List MCP calls
List the calls assistants made to your account through MCP, newest first.
The account log records what changed. This records what was asked: every tool call, whether it read, changed, was refused, was given wrong parameters or failed. A call made with no valid key is never attributed to an account, so it never appears here.
Calls are kept for 90 days by default, a period the instance sets, and then removed. Nothing here can be written, changed or removed through the API.
This endpoint also answers QUERY, with the same parameters sent in the request body. How QUERY works.
Query parameters
Sent as JSON. Anything not listed here is ignored.
How far back to look: day for the last 24 hours, week for the last 7 days, month for the last 30 days, or all for everything still kept. Any other value answers 422.
Only the calls of one connection, written as its type and its id joined by a colon, such as api_key:14 or oauth_client:9a1f4c2e-7b3d-4f8a-a6e1-3c5d9b2e7f14.
Only the calls with one result: succeeded, refused, invalid_parameters, error or interrupted. Any other value answers 422.
Only the calls to one tool, by its name, such as get-contact.
Which page to answer with.
How many calls one page holds, between 1 and 100. A number outside the range is brought back into it rather than refused.
Returns
A page of calls, and how many there are in all.
The calls on this page. An empty list rather than an error when there are none.
One call.
Always mcp_call. It names the kind of record, so a caller holding several kinds can tell them apart without looking at where they came from.
The id of this record, as a string even when it reads as a number. It is what every endpoint naming this record takes.
Everything the record holds.
When the call started, as a count of milliseconds since 1970. Milliseconds rather than seconds, because an assistant often makes several calls within the same second.
How long the tool took, or null while the call is interrupted.
What the assistant connected with.
api_key for an API key, oauth_client for an assistant connected through OAuth.
The id of the key or of the OAuth client. It stays after the key is revoked or the assistant disconnected.
The name the connection had when the call was made. It is the name you or the assistant chose, not a verified identity of whoever made the assistant.
The name of the tool, as the assistant sent it, such as get-contact.
read, create, update or delete, taken from the API endpoint the tool mirrors. Null for a tool that does not exist.
The vault the call reached, or null when it named none inside your account.
The id of the vault, or null once the vault is deleted.
The name the vault had when the call was made.
The record the call reached, or null when it reached none inside your account.
What kind of record it is, such as contact or life_event.
The name the record had when the call was made.
succeeded, refused, invalid_parameters, error or interrupted. It is read from what the tool answered, never from the HTTP status, which is 200 for a failed tool call too. A call that started and never finished stays interrupted.
Why the call did not succeed, or null when it did.
The JSON-RPC code: -32602 when the parameters or the record were refused, -32603 for anything else.
The reason sent to the assistant, word for word.
The id the assistant gave the request, as a string.
The address the request came from, up to 45 characters.
The parameters the assistant sent, as JSON, with secrets and the values of sensitive fields masked before they were stored.
Whether the parameters were longer than 16 384 bytes and only their beginning was kept, in which case parameters is not valid JSON.
The response the assistant received, as JSON, masked the same way, or null when the call ended in a JSON-RPC error.
Whether the response was cut the same way.
Where to find this record.
The absolute address this record is read from.
Where you are in the collection.
The page this is, counting from 1.
How many one page holds, which is the size that was asked for once it has been brought into the allowed range.
How many calls there are in all, not how many are on this page.
Rate limit
Sixty requests per minute, per user, shared by every key and device. Every answer carries the budget left in X-RateLimit-Remaining.
Keys act as you, with exactly the access you have. Make one per integration so you can revoke it alone.