Task, project, note, report, and crew workflows for MCP-capable coding agents.
The shared skills work across harnesses. This repository packages them for Claude Code and Codex, and provides configuration for Cursor, VS Code, and other Streamable HTTP MCP clients. Harness-specific features are documented separately; an MCP connection alone does not install commands, skills, or agents.
Canonical repository: AgentTaskLLC/AgentTask-plugin. Personal forks and mirrors are development copies. The plugin is independent of the AgentTask product source and connects to the hosted service.
Claude Code:
/plugin marketplace add AgentTaskLLC/AgentTask-plugin
/plugin install agent-task@agent-taskFor this local checkout, replace the marketplace source with its absolute path.
Open /mcp to authenticate. Use namespaced commands such as
/agent-task:start AI-123 and /agent-task:report.
Codex, on versions with the plugin CLI:
codex plugin marketplace add /absolute/path/to/agent-task-plugin
codex plugin add agent-task@agent-taskStart a new session after installation. The local path is intentional while these 0.4.0 changes await release. Once published, the canonical GitHub source can be used in the marketplace-add command. Do not install a second marketplace with the same name without first inspecting your existing sources.
For MCP-only setup, authentication, endpoint overrides, and other harnesses, use the connection guide and compatibility guide.
Related MCP server: task-orchestrator
Shared workflow skills, including natural-language access to command procedures.
Claude Code commands for start, plan, update, triage, standup, report, organize, note, init, finish, and crews.
A Claude Code reporter with an explicit read-only tool allowlist.
A local, read-only session hook that surfaces a branch ticket hint in Claude Code and compatible Codex versions after hook trust review.
Optional Git pre-push reminders and an optional Claude crew status line.
A configuration generator that emits credential references, never secret values.
The shipped connection is header-free OAuth at
https://app.agent-task.com/v1/public/mcp. API-key access and alternate endpoints
use user-owned client configuration. Tool availability comes from live discovery,
not a fixed tool count. See the plugin reference.
Python 3.11+ is used for repository checks and configuration generation. The installed plugin has no Python or Node runtime requirement. Optional shell helpers need Bash and Git 2.31+; session/status helpers also need jq.
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-dev.txt
.venv/bin/python scripts/validate.py
.venv/bin/python -m unittest discover -s tests -v
shellcheck plugins/agent-task/git-hooks/* plugins/agent-task/scripts/*.sh plugins/agent-task/statusline/*.shContributing covers verification and release preparation. Security explains trust boundaries and responsible reporting. Changelog records compatibility changes.
Proprietary. See LICENSE. Copyright (c) 2025 AgentTask LLC.
This server cannot be deployed
Task & board management for AI agents + humans. Kanban, comments, digests via MCP.
Work management where AI agents are first-class members: tasks, projects, memory over hosted MCP
Agent-first task marketplace MCP — discover, claim, and deliver paid workspace tasks.
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
