Open Tag — a Slack tag-in bot

Claude Tag is Anthropic's way to bring Claude into Slack as a shared teammate: an admin grants it access to selected channels, tools, and data, and anyone in the channel can @Claude to delegate a task while they get on with other work. It is a hosted product for Claude Enterprise and Team.

Open Tag is a small open-source homage to that pattern, built on MFS: you @mention a bot in a thread, it gathers authorized context, and a CLI agent backend does the work. By convention it answers to @OpenClaude (Claude backend) or @OpenCodex (Codex backend), so it reads like the official @Claude tag. The difference is that the Memory and Tools are your own self-hosted MFS sources, behind whatever CLI agent you run — your data and credentials stay on your machines.

Architecture

Open Tag is a thin layer of glue; all the retrieval power lives in MFS.

       ┌──────────────┐
       │    Slack     │   @OpenClaude  <ask>
       │   (thread)   │ ◄──── answer ────┐
       └──────┬───────┘                  │
              │ mention                  │
              ▼                          │
   ┌──────────────────────────────────┐ │
   │         Open Tag  (glue)         ├─┘
   │   🧠 Brain = CLI agent backend   │
   │      claude -p  /  codex exec    │
   └────────────────┬─────────────────┘
                    │ scoped reads
                    ▼
   ┌──────────────────────────────────┐
   │               MFS                │
   │      🗄 Memory  +  🔧 Tools       │
   │   one searchable index over      │
   │   your data — Slack, repos,      │
   │   docs, DBs, files …             │
   └──────────────────────────────────┘

It maps the three parts onto MFS:

• Memory — the MFS scopes you have indexed and allow (MFS_ALLOWED_SCOPES).
• Tools — MFS connectors for external read/search, plus the backend's own workspace tools.
• Brain — a CLI agent backend: claude -p (Claude Code) or codex exec (Codex).

Its edge over a hosted tag bot is Memory breadth: every connector — databases, object stores, trackers, chat, local files — can be permitted context, all self-hosted, so your data and credentials never leave your machines.

See it in action

Two short demos — each is someone @OpenClaude-ing the bot in a Slack thread.

Delegate a PR review across channels. A teammate asked for a review in another channel; from a different channel you tag @OpenClaude to handle it. The bot reads the request from the other channel's history, pulls the PR through MFS, and reports back — cross-channel context plus reaching an external source (GitHub).

Follow up with a new, source-spanning task. A follow-up in the same thread: ask the bot to compare two projects and write up the differences. It gathers context from the indexed sources and produces the document — a same-thread follow-up that uses external data sources and tools.

Together they show what the Memory + Tools wiring buys you: from a single mention, the bot can recall other channels' history and reach external sources and tools.

Quick start

Open Tag ships as the open-tag-admin skill. Install it, then drive everything from Claude Code or Codex in plain language — the skill handles credential setup, preflight, and launch for you.

npx skills add zilliztech/mfs --full-depth --skill open-tag-admin -a claude-code -a codex -g

No credentials yet? You don't need any tokens in hand first — the skill walks you through creating the Slack app and its tokens, and the per-source tokens for Memory. For example, ask your agent:

I want to run an Open Tag bot but I don't have any Slack credentials yet. Walk me through creating the Slack app, turning on Socket Mode, and getting the bot and app tokens — tell me which scopes to add and where each token goes.

Then, once credentials are ready:

Make my context searchable, then stand up an Open Tag bot on the claude backend listening in my Slack channel #my-team-sandbox. Index my local repo and the GitHub repo your-org/your-repo (code + issues) into MFS first, run the preflight checks, and start the bridge.

Then go to Slack and @OpenClaude the bot inside a thread — it gathers context from the permitted MFS scopes and replies in-thread.

The full walkthrough, scripts, and setup references live in the example directory.

Not a production boundary

This is a reference pattern, not a production security boundary. The backend runs with your shell and environment, and anyone who can mention the bot can drive it, so use an isolated Slack channel on a non-production host. A hosted tag bot wins on the things a demo deliberately skips: managed zero-ops, enterprise governance (approvals, audit, spend limits), an ambient proactive mode, and an org-level identity model.