Tasks as Git-native state

Traditional project management tools are designed for humans. They exist in external databases, behind proprietary APIs, and require authentication layers that create friction for AI agents. Backlog.md flips this model by treating the file system as the source of truth. It is a task manager where todos, boards, and decision logs are stored as plain Markdown files directly inside a Git repository. This approach ensures that the project's state is always within the local context of a coding agent, bypassing the latency and complexity of external SaaS platforms.

Developed as an open-source project, Backlog.md bridges the gap between human intent and agent execution. When a developer initializes the tool, it generates specific instruction files—such as CLAUDE.md or AGENTS.md—that inform AI assistants how to interact with the project's task list. This effectively gives the agent a set of "operating instructions" for its own workflow, allowing it to mark tasks as complete, update progress, or create new sub-tasks as it writes code.

The MCP advantage

The project is a first-class citizen in the burgeoning Model Context Protocol (MCP) ecosystem. By implementing an MCP server, Backlog.md allows tools like Claude Code, Gemini CLI, and other MCP-compatible assistants to discover and manipulate tasks through a standardized interface. Instead of the agent having to guess how to format a task or where to find the backlog, the MCP server provides a structured resource at backlog://docs/task-workflow. This makes the task management process programmable and predictable for LLMs, which struggle with the ambiguous UI elements of traditional PM web apps.

While Nulab's Backlog (a legacy SaaS product) serves general business collaboration, Backlog.md is focused strictly on the developer and agent relationship. It avoids the per-user pricing and heavy overhead of enterprise tools in favor of a zero-configuration, Go-based binary that runs locally. The use of Git for storage means that every task change is version-controlled and tied to the code changes it describes, providing a historical record that is useful for both human auditing and agent training.

Distributed and local-first

Because Backlog.md stores everything in Markdown, it is compatible with any text editor or terminal. A human can edit a task in VS Code while an agent updates a status via the CLI. The tool enforces consistency in metadata and field types, ensuring that human edits do not break the agent's ability to parse the file. This dual-access model is the project's core differentiator; it doesn't force the human to use an agent-only tool, nor does it force the agent to navigate a human-only web interface. It sits in the shared space of the repository, making project management a byproduct of the development process rather than a separate administrative chore.