Skip to content

AgentInk

AgentInk is a self‑evolving, proactive developer agent that lives in Slack. You talk to it like a teammate — @agentink do X — and it reasons about the request, calls the right tools, and reports back. It can research, write and ship code, drive an Android device, generate images, run live meetings, manage tasks, and more. Crucially, it can also modify and extend itself: you can ask it to build new capabilities, and it will branch, implement, test, and open a PR.

Ground truth (no marketing)

  • Reasoning: Google Gemini via Vertex AI (google‑genai SDK). Anthropic/Claude is not its reasoning engine — it's an optional per‑user credential feature only.
  • Memory & state: Neo4j (knowledge graph) + PostgreSQL with row‑level security (encrypted per‑user secrets, profiles, conversation history).
  • Runtime: a single inkwell.service on one GCP VM, reachable in Slack via Socket Mode — primarily #agentink-live-development.
  • Image generation: Nano Banana 2 (Gemini 3 Pro Image).
  • AgentInk is instructed to never fabricate capabilities or success — if it can't verify something, it says so.

Three ways to think about it

  • :material-message-text: Use it

    Ask it to do real work in Slack — research, images, code, meetings, tasks, dashboards. See Getting Started and What It Can Do.

  • :material-tools: Extend it

    Ask it to add a tool, fix a bug, or change its own behavior. It follows a real dev workflow (branch → test → PR → review → deploy). See Iterating & Extending.

  • :material-monitor-dashboard: Watch it

    Every agent's inner monologue and all subsystem activity stream to the Control Center dashboard — live and as searchable history. See The Dashboard.

A 30‑second tour

  1. In Slack, mention it: @agentink summarize what you worked on with Vera this week.
  2. It plans, calls tools (reading its knowledge graph, searching, etc.), and replies — its reasoning streams to the dashboard as it goes.
  3. Ask it to do something consequential (unlock a door, send an external message, run a trade): it pauses for a named human's approval first (the Consequence Gate).
  4. Ask it to improve itself: @agentink add a tool that posts a weekly RSS digest — it proposes the change, branches, implements, and opens a PR for review.

Continue to Getting Started →.