cairn

cairn is continuity for agents: a durable task graph your agent plans in and resumes into. The plan lives outside the conversation, so when the context resets the agent reloads it from the graph instead of reconstructing it from a transcript.

One command serves cairn to anything that speaks the Model Context Protocol:

kli mcp-serve cairn

Source and releases are on GitHub.

Learn

Start with the Quickstart. It installs cairn, whether baked into kli or served to any MCP client, then walks the task_create, observe, handoff, and task_bootstrap resume loop and a tiny plan you fork into phases and resume.

Concepts explain how the graph, the current-task pointer, and the capability boundary work. Read what it is, the task graph, the current-task pointer, the event log, the capability boundary, and the working method.

Reference

Tools & Query Language is the exact catalogue: the fourteen MCP tools, the TQ query language, and the built-in views. The closed enums pin the edge types, statuses, and queryable fields.

Commands & Serving documents the six slash commands you get under a kli host and how to serve cairn over MCP to any client.

These docs are written to be read by a machine mid-task as readily as by a person at a desk, and the two read different parts of them. A person evaluating or operating cairn wants the Quickstart and Concepts; the agent executing against it wants the Reference — the exact tools, query language, and enums — which the site also serves as Markdown and an llms.txt index so a machine can pull it in-band. Either way, every page leads with what a thing is and quotes tool names and return strings verbatim, since an agent cannot paraphrase a wrong string into a working call. These docs were themselves produced by an agent dogfooding the cairn-method, the same bootstrap, observe, and handoff loop that keeps any plan across a reset.