vikunja-mcp
A workflow-level MCP server for Vikunja with a fixed task pipeline (Backlog → Queue → Design → Build → Review → Done) that enables agents to claim, advance, and comment on tasks, while handling human callbacks and blocked items.
README
vikunja-mcp
A workflow-level MCP server for Vikunja — not a generic task CRUD wrapper. It exposes a small set of tools that push an agent through a fixed pipeline, and the gates (what's allowed from which stage, what evidence is required) live in the tools themselves, not in prompts:
Backlog → Queue → Design → Build → Review → Done
↕ ↕
Call to Human
BacklogandDoneare human territory (triage in, sign-off out) — agents never move a task intoDonethemselves.Queue → Design → Build → Reviewis the agent loop: claim, spec, build, hand off for review.Call to Humanis a side branch reachable fromDesign/Buildwhen an agent needs a decision or input; the human answers and moves it back.
Install
uvx --from git+https://github.com/ufna/vikunja-mcp@main vikunja-mcp
Register it with Claude Code via .mcp.json:
{
"mcpServers": {
"tracker": {
"command": "uvx",
"args": ["--from", "git+https://github.com/ufna/vikunja-mcp@main", "vikunja-mcp"]
}
}
}
Configuration
Config is resolved from three layers, in priority order:
- Environment variables —
VIKUNJA_URL,VIKUNJA_TOKEN,VIKUNJA_PROJECT_ID - Repo file
.vikunja-mcp.toml(found by walking up from the cwd) —urlandproject_id, safe to commit (no secret) - User env file
~/.config/vikunja-mcp/env(KEY=VALUElines,chmod 600) — the usual place forVIKUNJA_TOKEN
# .vikunja-mcp.toml (commit this)
[tracker]
url = "https://vikunja.example.com"
project_id = 12
project = "My Project" # informational label; not used for lookup
# ~/.config/vikunja-mcp/env (chmod 600, keep out of git)
VIKUNJA_URL=https://vikunja.example.com
VIKUNJA_TOKEN=tk_xxxxxxxxxxxx
VIKUNJA_PROJECT_ID=12
Note: the token is never read from the repo toml — only from
VIKUNJA_TOKEN (env) or the user env file — so .vikunja-mcp.toml has
nothing secret in it and is safe to commit.
Tools
| Tool | Gate / behavior |
|---|---|
next_task() |
Returns your active task in Design/Build first (incl. one bounced back from Call to Human), else the top-priority free task in Queue. Never returns Backlog or blocked-labeled tasks. One task at a time. |
claim(task_id) |
Queue → Design only. Assign-then-verify: assigns you, re-reads the task, and backs off if someone else was assigned in the same window (race lost — call next_task again). |
get_task(task_id) |
Dossier: description, stage, assignees, labels, full comment thread. |
comment(task_id, text) |
Adds a progress note to the task's comment log. |
advance(task_id, to, spec=, worklog=, evidence=) |
to="build": Design → Build, requires spec. to="review": Build → Review, requires worklog + evidence. to="done" is always rejected — Done is human-only. Task must be assigned to you. |
call_human(task_id, question) |
Design/Build → Call to Human. Keeps your assignment (not a review step, not an external block); posts question as a comment. |
return_task(task_id, reason) |
For external blockers (no access, missing dependency, someone else's service down). Unassigns you, adds a blocked label, moves the task to Backlog for human re-triage. |
decompose(task_id, subtasks) |
Requires ≥2 subtasks (each needs a title). Creates each as a new task with a parenttask relation to the parent and drops it in Queue. Parent is unassigned, labeled epic, and moved to Backlog. |
Project setup
VIKUNJA_TOKEN=<admin token> vikunja-mcp setup --project NAME [--share user:read|write|admin ...] --url URL
Creates the project if it doesn't exist (matched by title), then
creates/reconciles the seven canonical buckets in order, migrates known
default-Vikunja buckets (Todo/To-Do/To-do → Queue, Doing → Build,
moving their tasks), removes empty non-canonical buckets (leaves non-empty
unknown ones alone with a warning), sets Backlog as the default bucket and
Done as the done bucket, applies any --share grants, and prints ready-to-
commit .vikunja-mcp.toml + .mcp.json snippets.
vikunja-mcp install-skill
Copies the packaged tracker skill (queue discipline, comment-trail
expectations, call_human vs return_task) to ~/.claude/skills/tracker/SKILL.md.
Development
uv sync
uv run ruff check .
uv run pytest tests/unit -q
Integration tests exercise a real Vikunja instance and are skipped unless
VIKUNJA_TEST_URL is set:
docker run -d --name vikunja -p 3456:3456 \
-e VIKUNJA_DATABASE_TYPE=sqlite -e VIKUNJA_DATABASE_PATH=/tmp/vikunja.db \
-e VIKUNJA_FILES_BASEPATH=/tmp/files \
-e VIKUNJA_SERVICE_JWTSECRET=ci-secret \
-e VIKUNJA_SERVICE_PUBLICURL=http://localhost:3456/ \
-e VIKUNJA_SERVICE_ENABLEREGISTRATION=true \
vikunja/vikunja:2.3.0
timeout 60 bash -c 'until curl -sf http://localhost:3456/api/v1/info; do sleep 1; done'
VIKUNJA_TEST_URL=http://localhost:3456 uv run pytest tests/integration -q
Vikunja rate-limits /login (10 requests/60s shared with /register); the
integration conftest retries on HTTP 429 with backoff (up to ~150s worst
case across a full run) — expected, not a bug.
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。