An MCP server that answers from the books, and says where the answer came from.
Point Claude, Cursor or any MCP client at Gatlas and ask across every company in the portfolio in plain language. The agent reads documented views rather than raw tables, so it answers in the few rows that matter, and every figure it returns traces back to a row someone can open.
A question, and its basis
What the agent sees
- Curated datasets over the synced tables, each carrying its own context and column notes. The default way an agent gets a number, with filters applied server-side. Raw SQL exists, and is the last resort.
- Curated markdown kept current by scheduled syntheses over the raw files, with the raw files reachable underneath when no document covers the question.
- Which source feeds which company, how fresh it is, and what a stale connection looks like. A connection that may manage syncs can run one.
- Dashboards, documents and decks composed from published views, with a share link. An agent builds one from a closed set of components, so it cannot draw a figure the views do not hold.
What it will not do
- A connection is scoped to the companies you name. To the agent, a company outside that scope does not exist.
- A missing value is returned as missing, never as zero. A page of results says when it is one page of many, so an agent cannot sum a truncation as though it were the whole.
- A tool that changes state says so in its name and its description, and the one that issues a credential-bearing link needs a human’s explicit go-ahead. A Read-only connection cannot reach any of them.
- No credential passes through the agent. The person who owns a login authorizes it through a setup link, and the credential goes into a vault.
- Every call is logged with its tool, its shape and its outcome. Not its content.
The tools
Start here
- get_agent_context
- Operating rules, entity roster and freshness warnings for THIS org. Call it first in every session.
- get_help
- This tool: guidance, the starting libraries of views/documents/artifacts for a company being set up, the artifact component catalog, troubleshooting, the tool map, and the help center for people.
- list_entities
- The companies in scope for your key.
- submit_feedback
- Tell the Gatlas team what is broken, missing or confusing. Please use it.
Views (structured)
- list_views
- The curated view catalog you can read. The single source for what exists.
- search_views
- Ranked search over the view catalog when list_views is too long to scan.
- describe_view
- A view's columns, AI context and caveats. Read before querying it.
- query_view
- Read a view with server-side filters. The default way to get numbers.
- query_sql
- Raw SQL over synced tables. Last resort when no view covers the question.
- list_tables
- The raw tables behind the views, per entity schema.
- describe_table
- A raw table's columns and docs, when you must go below the views.
- create_view
- [WRITE] Persist a new view from SQL + AI context; dry-runs the SQL first.
- edit_view
- [WRITE] Change a view's SQL, AI context or column docs in place.
- delete_view
- [WRITE] Hard-delete a view; unconfirmed calls return an impact report instead.
Documents (unstructured)
- list_documents
- The curated markdown documents you can read.
- get_document
- Read one document (optionally a past version), plus the raw source files that version was generated from.
- search_documents
- Ranked passages from the RAW files under the curated documents. The last resort on this side, like query_sql on the other. It reports how far it searched.
- get_raw_document
- One raw file's text, a page at a time, by the document_id search_documents or get_document's Sources block printed.
- edit_document
- [WRITE] Update a document; the old body is kept as a version.
- list_transform_prompts
- The scheduled AI syntheses that keep documents current.
- create_transform_prompt
- [WRITE] Define a synthesis: prompt + inputs (raw documents, curated documents, published views) + output document + schedule.
- edit_transform_prompt
- [WRITE] Change a transform prompt's definition or schedule.
- run_transform_prompt
- [WRITE] Run a transform prompt now (preview or live).
Pipeline
- list_connections
- The source connections feeding each entity, with freshness.
- get_sync_status
- Sync health per connection (needs a key that can manage syncs).
- get_recent_syncs
- Recent sync runs and their outcomes (needs sync management).
- trigger_sync
- [WRITE] Re-sync a connection now (needs sync management; rate-limited).
- edit_sync_schedule
- [WRITE] Change how often a connection syncs.
- mint_setup_link
- [WRITE] Issue a credential-bearing setup URL for a human operator; requires confirm + purpose.
Activation
- list_artifacts
- Saved dashboards/reports built over published views.
- list_components
- The component library this key can see, with how many artifacts place each one.
- describe_artifact
- An artifact's components, bound views and embed URL.
- describe_component
- One library component's full definition and where it is placed.
- create_artifact
- [WRITE] Compose charts/tables/text over published views into a shareable artifact: a dashboard, an A4 document or a 16:9 deck (layout.format).
- create_component
- [WRITE] Save one chart, table or text block to the component library for artifacts to place by name.
- edit_artifact
- [WRITE] Change an artifact's components, its layout format (dashboard/document/deck) or its share link (the embed field).
- edit_component
- [WRITE] Change a library component everywhere it is placed.
- delete_artifact
- [WRITE] Hard-delete an artifact and its embed link; unconfirmed calls return an impact report instead.
- delete_component
- [WRITE] Remove a library component; refused while any artifact still places it. Unconfirmed calls return an impact report instead.