plandb

PlanDB is a local-first, single binary, SQLite-backed issue tracker aimed at AI agents, giving them a compound task graph, atomic claiming, and context that resurfaces on its own.

PlanDB is described as the issue tracker AI agents are missing, a Linear or Jira for machines rather than humans. The argument is that agents decompose tasks mid-flight, parallelize across branches, and pivot entire subtrees when an approach fails, so they need infrastructure that keeps up. The answer is local-first: a single binary, SQLite backed, with no cloud, no accounts, and no setup.

Installation is a single command. The installer downloads the binary and auto-configures your agents with full planning instructions, supporting Claude Code, Cursor, Codex, Gemini CLI, OpenCode, Windsurf, and Aider. Re-running it is safe because it is idempotent. For Claude Code specifically it installs a rules file plus a skill that adds a structured 6-phase workflow with command gotchas and anti-patterns, so agents plan automatically.

The design is a compound graph rather than a flat task list. Tasks contain subtasks to any depth, like folders, and dependencies connect tasks across those boundaries, like symlinks. A backend subtask can depend directly on a schema task in a different part of the tree, one split can turn a stuck task into three parallel subtasks, and a --dep flag wires them into the right execution order. Plans adapt mid-flight: splitting a task, inserting a missed step, or pivoting a subtree rewires dependencies automatically.

Atomic claiming means two agents cannot grab the same task, with no locks, no races, and no duplicate work. Critical path analysis shows the longest dependency chain, and a bottlenecks view shows what blocks the most downstream work. Context recorded with plandb context surfaces automatically through BM25 when a related task is claimed later, so knowledge compounds across agents and sessions without anyone searching for it.

The README shares a test run: one Claude Code instance was given a single sentence, build a GPT from scratch in Rust and train it to do tool calling, and PlanDB handled the planning with no human intervention. The agent built a 3,769-line transformer in pure Rust with zero ML frameworks, ran 7 RL experiments, and the plan grew from 6 to 20 tasks. Three interfaces are documented: a CLI for shell agents, an MCP server for Claude Code, Cursor, and Windsurf, and an HTTP API via plandb serve for custom agents and dashboards. PlanDB is the task planning layer of the AgentField project, and SWE-AF uses it internally to coordinate parallel workstreams.

The pitch is direct: agents have no issue tracker, and PlanDB installs with one command and no configuration. The case study of a single sentence growing into a 20 task plan is the README's best evidence that the approach holds up in practice.

• Source 1 · Project repository
• Source 2 · README