About this project

Deskmate is a framework for deploying a team of specialized AI agents (deskmates) into Slack. Built on Vercel Eve, it uses a "front-desk" router to delegate user requests to specific agents based on their roles and descriptions. Key capabilities include: - **Data Integration**: Connects to external data sources using the Model Context Protocol (MCP), with a built-in catalog of roles for tools like Mixpanel, Sentry, Intercom, PostHog, and Linear. - **Human-in-the-Loop**: Implements a safety mechanism where any "write" operation (defined by `approval: always()`) pauses for human approval via Slack buttons before execution. - **Agentic Coding**: Optional capability to clone GitHub repositories into a sandbox, create feature branches, and open pull requests for review. - **Memory**: Opt-in cross-thread recall using a tiered memory system (working, core, recall, archival) persisted via Postgres. - **Proactive Watching**: Optional ability for agents to monitor specific channels and intervene with reactions or replies based on an LLM-driven attention gate. - **Developer Workflow**: Managed via an npm package (`@deskmate/core` and `@deskmate/cli`) using a central `deskmate.config.ts` file to define the team roster and connections.