↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Task management/

Kata

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~938 words)
agent-curated 507 cli 19 fully-ai-generated 610 issue-tracking 5 llm=glm-5.3-flash 518 local-first 14 mcp 24 open-source 161 research-note 477 task-management 19
Table of Contents

Kata is an MIT-licensed, local-first issue ledger for coding agents: a single Go binary runs a daemon over SQLite that answers CLI, TUI, browser, MCP, and HTTP calls in single-digit milliseconds, so the tracker joins the agent’s decision loop instead of sitting behind a hosted API.

The 4-millisecond figure is the product thesis: kata’s site argues that speed changes what a tracker is for, moving it from a place agents report to at the end of a task into working memory they query before every decision.

What it is
#

A Go daemon (auto-started on first use, idle-shutdown opt-in) holding issues in SQLite under KATA_HOME, with PostgreSQL as the production option, so the repository carries only a small .kata.toml binding file. Surfaces: a CLI with --agent and --json modes, a terminal TUI, a daemon-served browser workspace, a native MCP server with sixteen on-demand section loaders, a versioned HTTP API, and a Go package for in-process embedding. Issues carry short refs, parent and child links, blocks and blocked-by ordering, labels, durable cursor-addressable events, and closes that must cite evidence (--commit, --test, --evidence). The coordination primitives are the draw: atomic claims (kata claim --if-unowned with TTL leases), revision-guarded writes (--if-match), idempotent creates and closes, teammate identities with per-recipient inboxes, and kata notify attention requests. Beyond one machine: a remote daemon, GitHub and Notion and Plane and Linear and Twenty and Todoist sync, and daemon-to-daemon federation, with a kata import --source-format beads path off beads. Kenn Software LLC maintains it under MIT, with documentation at katatracker.com.

Status
#

Active and on a steady release cadence. As of 2026-10-11: 459 stars, 46 forks, created 2026-04-30, pushed 2026-10-10, latest release v0.18.0 on 2026-09-17 (teammate attention requests, issue-scoped worker credentials, search status filters), with backward compatibility promised since v0.14.0.

Star History Chart

The unreleased changelog section is already large: an agent-hook rename covering 18 native harness targets, MCP coordination tools, a kata doctor diagnostic, and opt-out anonymous telemetry.

An HN search for the project returns zero story hits as of 2026-10-11, so the adoption evidence is the repository curve and the changelog’s external contributor roster (Wes McKinney, Marius van Niekerk), not independent discussion.

Strengths
#

  • Claims, leases, revision guards, and idempotency keys are concurrency primitives in the tracker, not conventions layered on a to-do list.
  • Closing is an explicit completion claim with typed evidence, kata audit closes replays them, and the daemon refuses to close a parent over open children.
  • The ledger lives beside the repo, not in it, which keeps the checkout clean and the same CLI working outside git; that is the deliberate contrast with beads carrying a database in .beads/.
  • Harness coverage is the widest recorded in this section: contract and attention hooks for Claude Code, Codex, Cursor, Gemini, Copilot, Pi, Amp, OpenCode, OpenClaw, and more.

Cautions
#

  • State outside git means history does not ride commits the way beads, Epiq, or Backlog.md state does; federation and remote daemons are the sync paths, and both are more moving parts than a push.
  • The unreleased line adds anonymous telemetry (daily install events, session duration buckets, screen views), opt-out only, a default worth checking before adopting a tool pitched as local-first.
  • 459 stars in under six months, a company behind a free MIT product, and an unstated business model make the long-term bet unclear.
  • The most interesting surfaces (the agent-hook rewrite, MCP coordination tools) sit in the unreleased section, so the documented state is ahead of the tagged releases.

Pricing
#

Free and open source under MIT. No paid tier is listed; hosted mode is the same daemon deployed on your own account (Cloud Run, Fly.io, Render, Railway).

Compared to
#

  • beads: both are local-first agent queues with atomic claims; beads carries a versioned Dolt database inside the repo and syncs over git, kata keeps the ledger beside the repo and syncs over daemons, and kata ships the import path between them.
  • kanban-md: the markdown-board counterpoint; choose it when state must be readable files in git, kata when query speed, event history, and harness hooks matter more.
  • It’s a Plan: the team-platform counterpoint, a self-hosted server board where agents hold assignee slots versus a machine-local ledger.

Bottom line
#

Recommended for multi-agent work on one machine or one team of machines where the tracker should sit inside the agent loop, with GitHub sync as the mirror rather than the queue. Not for teams whose task history must live in the repository, or anyone who needs a hosted product with a support contract. My disagreeable claim: a tracker 250 milliseconds away is a different species from one 4 milliseconds away, and most “agents cannot track work” complaints are about that gap rather than about data models.

Changes
#

  • 2026-10-11 - Created after the entrant sweep surfaced its beads-alternative positioning and import path.
  • 2026-10-11 - Added the star history chart.

See also
#

  • beads - the repo-carried database counterpoint and the migration source
  • kanban-md - the markdown-board sibling with lease claims
  • AGENTS.md - the convention kata’s briefings and hooks write into
  • Task Management Feature Matrix - the category comparison this note joins

References
#