAIMS_PATH` identify the remote peer and coordination data; the documented defaults are `jonny-mobile`, `agentsync`, and `claims.json`.

Add `AUTH_TOKEN` and `GH_PAT` as Worker secrets after deployment. The GitHub token should be fine-grained, restricted to the coordination repository, and limited to Contents read and write access. Add the Worker URL with `/mcp/<AUTH_TOKEN>` to claude.ai as a custom connector, then ask Claude to call `survey` to test it.

## Tools and capabilities

The agentsync (remote) MCP server provides these tools:

- `survey` returns the board, including peer claims, conflicts, and mailbox contents.
- `claim` requests work while checking for active overlapping claims.
- `check_conflicts` rechecks a claim, optionally for a specific branch.
- `update_status` changes a claim to `planning`, `in-progress`, or `done`.
- `release` marks a claim complete.
- `history` lists recent commits on the coordination branch.
- `mailbox` supports notes between human or agent participants.

## Limitations and notes

The connector URL is itself a credential. Anyone who obtains it can call the Worker as a full mesh peer, including claiming work, releasing claims, and posting mailbox messages. The Worker masks the token in its own route logs, but URLs may still appear in browser history, proxy logs, screenshots, transcripts, or other external records.

Rotate `AUTH_TOKEN` by changing the secret and redeploying the Worker; old connector URLs then return `404`. Rotate it immediately after exposing the URL. The Worker also returns `404` for every request until `AUTH_TOKEN` is configured, and tools report a missing-secret error when `GH_PAT` has not been set.

_Full upstream README: https://allmcps.com/mcp/agentsync-remote/readme_