vibedasher-mcp

An MCP server that exposes the Vibedasher data engine — datasets, CSV upload + ETL, headless SQL query, and dashboard (viz) management — as tools a customer's AI (Claude Code, Cursor, ...) can drive directly.

This is the control plane (VD-601, PIVOT/PLAN.md EP-6). It wraps the published vibedasher Python SDK and calls only the public, metered /api/v1/* API.

Install: pip install 'vibedasher-mcp>=0.6.0', then vibedasher login — no API key to mint. Live on PyPI since 2026-08-09, as are vibedasher (PyPI) and @vibedasher/client (npm).

The eject tools (VD-602) ship in this package, not separately: eject_viz, eject_viz_files, eject_instructions and get_viz_files are registered by register_eject_tools() in eject.py. They pull a viz's code plus an SDK-wiring manifest so your AI can recreate the dashboard natively in your stack.

Tools

Delivery: an ejected tree does not fit in one response

Measured against prod on 2026-08-18: eject_viz(1275) is 44 files / 517,351 characters (~129k tokens) and eject_viz(1282) is 45 files / 556,363 characters (~139k tokens). MCP clients cap a tool result, so the single-shot payload was either dropped outright by a strict client or ate the agent's entire context. Every eject response now carries a delivery block, and the three file tools take the same three parameters.

eject_viz(viz_id=1275, out_dir="/abs/path/to/app/dashboard")   # one call, whole tree on disk

out_dir resolves on the machine running the MCP server; with the standard stdio launch (command: vibedasher-mcp) that is your machine, because the client spawns the server as a local subprocess. Relative paths are refused, existing files are refused unless overwrite=True, and any path escaping out_dir aborts the write before a byte lands.

Paging is the fallback, not the goal: round trips do not reduce total tokens (the content still crosses the model), they only keep each individual response deliverable. Hand delivery.pendingPaths straight back to eject_viz_files(viz_id, paths=...) and repeat until it is empty. A response with a non-empty pendingPaths is a page, not a tree.

Override the budget per call with max_chars or globally with VIBEDASHER_MCP_MAX_RESPONSE_CHARS.

Three data lanes

Pick by what you will do with the rows:

run_query_to_file writes on the machine running the MCP server, which with the standard stdio launch is the agent's own machine (the client spawns the server as a local subprocess). It is refused over --http, where the path would resolve on the server's disk. Both lanes that return counts honour the backend's 100,000-row cap: truncated: true means the rows — in the response or in the file — are a prefix. Parquet output is deliberately not offered: the wasm lane already returns Parquet the server built, and pyarrow is a 40 MB wheel to install for one option.

run_query reuses the SDK's hand-written query() transport (VD-301): the caller never sees inline-vs-presigned delivery, MessagePack, or retries — one call in, typed rows out. Each dataset id resolves server-side to that dataset's cleanSQLName alias under the VD-203 RLS/alias-rewrite; SQL references only aliases.

Auth

Two lanes. The first one is the default because it asks nothing of you.

Sign in through your browser (first run)

pip install 'vibedasher-mcp>=0.6.0'   # `login` does not exist below 0.6.0
vibedasher login                      # opens a browser; approve; done

vibedasher login opens your browser, captures the authorization code on a loopback listener bound to 127.0.0.1, and stores a short-lived refresh token in your OS keyring (with a 0600 file fallback where no keyring is available). There is no key to mint, copy, paste or rotate, and nothing lands in a config file or your shell history. Later runs are silent until the session lapses, at which point login asks again. The refresh token is single-use and rotates on every refresh; if a rotated-away copy is ever presented again — someone read the keyring or the credentials file — the whole session is revoked on the spot and the next run asks you to login again.

Two companions: vibedasher whoami prints the account, user, region and the date you will be asked to sign in again; vibedasher logout revokes this machine's session and forgets it.

The MCP client block then carries no credential at all:

{
  "mcpServers": {
    "vibedasher": {
      "command": "vibedasher-mcp",
      "env": { "VIBEDASHER_REGION": "eu-central-1" }
    }
  }
}

VIBEDASHER_REGION is optional here — login records the region the account actually lives in — but pinning it costs nothing and makes the shard explicit.

API key (CI, servers, headless agents)

Anywhere a browser cannot open, an API key is still the answer. Mint one in the console (Settings → API keys) or via POST /v1/api-keys, and give it to the server as an environment variable:

export VIBEDASHER_API_KEY=...           # X-Api-Key, handled by the SDK's create_client
export VIBEDASHER_REGION=eu-central-1   # or us-east-1
# export VIBEDASHER_BASE_URL=...        # optional override (on-prem/staging)

{
  "mcpServers": {
    "vibedasher": {
      "command": "vibedasher-mcp",
      "env": {
        "VIBEDASHER_API_KEY": "...",
        "VIBEDASHER_REGION": "eu-central-1"
      }
    }
  }
}

An API key authenticates as the user who created it, has no expiry unless one is set, and is a bearer secret in plaintext on whatever disk holds that config. That is a fine trade for a build server and a poor one for your laptop — which is why it is the second lane and not the first.

Either way, the server checks before it serves

vibedasher-mcp resolves a credential — stored session first, then VIBEDASHER_API_KEY — at startup. With neither, it refuses to start and prints what to do on stderr, so a credential-less server never connects and never advertises tools it cannot serve. Neither credential is ever logged or echoed in a tool result.

Metering

Every tool hits the public, metered endpoints (create_dataset is credit-gated; run_query is metered via the usage/credits event). The MCP adds no side channel and bypasses no metering.

Run

pip install 'vibedasher-mcp>=0.6.0'   # (monorepo dev: also make `vibedasher` importable)
vibedasher login                       # sign in once, in a browser
vibedasher-mcp                         # stdio MCP server
# or:  python -m vibedasher_mcp

Wiring Claude Code in one line, on a machine that has signed in:

claude mcp add vibedasher --env VIBEDASHER_REGION=eu-central-1 -- vibedasher-mcp

Then ask your agent for something. It has 82 tools; a good first prompt is "run whoami, then list_datasets, and build me a dashboard from the most interesting one — verify with get_build_status and probe_viz before telling me it's done."

Two verifiers, two worlds, and they do not overlap:

render_probe walks the WHOLE dashboard, not one screenful: it scrolls the frame to the bottom in steps (re-settling after each one, so a lazily-mounted tile is counted and a permanently skeletal one still goes red) and, when the tree exposes the generated page control [data-selector="page"] [data-value], opens every page and sweeps each. Tiles are unioned by node identity, so a tile carrying three conventions is one tile and two conventions marking different tiles add up. tiles.covers states in one sentence what the count is a count OF — all pages fully scrolled, one page fully scrolled, or one viewport — and whatever it could not walk stays in notVerified. Pass scroll=False / traverse_pages=False to narrow it; the caveat comes back when you do.

render_probe needs a browser, so Playwright is an extra:

pip install 'vibedasher-mcp[render]>=0.6.0' && python -m playwright install chromium

Without it the tool returns checked: false and that install command. It never reports a pass it did not earn.

For the JSON mcpServers form of both lanes, see Auth.

Development

The package depends on the sibling SDK at ../sdk/py. Tests wire that path automatically (tests/conftest.py), so from packages/mcp:

python -m pytest tests/ -q

Note: fastmcp is a client-side dependency of this standalone package; it is not bundled into the API Lambdas, so the export_requirements.py step in the root CLAUDE.md does not apply here.