Altiplano

A small, dependable MCP server for Vikunja. Named after the Andean altiplano, the high plateau that is the Vicuña's native habitat.

Filtering and sorting are passed straight to the Vikunja API (server-side), so there is no client-side filtering engine and no paginate-then-filter pitfall.

Tools

Projects:

• list_projects (includes parent_project_id, shows sub-project nesting)
• create_project (title, parent_project_id?, description?): pass parent_project_id for a sub-project

Tasks:

• list_tasks (project_id, filter, sort_by, page, per_page)
• get_task (task_id)
• create_task (project_id, title, description?, priority?, due_date?, start_date?, end_date?)
• update_task (task_id, title?, description?, done?, priority?, due_date?, start_date?, end_date?): only the fields you pass change. Pass an empty string to any of the three dates to clear it
• set_reminders (task_id, reminders): replaces the task's reminders with the given ISO 8601 datetimes; empty list clears
• delete_task (task_id): soft-deletes the task along with its comments, labels and assignees. Vikunja keeps it for 30 days but exposes no way to restore it, so treat this as irreversible

Relations:

• add_relation (task_id, other_task_id, relation_kind?): relates two tasks, defaulting to a plain related link
• remove_relation (task_id, other_task_id, relation_kind?): the kind must match the one the relation was created with

There is no list_relations: get_task already returns related_tasks, grouped by kind.

Labels:

• list_labels
• add_label (task_id, label_id)
• remove_label (task_id, label_id)

Comments:

• list_comments (task_id)
• add_comment (task_id, comment)
• update_comment (task_id, comment_id, comment): replaces the comment text; get comment_id from list_comments
• delete_comment (task_id, comment_id)

Assignees:

• search_users (query): find a user_id to assign
• list_assignees (task_id)
• add_assignee (task_id, user_id)
• remove_assignee (task_id, user_id)

Credentials (no secrets in mcp.json)

The server resolves two values, in order:

1. Environment variables VIKUNJA_URL and VIKUNJA_API_TOKEN.
2. A per-device file of KEY=VALUE lines, default ~/.config/altiplano/env (override the path with ALTIPLANO_CONFIG).

VIKUNJA_URL is the base API URL including the version prefix (e.g. https://todo.example.com/api/v2).

Recommended so the your mcp.json carries no secrets:

• Drop a per-device file and lock it down:

  mkdir -p ~/.config/altiplano
  printf 'VIKUNJA_URL=https://todo.example.com/api/v2\nVIKUNJA_API_TOKEN=tk_xxx\n' > ~/.config/altiplano/env
  chmod 600 ~/.config/altiplano/env
  

• Or inject via the launcher's environment (e.g. a systemd unit EnvironmentFile= pointing at a chmod 600 file), which the server inherits.
• For stronger setups, source the token from a secret manager/keychain at launch and export it into the environment.

Then mcp.json only needs the command, no env block, no plain-text secrets:

{
  "altiplano": {
    "command": "uvx",
    "args": ["--refresh-package", "altiplano", "altiplano@latest"]
  }
}

uvx caches aggressively. It fetches a version on first run and reuses it afterwards, and it separately caches the list of versions it knows about. So altiplano@latest on its own can keep launching an older build for a while after a release, and restarting your client does not help.

--refresh-package altiplano revalidates against PyPI on each start, which is what makes @latest actually mean latest. It costs roughly a tenth of a second.

If you are already stuck on an older build, quit your MCP client first, then run uv cache clean altiplano. It has to be quit: while the client is running it holds that cache, and the command will sit and wait for it.

Choosing the API version

Vikunja 2.4.0 added a v2 API alongside v1, and this MCP deals with both. The version comes from the URL you configure, so there is nothing else to set:

Prefer /api/v2 if your server has it. Stay on /api/v1 only for older servers; every tool behaves the same either way, though v1 pays for it in requests and cannot exchange Markdown.

On v1, an update costs two requests

Only the fields you pass ever change, on either version. Getting there differs.

v1 has no partial update: POST /tasks/{id} is a replace, so a body carrying only the changed fields resets everything else to its zero value. Left alone, update_task(task_id=42, priority=4) would blank the description, update_task(task_id=42, done=True) would discard the description, priority and dates, and set_reminders would do the same through the same endpoint.

So on v1, update_task and set_reminders read the task first and write it back with your changes merged in. That is one extra request per update, and it is why v2 is worth pointing at.

v2 needs no read. It uses PATCH, omitted fields survive, and only a description forces the longer path, for the Markdown reason below.

One thing v1 cannot do is notice a concurrent edit. The read-then-write on v2 sends the ETag back as If-Match, so a task that changed in between fails loudly. v1 has no ETag to send, so a simultaneous edit from elsewhere is overwritten.

Markdown descriptions and comments (v2 only)

Vikunja stores task descriptions and comments as HTML. On v2 this server asks it to convert, so you write and read Markdown:

create_task(project_id=12, title="Ship it", description="**bold** and a [link](https://example.com)")

Vikunja stores that as <p><strong>bold</strong> and a <a href="...">link</a></p>, and get_task hands it back as the Markdown you wrote. The same applies to update_task, create_project, add_comment and update_comment.

On v1 there is no conversion and the fields are HTML, so Markdown you send is stored verbatim and renders as literal asterisks.

Two things worth mentioning:

• v2 only converts on create and on full replace, never on a partial update. So changing a description reads the task first and writes it back whole, which costs one extra request. The ETag from that read goes back as If-Match, so a task that something else wrote to in between fails with a message telling you to read it again, rather than having that edit quietly overwritten. Updates that do not touch the description stay a single partial update.
• Vikunja resolves @mentions during conversion, so writing @someone in a description notifies them.

Run

uv run altiplano                        # dev, from this directory
uvx --from /your/local/path altiplano   # local path
uvx altiplano@latest                    # from PyPI, refreshing the cache

Notes

• Vikunja priority scale: 0 Unset, 1 Low, 2 Medium, 3 High, 4 Urgent, 5 DO NOW.
• Dates are ISO 8601 datetimes. start_date/end_date mark the window you plan to work on a task (start work / finish work); due_date is the deadline.
• To clear a date, pass an empty string. Vikunja has no null for one: an unset date is the zero time, 0001-01-01T00:00:00Z, and that is what gets written.
• When a call is rejected, the error carries Vikunja's own explanation, plus its numeric error code on v2, instead of only the HTTP status.
• Relation kinds: subtask, parenttask, related, duplicateof, duplicates, blocking, blocked, precedes, follows, copiedfrom, copiedto. Direction matters for the asymmetric ones: add_relation(task_id, other_task_id, "subtask") makes the other task a child of task_id.
• The UI shows tasks by their project-local identifier (e.g. #50), which is not the global id the API uses.
• Verified end to end against Vikunja v2.5.0 on both /api/v1 and /api/v2.
• list_assignees needs a server where GET /tasks/{id}/assignees works. It answers 500 on v2.3.0, which was a server-side bug, and works on v2.5.0. Every other tool works on both.

Contributing

Enable the pre-commit hook once per clone:

git config core.hooksPath hooks

Git hook configuration lives in .git/config, which is not version controlled, so I can't do it for you. The hook itself is committed in hooks/, and git will not look there until you point it.

The same checks run on every pull request. The suite must pass and coverage must stay above 90 percent.

Coverage is also reported to Codecov on each pull request, as a comment and as inline annotations. That reporting is purely informational and never fails a build.

Licence

MIT.

Support

RTFM, then RTFC... If you are still stuck or just need an additional feature, file an issue.