MCP
Flexkit runs a hosted Model Context Protocol server. Add it to an MCP client and the client’s agent can search your schema, read and change records, search your Knowledge, and use your skills and automations, with your own Flexkit role and space memberships.
https://mcp.flexkit.ioAdd the URL as a remote (Streamable HTTP) MCP server in your client and sign in with your Flexkit account when the client asks. There is nothing to paste besides the URL: the client discovers the sign-in flow on its own.
Choose what the connection reaches
The path after the host limits which projects the agent works on. Your account authorizes the client either way; the path decides what each connection reaches.
TEAM_ID and PROJECT_ID are the IDs in your Dashboard URL (/dashboard/TEAM_ID/PROJECT_ID).
On the account and team URLs, the agent calls list_projects and passes the project to each tool. When you have a single project there, the tools use it automatically. On a project URL, every tool acts on that project.
Read-only connections
Add ?read_only=true to any of the URLs to leave out the tools that change project data. It is a setting of that connection, not a restriction on your account: a client you authorized can reach every URL, so revoke clients you do not trust instead of relying on it.
Example:
https://mcp.flexkit.io/TEAM_ID/PROJECT_ID?read_only=trueQueries, schema and Knowledge search, skills and automations still work. Plugin tools run only when their plugin marks them read-only (searching mail, reading a calendar). Custom tools are left out, since Flexkit cannot tell which of them change data. Updating your agent memory stays available, since it only edits your own notes.
Connect and verify
In the Dashboard, open your project’s API → MCP page for the URL to copy and each client’s setup steps (Claude Code, Claude, ChatGPT and others). It also lists the clients you authorized and lets you revoke them.
- Add the URL in your client’s remote MCP configuration.
- Follow the OAuth sign-in and consent flow with the intended Flexkit account.
- Start with schema inspection or a bounded read-only query.
- Verify the returned data against Studio before letting the client change data.
Client settings differ; use your installed client’s remote-server and OAuth instructions. Do not paste a Dashboard session cookie into its configuration. Opening the URL in a browser brings you to this page.
Capabilities and authority
Every tool acts with your role and space memberships in the project it touches, checked on each call, so a role change applies immediately. Memory is per project and shared with your Chat context.
search_schema without a query returns the entity overview: for every attribute its data type and scope, the unique, required, readOnly and hidden flags, the relationship target, the label when it differs from the name, the option values of a select, and a compact validation JSON Schema. That is enough to write a valid record without reading the GraphQL SDL; pass a query to receive the matching SDL definitions instead. The same deployed schema is available to member sessions through the schema endpoint.
export_records pages through a query on the server and writes every matching record to the file, so the data does not pass through the conversation; it stops after about a minute of paging, or at 250,000 rows or 50 MB, and says so. generate_report stores the HTML page or the PDF rendered from it. Both run with your access, need no approval since they change no project data, and work on read-only connections. Their links are permanent and unguessable, and anyone with a link can open the file, so share them only with people who may see the data.
Plugin tools run with your own connections, as in the Studio chat: your personal account when you connected one, else the project’s. Connect and reconnect plugins in Studio → AI → Plugins; a plugin that is not connected is listed but cannot run. Custom tools run at the project’s production tools URL, with your role and spaces as the actor; a local flexkit dev runtime is only used by the Studio chat and automations.
Sandboxes and monitors are part of the Studio chat and are not offered over MCP.
Approving changes
Before any tool changes the project (records, automations, skills or the schema) or calls a plugin tool that can make changes, Flexkit asks you in your client: the question names the project and shows what will change, like the review card in the Studio chat. Records show their values before and after, schema changes list breaking changes first, automation and skill changes show each setting that changes, and a plugin call shows the account and its arguments. Plugin tools their plugin marks read-only, such as searching or reading email, run without asking, so a task that reads many messages does not ask for each one; anything that sends, changes, labels or deletes asks. Custom tools run without asking, as in the Studio chat, since your application authorizes each call. Nothing runs until you approve, a declined change is not run, and an approval covers exactly the change it showed, once: if the records, automation, skill or schema changed while you were deciding, the change is refused instead. Unlike the Studio review card, the question cannot be edited; decline it and ask the agent for a different version. The question comes from Flexkit through your client, not from the model, so the agent cannot answer it for you.
Asking requires a client that supports MCP elicitation on protocol version 2026-07-28, such as Claude Code. Clients without it can still read data, but Flexkit refuses changes from them; make those changes in Studio instead. Approvals expire after ten minutes.
Revoke an unwanted connection through the supported account authorization controls and remove it from the client. If access fails, check the account, your membership in the team or project in the URL, the consent, and project restrictions.