KanAI: an AI assistant for your Kanboard, that runs on your own machine
Hi everyone
I’d like to share a plugin I’ve been building as a hobby-project: KanAI connects Kanboard to an LLM so you can "chat about a project" and let an "AI-assistant help you maintain the board", while keeping your data on your own server.
SIDE NOTE: I’m still testing it myself with different LLM’s, smaller models work faster, but depending on the tasks, bigger models give better quality suggestions. I’ve tested 7B, 12B, 14B and 16B LLM’s, and want to try 27B+ at a later stage later.
It’s local-first: the default provider is your own OpenAI-compatible server (Ollama, LM Studio, vLLM, …), so by default no project data ever leaves your machine. External providers (OpenAI, Anthropic Claude) are optional and off until an admin explicitly enables them.
Repo: GitHub - k1bot2026/kanboard-plugin-kanai: AI assistant & project Q&A (RAG) for Kanboard — local LLM first, optional external providers · GitHub · License: MIT
What it does
- Shared chat per project — a KanAI tab next to Overview / Board / List. Multiple conversations (create, rename, delete), shared with every project member so a colleague can pick up where you left off. Each message shows the sender’s avatar; answers render as Markdown and turn task references like
#42into clickable links. - Ask anything (RAG) — questions are answered from the project’s own tasks, descriptions, comments and subtasks.
- Assistant with human approval — ask it to tidy up and it proposes actions (create / update / close / move / assign tasks, tags, due dates, comments, subtasks, task links). Nothing is applied until you tick the proposals you want; they’re then executed through Kanboard’s own models as your user, so permissions and activity streams work normally.
- Quick actions — one-click presets: project summary, what’s next, board health, stand-up notes, cleanup suggestions, risks & blockers, workload, organize, enrich, help.
- Optional daily digest — a
kanai:digestconsole command (for cron) posts a status/cleanup/risks digest with proposals per opted-in project.
Privacy & safety
I care about self-hosting, because project data is very sensitive and confidential. So this was the priority:
- Local by default — nothing leaves your server unless an admin turns on an external provider.
- Admin-controlled policy — only the application administrator decides the external-AI policy (disabled / per-project opt-in / forced-on). Project admins cannot enable external AI the app admin has locked down; it’s enforced server-side.
- Human-approval gate — no model output ever changes the board on its own; project content is treated as untrusted input (prompt-injection framing).
- API keys encrypted at rest (AES-256-GCM), masked in the UI, never sent to the browser.
- Rate limiting per user, CSRF-validated actions, and Kanboard’s normal ACLs (members chat, managers configure, admins manage providers).
Install
Copy or symlink the repo into your Kanboard plugins/ directory as KanAI:
ln -s /path/to/KanAI <kanboard>/plugins/KanAI
Then enable it per project under Project settings → KanAI Settings, and point it at your LLM under Settings → KanAI (there’s a “Test connection” button that also discovers your local models).
Requirements: Kanboard ≥ 1.2.46, PHP ≥ 7.4 with cURL and OpenSSL, and an LLM endpoint (e.g. Ollama). A ~7–12B model answers in seconds on modest hardware; the repo includes a model guide with measured latency/quality from 7B to predictions up to 80B.
Feedback welcome
It’s fully open-source (MIT) and I’d love feedback, bug reports, and ideas. Especially from people running it against different local models or larger boards. Issues and PRs are open on GitHub.
Thanks for taking a look!
Kind regards,
K1.
