↓ Skip to main content
tomrochette.com
  1. Agents/

Backlog.md

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~639 words)
agent-curated 173 byok 9 fully-ai-generated 263 kanban 3 llm=glm-5.3 170 llm=glm-5.3-flash 178 markdown 1 open-source 62 research-note 148 task-management 8
Table of Contents

Backlog.md is an MIT-licensed, markdown-native task manager and kanban visualizer for any git repository: a zero-config CLI where every task is a plain .md file and agents are steered through three human review checkpoints. Facts below verified as of 2026-09-13.

The tracker’s real product is the review gates, spec, plan, then code, each a screenful a human can actually read, which makes it the first line of code review rather than a to-do list.

What it is
#

Install with npm i -g backlog.md and any folder becomes a project board: tasks live as markdown files with acceptance criteria and a reusable Definition of Done, backlog board paints a live kanban in the terminal, and backlog browser serves a local drag-and-drop web UI. The pitch is aimed squarely at the agent era: agents produce more code than you can read, so you review the task spec and the implementation plan before the code exists, one task equals one context window equals one PR. It works with Claude Code, Codex, Gemini CLI, Kiro, Cursor, and anything MCP-compatible, auto-configuring them on demand, and the project dogfoods hard: nearly all of its own code is written by agents working through its own backlog. It is maintained under the MrLesk identity, with conference talks (Devoxx Belgium 2025) carrying the method.

Status
#

Active and healthy at mid-scale. As of 2026-09-13: 6,719 stars, 54 open issues, pushed 2026-09-12, MIT-licensed, 57,360 npm downloads last month. The July 2025 launch thread reached 254 points, the largest HN footprint of the task trackers profiled in this section.

Strengths
#

  • Everything is markdown in git: diffable, greppable, portable, and readable by humans and agents with no adapter.
  • The three-checkpoint model (spec, plan, code) is the most explicit human-oversight design in this category.
  • Terminal kanban plus local web UI covers both tastes without accounts or hosting.
  • Dogfooding through its own backlog is verifiable in the repo, not a marketing claim.

Cautions
#

  • Markdown files have no atomic claims or dependency graph, so several agents sharing one board will race where beads serializes them.
  • A board of thousands of flat files strains both the kanban view and the human reviewer it exists to protect.
  • The review gates depend on agent discipline to seek approval; a yolo-configured agent can steamroll them.
  • Bus factor looks small for a tool you might standardize a team on.

Pricing
#

Free and open source under MIT. No paid tier, hosting, or account exists.

Compared to
#

  • beads: graph database with atomic claims; choose beads for multi-agent concurrency, Backlog.md for human reviewability.
  • Task Master: PRD-driven pipeline with a company behind it; choose it when you want structure imposed, Backlog.md when you want to impose your own.
  • GitHub Issues: hosted and agent-assignable, but the plan lives outside the repo and outside your diffs.

Bottom line
#

Recommended as the default first task tracker for agent-heavy repos, solo or small team, where reading the plan beats racing the agents. Not for multi-agent shared queues (no claims) or organizations needing a vendor to blame. My disagreeable claim: its three checkpoints are what spec-driven development actually needed, a enforced review cadence rather than more templates, and spec-kit’s constitution files are weaker oversight than one gate a human actually attends.

Changes
#

  • 2026-08-27 - Created in the Task management category seed with the three-checkpoint review model, dogfooding claim, and terminal kanban recorded.

See also
#

  • beads - the graph-database counterpoint from the same era
  • Task Master - the productized pipeline
  • GitHub Spec Kit - the spec-first process layer it complements
  • Vibe Kanban - the kanban pattern applied to agents themselves

References
#