Room MCP - Corvic Platform

Room MCP gives you programmatic access to everything a room holds. One MCP endpoint exposes the whole platform surface: enumerate and run pipelines, ask Virtual Memories questions, read documents through Document Lenses, delegate open-ended work to the room’s agents, and pull results back out as rows or files. Unlike Agent MCP — which talks to a single deployed agent — Room MCP is the API for the room itself.

Get connected from the MCP Playground

Every room has an MCP Playground (open it from the room’s surface switcher, next to Chat, Live Apps, and Memory). It gives you everything you need to connect:

The MCP endpoint — https://api.corvic.ai/mcp/
An access token with a chosen expiry, sent as Authorization: Bearer <token>
The room’s ID, which most tools take as room_id

The Playground also lets you call the room’s tools live — the exact requests and raw content an MCP-aware client (Claude, Cursor, or your own) would send and receive. It is the fastest way to explore a tool’s arguments and see real responses before wiring anything up.

Treat access tokens like passwords. Don’t commit them to source control — load them from environment variables or a secrets manager, and pick the shortest expiry that works for your use.

Any MCP-compatible client can connect: point it at the endpoint with your Bearer token, list the tools, and call them.

Conventions

A few rules hold across the whole catalog:

Pagination — any list_* tool may paginate. Omit cursor on the first call; the response carries the next one, and an empty cursor means done. Omit entries_per_page (or pass 0) for the server default.
IDs and refs — every ID is a string. Feature view refs look like 75:code_augment; table IDs look like 117:file. Refs hold a colon, so SQL must quote them.
Async runs — run_pipeline returns immediately with the feature views to watch and no run ID. Poll get_pipeline, whose statuses are not_run, processing, completed, failed, stopped (per feature view and rolled up).
Metering — query, ask_memory, run_pipeline, and the extraction calls are metered. Reads — listings, get_*, fetch_*, get_extraction — are not.

Rooms

Tool	Description
`list_rooms`	List rooms (paginated).
`get_room`	A room’s name, creation time, and visibility.
`create_room`	Create a room with `private` or `organization` visibility.

A room is Corvic’s isolation boundary — almost every other tool takes a room_id.

Data sources

Tool	Description
`list_data_sources`	List a room’s data sources (paginated).
`get_data_source`	One data source.
`create_data_source`	Create a source for file uploads.
`generate_access_token`	A 15-minute read/write token for uploading to a source or reading a feature view.

Pipelines

Tool	Description
`list_pipelines`	List a room’s pipelines (paginated).
`get_pipeline`	Status overall and per feature view, action-graph lineage, and column schemas.
`create_pipeline`	Create an empty pipeline.
`add_pipeline_source`	Read a data source into a pipeline; returns the refs it produces.
`add_pipeline_action`	Add an action (parse, augment, join, embed, Python/SQL, …); returns the refs it will produce.
`delete_pipeline_action`	Remove an action; refuses (naming blockers) if anything downstream reads its output.
`run_pipeline`	Run the pipeline, or just the actions named in `action_ids` plus their upstream dependencies.
`interrupt_pipeline`	Cancel running actions; finished work is kept, unfinished feature views report `stopped`.
`schedule_pipeline`	Put the pipeline on a cron schedule (5-field cron plus extended forms, with an IANA `timezone`). Replaces any existing schedule.

Pipelines correspond to the Workflows you build on the canvas. Nothing runs as you add sources and actions, so build the whole chain and check it before it costs anything.

Memories

The tools behind Virtual Memory — ask in words, query precisely, or search by meaning.

Tool	Description
`list_memories`	A room’s memories: `memory_id`, title, subject, kind (`relational` \| `semantic`), searchability, and the `artifact` table ID the graph tools take.
`ask_memory`	Ask one memory a question in words; it answers only from what it declares. Runs a model, so it is slow and metered.
`knowledge_graph_catalog`	The sources, dimensions, and measures a memory (or app) declares — read this before searching.
`knowledge_graph_search`	Compile a precise statement against those declarations; results land as a new table.
`semantic_search`	Search an embedding space by meaning; the top 20 most similar rows land as a new table.

Document lenses

The tools behind Document Lens — repeatable readings of one kind of document.

Tool	Description
`list_lenses`	A room’s lenses: what kind of document each reads, the fields it returns, and whether it is ready to call.
`extract_document`	Read one document through a lens; returns `fields`, `reasoning`, `repairs`, `missing_required`, and `extraction_version`. Blocks for the whole reading.
`start_extraction`	The same reading, started asynchronously; returns a `reading_id`.
`get_extraction`	Poll for the reading. Not metered, no deadline — the reading survives a client that walks away.

If missing_required names anything, the lens could not read that document. For long or scanned documents, prefer start_extraction + get_extraction over extract_document, which is bounded by your client’s tool-call deadline.

Agents, threads, and playbooks

For anything open-ended, hand the work to the room’s own agent.

Tool	Description
`list_agent_configs`	List a room’s agent configurations (paginated).
`get_agent_config`	One agent config: name, type, mode, model, schedule.
`list_playbook_templates`	The Playbooks available in a room.
`get_playbook_template`	A playbook’s required inputs, steps, examples, and connector requirements.
`create_agent_config_for_playbook`	An agent config that runs a given playbook.
`list_chat_threads`	Threads, filtered by exactly one of `room_id` or `agent_config_id`.
`get_chat_thread`	One thread’s description and visibility.
`create_chat_thread`	A new thread to ask on.
`query`	Ask the agent; blocks until the turn ends and returns its final message.

query is a full agent turn — minutes for real work, and metered. Whatever it builds lands as tables in the room; find them through get_room_feature_view_registry_ref. Reuse the thread ID for follow-ups, since a new thread starts with no context.

Reading data out

Tool	Description
`get_room_feature_view_registry_ref`	The room’s registry: one row per feature view, with its ref, name, description, and traits.
`fetch_feature_view`	Rows as JSON. `offset`, `limit`, and `columns` are required (an empty `columns` list means every column); `cel_filter` is optional.
`fetch_artifact`	A generated file, or the artifact’s file listing.
`generate_feature_view_shared_url`	An opaque public URL (`table`, `table-json`, or `artifact` mode), valid 15 minutes.
`get_user_content_origin`	The user-content domain and the paths it serves.

Never fetch a table unbounded — ask for a small window first, look at the schema, then widen. Binary columns come back as placeholders, so project with columns.

Models

Call these before writing an action that names a model — the offered set is what your deployment configured.

Tool	Description
`list_completion_models`	Models `llm_augment` accepts. One is marked `is_default`.
`list_text_embedding_models`	Models the text embedding actions accept.
`list_image_embedding_models`	Models `embed_image` accepts.

Custom APIs

Tool	Description
`list_custom_apis`	The external APIs configured in the API Secret Vault and their secret names, injected during Python execution.

Deprecated names

The catalog still answers to old names for old clients. Do not call these in new code:

Deprecated	Use instead
`list_data_apps`	`list_pipelines`
`get_data_app`, `get_workflow`	`get_pipeline`
`run_data_app`, `run_workflow`	`run_pipeline`
`schedule_workflow`	`schedule_pipeline`
`list_feature_views`	`get_pipeline`, or the room registry

MCP API Overview

Agent vs. room access and credentials.

Agent MCP

Talk to a single deployed agent instead.

Virtual Memory

The memories behind ask_memory and the graph tools.

Document Lens

The lenses behind extract_document.

Workflows

Build the pipelines exposed here.

Corvic Tables

Understand feature views and tables.

​Get connected from the MCP Playground

​Conventions

​Rooms

​Data sources

​Pipelines

​Memories

​Document lenses

​Agents, threads, and playbooks

​Reading data out

​Models

​Custom APIs

​Deprecated names

​Related

MCP API Overview

Agent MCP

Virtual Memory

Document Lens

Workflows

Corvic Tables

Get connected from the MCP Playground

Conventions

Rooms

Data sources

Pipelines

Memories

Document lenses

Agents, threads, and playbooks

Reading data out

Models

Custom APIs

Deprecated names

Related