MCP
The endpoint an AI agent connects to, and the deployment's switch for it.
Every project can expose an endpoint an AI agent connects to, so the agent reads the project's own documentation and queries its datasources under the same rules a person gets. The opt-in is per project and off until an owner turns it on in the sidebar, read-only until they say otherwise.
That is why the deployment switch defaults to on, unlike git and
repository commands: a second default-off gate would only mean editing
.env to try it on your own machine.
| Default | ||
|---|---|---|
MCP_ENABLED |
true |
The deployment's kill switch. Off means the route and the panel are both absent. |
MCP_PUBLIC_URL |
WEB_ORIGIN |
What the panel tells people to point their client at. The endpoint is the same process on the same origin as the app, so the address you already gave browsers is the right one — including behind a proxy. Set this only when MCP answers on a different hostname. |
MCP_MAX_ROWS |
200 |
Much lower than the grid's, because the consumer is a context window. A call asking for more is clamped, not refused. |
MCP_MAX_BYTES |
1000000 |
Serialized bytes for one tool call's rows. |
MCP_READ_MAX_BYTES |
65536 |
Per call, reading a file or a resource. The reply says how to continue rather than silently ending. |
MCP_INSTRUCTIONS_MAX_BYTES |
16384 |
Cap on the briefing handed to every client at initialize. |
The sidebar section says what is true of it without being opened: a grey no tokens while nothing can connect, green read only once something can, and violet read/write when what connects can commit. Three words, three colours, and the words are there because colour on its own is not a label anybody can rely on.
An MCP query is an execution like any other: it appears in history under the token's name, with the same timeouts and the same concurrency limit a person has. A read-only project cannot change a row no matter what an agent is asked to run — and a datasource marked read-only stays read-only however the project's mode is set.
Click MCP Server in the sidebar to open its management tab. The sidebar keeps the on/off switch and read/write status; the tab holds mode, functionality switches, tokens, endpoint and client configuration. You can configure it and create or revoke tokens while it is off.
A failed status read keeps the MCP Server launcher visible for owners and shows status unknown, with retry. Open the tab to see the error and retry loading settings. Unknown does not mean off. The header is hidden only when the deployment explicitly reports MCP disabled or the account cannot manage MCP. If settings report a pending migration, follow Upgrading and retry after applying it.
Domain management, sync and Git are separate opt-ins, initially off:
- Domains: list, create, edit and delete domains; assign or unassign objects; set descriptions, ordering, hidden and include-data settings. Hidden domains are excluded from sync.
- Sync: preview or write a datasource snapshot to its configured export directory. Preview is the default. Sync writes files and does not touch Git; commit and push them with the Git tools.
- Git: inspect status, commit, pull and push as separate tool calls. Status fetches first, so behind is about the remote as it is now. Commit stages named paths then commits the whole staged index, including files staged earlier. Inspect status first. Deployment Git must also be enabled, and the datasource must have a repository.
- Pull and push keep a shared repository moving. When other people push to the same branch — hand-written migrations next to DataGripe's export, say — pull fast-forwards, or rebases the clone's unpushed commits onto the remote. Push does the same first, so a remote that moved on other files does not reject it. Neither ever forces or makes a merge commit. A real conflict, uncommitted changes to tracked files when a rebase is needed, or incoming changes to a file with unsaved edits in DataGripe refuse with nothing changed and nothing pushed; that one needs a person with a terminal. The sidebar's own pull stays fast-forward only.
Domain changes, sync writes, commits, pulls and pushes additionally require read/write mode and editor access. Reads and sync previews work in read-only mode once their functionality is enabled. The account must also hold the corresponding domain, sync or Git capability; custom role restrictions still apply. Agents should reuse the same idempotency key when retrying a completed mutation.
The design, the tool list and the token model are in the MCP spec.
This page as Markdown, for an agent or a diff.