Antigravity Taskboard
github.com/svylabs/antigravity-taskboard- Category
- Developer Tools
- Rank
- No. 1948Tools index
Previous survey · No. 2412 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Agent Building
- Interfaces
- Agent Skill / Plugin
- GitHub
- 1 stars
- Date
About
A Kanban-style plugin for Google Antigravity that gives autonomous coding subagents a Trello-like task board backed by project-scoped SQLite persistence. It orchestrates a continuous supervisor loop with pre-execution verification gates, scope-aware parallel task execution, file locking during review, and a feedback loop for revisions—all without external daemons or SDKs.
What it does
A local work queue for Google's Antigravity agent. You write cards in a browser, each with a goal, a test command or checklist, and optional ordered steps. A bundled skill tells Antigravity to wake every minute, check the queue cheaply, write a plan for the next card, hand the work to a helper agent on its own git branch, run the checks, then park the card for your approval. Agents can drive the queue from a command line tool or a small MCP (Model Context Protocol) server, and the board chimes when a card needs you.
Why it's ranked here
The idea is right: agents should not start work without a stated way to prove it is done, and this makes that the gate. The storage is plain and inspectable, one small database file per project with one declared dependency and a fallback to Node's built-in driver. Against that, much of the scheduling discipline, including parallel limits and review-time file locks, is written as instructions for the agent to follow rather than something the code visibly enforces, and the bundled instruction files contradict each other on whether tasks may run in parallel. A promising MIT licensed tool that is still settling its own rules.
What's good
The idle check is designed to cost almost nothing: it reads unread flags and returns a tiny reply when there is no work, so a one-minute wake schedule does not burn tokens. The loop shuts itself off after eight hours of inactivity by default, and that timeout is configurable. Every card can carry a test command or a written checklist, and a card with neither gets a question posted back instead of code. Attachments up to 10MB live inside the same database, so screenshots and logs travel with the task.
Tradeoffs
The board server listens on all network interfaces by default and answers every origin, and the code read shows no login step, so anyone on the same network may reach the board and the built-in repository file viewer. The docs disagree with each other: one instruction file demands strictly one task at a time, two others allow two in parallel, and the README advertises a revision column that the database code quietly renames on startup. The plugin manifest still points at a placeholder repository address. The manifest declares no test script and no test files appear in the tree.
How to use it well
Best for someone already running Antigravity who wants a visible queue and a review checkpoint instead of one long chat. Give every card a real test command; that is where the tool pays off. Start the board bound to your own machine only, since the default is open to the network. Pick one of the bundled rule files and delete the contradictions before trusting the parallel behaviour. It does not replace a real issue tracker, has no multi-user model, and the scheduling assumes Antigravity's own cron and helper-agent features.
Technical notes+
taskboard/tasks.mjs creates five tables (agent_tasks, task_comments, task_subtasks, task_attachments, board_metadata) at <cwd>/.agents/taskboard/tasks.sqlite, overridable with TASKBOARD_DB_PATH, opens it in WAL mode, loads better-sqlite3 and falls back to node:sqlite, and on every start rewrites status 'needs_revision' to 'planned'. getNextTodoTask orders planned before todo, then by priority. taskboard/server.mjs defaults HOST to 0.0.0.0 and port 4040 (scanning 4041 to 4099 on conflict), sends Access-Control-Allow-Origin: *, caps request bodies at 15MB, and keeps a cross-project registry under the home directory. taskboard/repo_viewer.mjs serves files from the detected repo root and loads the Markdown renderer from a CDN. taskboard/mcp.mjs is a hand-rolled stdio JSON-RPC server exposing nine taskboard_* tools. CLAUDE.md mandates strict sequential execution while rules/AGENTS.md and skills/task-loop/SKILL.md allow a concurrency cap of 2. plugin.json's repository url is github.com/your-org.
Observed
- License
- MIT (LICENSE, package.json, plugin.json)
- Language and runtime
- JavaScript ES modules on Node 18 or later (plugin.json engines)
- Dependencies
- One declared dependency, better-sqlite3, with a node:sqlite fallback in taskboard/tasks.mjs
- Interfaces
- Web board, command line tool (taskctl bin), stdio MCP server in taskboard/mcp.mjs
- Install surface
- git clone into the project's .agents/plugins directory, per README.md
- Network default
- Board server binds 0.0.0.0 with a wildcard CORS header (taskboard/server.mjs)
- Storage
- One SQLite file per project, attachments stored as BLOBs
- Tests
- No test script in package.json and no test files among the repository paths
Read from README.md, package.json, plugin.json, LICENSE, taskboard/server.mjs, taskboard/mcp.mjs, taskboard/tasks.mjs, taskboard/repo_viewer.mjs, skills/task-loop/SKILL.md, rules/AGENTS.md, CLAUDE.md.
What it can do
Track and manage subagent tasks on a Kanban-style board
Subagent tasks → Task board view
Persist project task data in project-scoped SQLite database
Task data → SQLite database records
Run a continuous supervisor loop that verifies tasks before execution
Pending tasks → Verified tasks
Execute multiple tasks in parallel with scope awareness
Task queue → Parallel task execution results
Lock files during review to prevent conflicting edits
Files under review → Locked file state
Provide a feedback loop for requesting task revisions
Review feedback → Revised task
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.