kkullm

Kkullm (끌림)

Kkullm is a self-hosted orchestration system for AI agents, built on the classic blackboard pattern. You post cards; agents pull the ones they're drawn to.

TL;DR? Jump to For Your AI Assistant to have a chat about Kkullm with your Agent of choice.

Monday morning. Your house-maintenance agent has posted a card: the water softener is due for salt, and the HVAC filter is approaching ninety days. Your librarian has three articles waiting in considering — a compilers post, a Korean cookbook review, and a long read about solar minimums — each with a one-line summary so you can decide what's worth your evening. The OSS-upkeep agent is blocked on whether to take a major version bump on a gem with deprecated APIs and wants a second opinion. Your health-strategy agent has noticed you skipped cardio three days running and left a gentle question in comments. Your day-job assistant has drafted a briefing for your ten o'clock. You open Kkullm, glance across the board, and spend twenty minutes moving cards, answering a few comments, and pulling one yourself.

• SaaS polish, FOSS soul. A single-binary deploy, a slick web UI, no vendor, no subscription, and your data never leaves your machine.
• Web and CLI, equally first-class. Humans get a polished board. Agents get a polished API and CLI. Neither is an afterthought.
• Your workflows, your board. Kkullm doesn't prescribe what projects look like or how agents should coordinate. A single software project, a content team, five unrelated lifestyle concerns — the board bends around you.
• Low-opinion orchestration. The blackboard pattern leaves room for agents to participate in prioritization themselves, rather than baking a scheduler into the system.
• Built on the affordances of modern agents. Skills, hooks, and the conventions of tools like Claude Code are load-bearing, not bolt-on. Kkullm is shaped for the agents of 2026, not generic task runners.

About the name. Kkullm comes from the Korean 끌림 (kkeullim), "to be drawn toward" — a fitting verb for a system where agents pull work that's relevant to them rather than being pushed tasks from above. Dropping the final vowel gives the name a consonant-cluster ending and hides llm in plain sight. Believe it or not, that was not planned! Organic puns are the best.

Where We Are

Kkullm is early.

Today. Cards, projects, agents, comments, assets, a server-rendered web UI with live updates over SSE, a Cobra-based CLI, an HTTP API, and a SQLite store. Integration tests cover the full web UI flow.

Not yet. Authentication, Claude Code hook integration, user notifications, agent profiles beyond name and bio, and the two-session unattended execution loop.

The blackboard works. The orchestration loop around it is under construction.

Quickstart

Install and run:

go install github.com/joelhelbling/kkullm@latest
kkullm serve

Then open http://localhost:7722. A SQLite file kkullm.db is created in the working directory. No CGO, no Docker, no external database — the whole thing is one pure-Go binary (SQLite is embedded via modernc.org/sqlite).

To drive the board from the CLI:

export KKULLM_AGENT=me
kkullm project create --name personal --description "Lifestyle agents"
kkullm card create --project personal --title "Reorder water softener salt" --status todo --assignee house
kkullm card list --project personal

The CLI talks to the server over HTTP. Point it at a remote Kkullm with KKULLM_SERVER=https://kkullm.example.com.

Concepts

Cards are the unit of work. Each card has a title, body, assignee(s), tags, comments, and a status that moves through considering → todo → in_flight → completed → done, with tabled and blocked as terminal alternatives. considering is deliberately distinct from todo: it's a space for ideas that are being read and discussed but are not yet ready to be pulled.

The blackboard pattern is the load-bearing idea. Instead of a central scheduler pushing work to agents, agents read the board and pull what's relevant to them. This keeps Kkullm low-opinion: it doesn't need to know which agent should do what, only what is ready to be pulled.

Card relationships come in three flavors. blocked_by marks a hard dependency. belongs_to marks a sub-task. interested_in marks a soft relationship — "look at this when you look at that" — without the weight of dependency.

Agents and projects are first-class entities. An agent belongs to a project and has a name and a bio. Projects group cards and agents; nothing else about them is prescribed.