Several unrelated projects use the name OpenTag. This guide covers Amplift OpenTag, an MIT-licensed system for invoking AI agents from collaboration tools and returning results to the originating conversation.
Its value is contextual continuity. A request can begin in Slack, GitHub, GitLab or another supported channel, be dispatched to an agent runtime, and return where the team already works. The repository includes public connector packages and an ACP-oriented runtime.
The operational model
OpenTag handles inbound events, maps identities and conversations, invokes a runner, and publishes the response. Before deployment, decide which channels and repositories may trigger it, which users are authorized, what credentials the runner can access and what audit record is retained.
Start with a private test channel and read-only repository token. Configure one connector, restrict triggers to a small group, and test duplicate delivery, retries and thread mapping. Collaboration platforms can resend webhooks; the integration must avoid duplicated work and comments.
Security boundaries
Messages and issue comments are untrusted input. A prompt in a thread must not broaden repository permissions or reveal connector secrets. Keep credentials out of conversation context, use separate service accounts and apply least privilege.
Human review remains appropriate before an agent merges code, changes infrastructure or contacts people outside the original thread. OpenTag shortens the path from discussion to action, which makes approval boundaries more important.
It fits teams whose engineering work already lives in issue trackers and chat. A direct client is simpler for individual use. Compare it with OpenWork’s workspace model: OpenTag brings agents into conversations, while OpenWork organizes reusable capabilities.
Primary source: Amplift OpenTag repository.
Deployment checklist
Use signed webhook verification, idempotency keys and bounded retries. Map platform identities to agent permissions explicitly instead of trusting display names. Log the triggering message, selected runner and resulting action without recording secrets. Add rate limits per user and channel, plus a kill switch for outbound actions. Test edited comments, deleted messages and permission changes because real collaboration threads rarely stay static.