MCP server

SQLAnvil Cloud exposes an MCP server (Model Context Protocol, Streamable HTTP) so AI agents can operate and debug your hosted projects directly: inspect run history, pull a failure’s executed SQL and verbatim warehouse errors, trigger workflow runs, cut releases, and refresh catalogs.

It is deliberately not an authoring surface — you edit and compile your project in your own repo with your own tools. The MCP covers what’s hosted: runs, workflows, releases, and their state.

Connect

1. Create a personal API token at app.sqlanvil.com → API tokens (shown once).
2. Add the server to Claude Code:

claude mcp add --transport http sqlanvil https://app.sqlanvil.com/api/mcp \
  --header "Authorization: Bearer sa_…"

Any MCP client that speaks Streamable HTTP works the same way. Revoking the token in the dashboard cuts access immediately.

Tools

Tool	What it does
list_projects / get_project	Your projects; config, warehouse connection, per-branch catalogs
list_workflows	Each workflow’s selection, schedule, release mode, enabled state
list_runs / get_run	Run history with filters; full per-action detail + log
get_run_failure	The debugging bundle: failed actions with executed SQL + verbatim errors
list_releases	Pinned compilation snapshots (branch, environment, commit, core version)
run_workflow_now	Enqueue an immediate run; returns a run id to poll
cut_release	Compile the branch head into a new pinned snapshot
refresh_catalog	Recompile a branch’s tag/action/environment catalog
set_workflow_enabled	Pause or resume a workflow’s schedule

Runs are asynchronous: operate tools return a run_id immediately; poll get_run for the outcome (runs typically take one to a few minutes).

Data shapes

Graph-adjacent payloads reuse the engine’s own JSON forms — action targets as schema.name, compiled SQL as sqlanvil compile --json emits it — so an agent that already knows the CLI’s output needs no second vocabulary. Everything an MCP tool returns is data the dashboard already shows you, scoped to your account by your token.