tasktrooper

Website: tasktrooper.ai · Docs: tasktrooper.ai/docs · Download: Releases

A local-first agent platform for software teams of one. A board of tasks, a set of role agents (product manager, architect, backend, frontend, QA), and a runtime that hands each task to an agent CLI on your own machine — Claude Code, Cursor, Antigravity or OpenCode — or to a model API you bring a key for: clone, plan, implement, test, open the PR, while you watch the run.

Everything runs on your machine: the desktop app starts an embedded Postgres and the Go backend, serves the UI, and runs the agent sessions locally. No account, no cloud, no login.

You don't drive the agents, the board does. Put a task on the board and the agent that owns its column picks it up on its own, does the work and hands the task on to the next column, where the next agent takes over. Nobody has to press run.

A usage limit doesn't lose work. When an agent CLI runs out of its usage limit in the middle of a task, the agent stops and the task waits on Blocked instead of failing. Once the limit resets, TaskTrooper picks the task back up by itself; on a CLI that can resume a session, such as Claude Code, the agent carries on from where it stopped instead of starting over.

Features

Board

• A Kanban board with thirteen columns out of the box, from Backlog to Released, including analysis review, code review, QA, PM UAT and human UAT. The columns, and which agents pick up work in each, are configurable.
• Agents take tasks by themselves. A task that lands in a column is dispatched to that column's agent automatically and moves on when the agent is done.
• Tasks carry acceptance criteria. A task cannot move forward out of a review column until every criterion has a verdict.
• Tasks can block each other; a blocked task waits until its blocker is done.
• Every task gets its own branch and pull request. Code review reads the PR; once it is signed off, a dedicated release engineer merges it, ships it per the component's delivery profile, verifies production afterward, and finishes or rolls back the release — never on a green deploy alone.
• When an agent CLI hits its usage limit, the task is parked on Blocked until the limit resets, and other runs on the same CLI are held instead of hitting the same wall. Then the task continues by itself. Claude Code resumes the parked session (--resume), so the agent keeps what it already read and wrote.

Role agents

Seven role agents ship with a catalog-driven library of skills, rules and tool policies — the repo's catalog/ directory is the source of truth and the backend syncs it into the database at boot and on an interval. They run on the connected agent CLI (Claude Code, Cursor, Antigravity or OpenCode) or on a configured API provider, each taking that runtime's default model unless an agent is given its own.

Every agent is editable: provider and model, tool policy, effort, skills, rules, the columns it works, and its memory. You can chat with any agent directly, or about a specific task.

• Your own agents. Create an agent from a template or from scratch with its own prompt, skills, rules, tool policy, columns and memory, and save any agent as a template for the next one.
• One board, several CLIs. The runtime is chosen per agent, not per install. A developer on Claude Code, a reviewer on the Cursor CLI, a QA agent on OpenCode and an API-only agent on Gemini can all work the same board, each in its own session.

Self-evolution

Agents rewrite their own playbooks from how their work actually went.

• Reflection. On a schedule, whenever a task is sent back to Need Revision, or on demand, an agent reviews everything since its last reflection: its runs, chat messages, revision comments, scores, KPI results and its current skills, rules and memories. It proposes changes to all three. Skill and rule changes are applied only for agents with self-evolution turned on.
• Golden gate. With the gate enabled, a golden task suite runs before and after a proposed change, and an independent judge model decides whether to keep it. If the pass rate drops, the whole change set is reverted automatically.
• Impact tracking. Each applied change is later classified as effective, regressed or neutral by comparing scores before and after it. Regressions are put in front of the agent's next reflection, which decides whether to revert.
• Budgets and history. Skills and rules are capped per agent (25 and 15 by default), so an agent merges and updates instead of piling up. Every write to a skill or rule is versioned with its source (you, self-evolution or the upstream catalog) and any version can be restored.
• KPIs. Each agent has targets such as tasks completed, first-pass rate, revisions received, UAT failures, failed runs and time spent per column, measured per day, week or month. The targets are part of the agent's prompt, and the Performance page shows how it is doing.

Memory and code understanding

• Agents save and search memories in four scopes: personal or team-wide, for one repository or for every repository. Notes about a single run are refused, and a near-duplicate is not saved twice.
• Repositories are parsed with tree-sitter and embedded on your machine with the bundled nomic-embed-text-v1.5 model, so agents search code semantically and see uncommitted edits without a re-index.
• Skills are loaded on demand, so a long skill list does not fill the prompt.
• Files uploaded to a workspace are available to agents through retrieval.

Runtimes, models and tools

• Agent CLIs run as local, headless processes: Claude Code, Cursor, Antigravity and OpenCode. Every session gets TaskTrooper's board tools over MCP with a per-run token that is revoked when the run ends.
• API providers for agents that do not use a CLI: OpenAI, Anthropic, Google Gemini, Groq, or any OpenAI-compatible endpoint such as LM Studio, Ollama or vLLM.
• Connect your own MCP servers.
• Built-in tools: terminal, file editing, web search with no API key, page fetch, headless-browser QA, and a boilerplate catalog to start new projects from.
• A repository's own version pins (.tool-versions, go.mod, .nvmrc and others) are honoured in the agent session.

Integrations and operations

• GitHub: clone, branches, pull requests, review comments, merge, and CI status from GitHub Actions.
• Deploys: a recipe catalog for Google Cloud Run and GKE, AWS ECS and Lambda, Vercel and Fly, rendered into a workflow per environment with a health check, plus a deployment matrix. A Vercel account can be connected to bind existing projects.
• Production incidents: alerts from Alertmanager, Sentry, Cloud Monitoring or any JSON webhook, together with a health monitor, fold into deduplicated incidents with a suggested remedy: a rollback, a config, dependency or capacity fix, or a code defect. Per repository you choose whether an incident is only recorded, becomes a diagnosis task, or is fixed through the board.
• Mobile releases: connect App Store Connect and Google Play and promote builds through internal, external and production channels. QA can drive iOS simulators (macOS only) and Android emulators on this machine through Appium.
• Usage: token usage per model and per day.

First run

A guided setup checks this machine (git, the agent CLIs you have, and optionally Chrome, Xcode, Appium and the Android SDK), lets you connect any agent CLI you have (Claude Code, Cursor, Antigravity, OpenCode) or an API provider with your own key, then connects GitHub and imports your first project.

Install

cu