External Invocations
A log of the MCP tools and Embedded API actions and tools run for your connected users: what ran, for whom, how it ended, and how long it took.
When a connected user's AI agent calls a tool on an embedded MCP server, or your backend executes an action or a tool through the Component Kit, ByteChef records the call. External Invocations lists those records, so you can see which tools your customers use, which ones fail, and how long they take. Open it under Embedded → External Invocations.
The page is available on Enterprise Edition only.
A record holds the identity and outcome of a call only: which tool or action ran, for which connected user, how it ended, and how long it took. The input parameters and the output are never stored.
What is recorded
A record is written for every call made on behalf of a connected user through one of three sources:
| Source | Calls recorded |
|---|---|
| MCP | Component tools and workflow tools an AI agent calls through an embedded MCP server. |
| API (Action) | Actions executed with POST /{externalUserId}/components/{componentName}/versions/{componentVersion}/actions/{actionName}. |
| API (Tool) | Tools executed with POST /{externalUserId}/tools. |
The call is recorded whether it succeeds or fails, including when the connected user has not connected the integration the tool needs. A call that is rejected before it runs, such as a call to a server that is turned off or a call without a connected user, is not recorded.
For a workflow tool, the record also carries the id of the workflow execution the call started. The
API returns it as jobId, so you can find the run in
Workflow Executions.
The page lists embedded calls only. Calls made through an automation MCP server or an A2A server are recorded separately and appear on the automation External Invocations page.
The invocations table
Each call is one row, newest first:
| Column | What it shows |
|---|---|
| Status | How the call ended, for example SUCCESS, ERROR, CONNECTION_REQUIRED or TIMEOUT. |
| Name | The tool name the AI agent called, or the action name your code executed. |
| Type | What was called: a Component tool or a Workflow tool through embedded MCP or POST /{externalUserId}/tools, or an Action through the Embedded API action endpoint. |
| Source | Where the call came in: MCP, API (Action) or API (Tool). |
| Server | The embedded MCP server the call came through. Embedded API calls show a dash. |
| Connected User | The connected user the call was made for. |
| Target | What ran: the component and operation, such as slack / sendMessage, or the name of the workflow a workflow tool ran. |
| Duration | How long the call took, in milliseconds. |
| Date | When the call was made. |
| Actions | View opens the workflow run a workflow tool started, with the input and output of each step. Stop stops that run while it is still going: the call ended with TIMEOUT, APPROVAL_REQUIRED or SUSPENDED. Both are unavailable for component tool and action calls, which start no workflow run. |
The list shows 50 rows per page. Click the refresh button in the page header to load the latest calls.
Click a row to open its details: Status, Target, Component version, Source, Duration, External user, and for a failed call the Error type and Error message.
How a workflow tool call ends
A workflow tool starts a workflow run and waits for it. The status reflects where the run was when the call returned:
| Status | The workflow run |
|---|---|
SUCCESS | Completed, and the call returned its output. |
ERROR | Failed, could not start, or was stopped before it finished. |
TIMEOUT | Was still running when the call stopped waiting for it. |
APPROVAL_REQUIRED | Paused at an approval step and waits for a person to decide. |
SUSPENDED | Paused at a step such as a delay and continues on its own later. |
Filter invocations
| Filter | Narrows the list to |
|---|---|
| Source | Calls that came in through one source: MCP Servers, API (Action) or API (Tool). |
| Status | Calls that ended one way. |
| MCP Server | Calls made through one embedded MCP server. The filter appears when there is at least one embedded MCP server. |
| Start date / End date | Calls made within a date range. |
One connected user's calls
To see the calls made for a single connected user, open the user from Connected Users and select the External Invocations tab in the user's sheet. It lists that user's most recent calls with the same columns.
Your own application can read the same history for a connected user from the embedded API:
GET /api/embedded/v1/{externalUserId}/external-invocations HTTP/1.1
Host: your-bytechef-host.example.com
Authorization: Bearer <API Key>
X-Environment: PRODUCTIONThe endpoint accepts the same source, status, startDate and endDate filters as the page, plus
pageNumber. It returns only the calls of the user named in the path. See the
API reference.
For example, to list the failed tool calls of user-123 since May 1:
curl -G "https://your-bytechef-host.example.com/api/embedded/v1/user-123/external-invocations" \
-H "Authorization: Bearer $API_KEY" \
-H "X-Environment: PRODUCTION" \
--data-urlencode "source=API_TOOL" \
--data-urlencode "status=ERROR" \
--data-urlencode "startDate=2026-05-01T00:00:00Z"source takes MCP_SERVER, API_ACTION or API_TOOL. status takes SUCCESS, ERROR, CONNECTION_REQUIRED,
TIMEOUT, APPROVAL_REQUIRED, APPROVAL_DENIED, RULE_BLOCKED or SUSPENDED. The response is a page of up to 50 records,
without the tool inputs and outputs:
{
"content": [
{
"id": 3107,
"createdDate": "2026-05-18T10:00:00Z",
"source": "API_TOOL",
"kind": "COMPONENT",
"status": "ERROR",
"invocationName": "SLACK_SEND_CHANNEL_MESSAGE",
"componentName": "slack",
"componentVersion": 1,
"operationName": "sendChannelMessage",
"integrationInstanceId": 1052,
"durationMs": 412,
"errorType": "<error type>",
"errorMessage": "<error message>"
}
],
"number": 0,
"numberOfElements": 1,
"size": 50,
"totalElements": 1,
"totalPages": 1
}A workflow tool's record has kind WORKFLOW and carries the jobId of the run it started.
Scope
The list is scoped to the current environment. Switching the environment in the application sidebar reloads the list and clears the filters. A call made in Production does not appear while you are looking at Development.
Who can see it
Reading the log requires a tenant admin. Connected users cannot read the log through the ByteChef UI. Your backend reads a user's history with the API call above.
Retention
Records are deleted after 90 days. A cleanup job runs once a day, at 02:30 server time. In a multi-tenant installation it cleans up every tenant's records, and one tenant's failure does not stop the others.
| Property | Purpose |
|---|---|
bytechef.external-invocation-log.retention-days | How many days a record is kept. Defaults to 90. |
See also
- MCP Servers - expose your integrations as tools to your connected users' AI agents.
- Component Kit - execute actions and tools for a connected user from your own code.
- Workflow Executions - the runs started when an AI agent calls a workflow tool.
How is this guide?
Last updated on