Agent Context Sync pools sessions from Hermes, DeepSeek Harness, OpenCode and other agents into one workspace and syncs them across devices — conversations started at the office are waiting when you get home, and any agent's sessions can be viewed, searched or exported on the web.
$ python mcp/server.py · syncs automatically on start
$ python mcp/server.py # start the MCP client → incremental pull on start · then auto-sync every 5 minutes → or trigger manually from your agent with sync_full ✓ sync complete — incremental merge · idempotent dedup · full-pool Full-pool sync local: 12 sessions / 348 messages remote: 128 sessions / 3,462 messages · shared by 7 agents
Session infrastructure for heavy AI users: secure, isolated, controlled.
Sessions from Hermes, WorkBuddy, DeepSeek Harness, OpenCode, Reasonix, OpenClaw and Oh My Pi land in one workspace; each client pulls the whole session pool (every agent), and the web UI lets you view, search and export everything.
Conversations at the office are pulled to your home machine automatically. Cross-device continuations merge new messages seamlessly — incremental sync plus idempotent dedup, instant and safe.
Self-hosted: all data stays on your own server. Workspaces are isolated, each with its own API key — even admins cannot read your content.
A user → workspace two-level model: a team shares one server while staying fully isolated. Registration is open by default, invite codes optional, and quota enforcement works per plan.
Adding an agent is a single adapter file — zero server or sync-engine changes. Claude Code, Cursor… plug in anytime.
Clients pull new versions on a schedule, verify each file with SHA256, replace atomically, and take effect on the next agent restart — old versions are backed up for rollback.
Full-pool sync, cross-device continuations, huge batches that never time out — every layer is designed for stability and idempotency.
Clients pull the entire session pool (every agent) and route sessions back into each local store — a session from machine A can be pulled and written locally by machine B.
When a session continues on another device, only new messages are pushed and merged into the existing remote session — cross-device conversations stay seamless.
Pulls are paged 15 sessions at a time; pushes are batched by both session count and message count — large syncs never time out.
Each watermark records which server it belongs to; switching servers triggers a full re-pull, so stale watermarks can never suppress sessions on a new server.
Messages are deduped by (session, role, timestamp) triple plus content fallback, with a unique index as the last line of defense — never stored twice.
With dual instances only one process runs background sync, avoiding local-store contention; write-lock conflicts retry instead of blocking.
Every agent — Hermes, WorkBuddy, DeepSeek Harness, OpenCode, Reasonix, OpenClaw, Oh My Pi — is supported. Deploy one client per agent; point them at the same workspace and they sync with each other.
The onboarding an open-source project deserves: one command, one config.
One-command deployment on your own server; an admin account and workspace API key are created automatically:bash deploy-server.sh
In the Setup Help page, download the client package for your agent (server address pre-filled), then fill in the workspace key:export HERMES_SYNC_API_KEY=ws_xxx
It pulls incrementally on start, pushes local data for first-time pairing when the remote is empty, then syncs every 5 minutes.
Open the management UI in your browser: session viewer with Markdown rendering, one-click export/import, and a clear device overview.
MIT-licensed client · open server source · free self-hosting
Agent Context Sync promises: free forever for personal self-hosting, transparent and auditable code. Join the community and turn your agents' context into an asset.
From today, every AI assistant you use has full context memory — across devices, across tools, never lost.
Sign in →