What the agent does
This is for an accounting firm that runs its client admin on n8n: a workflow that files documents from the client portal into each client's folder, one that sends the engagement-letter reminder, one that copies approved time entries into the billing system, one that emails a client when their return is ready to sign. Two or three people at the firm edit these workflows, and the partner who answers for client communication finds out about a change only when a client replies to an email nobody meant to send. The agent writes the partner a short weekly log of every change, in the firm's words. Someone with access to the firm's instance connects it once with POST /me/connections {"type": "mcp", "name": "n8n", "url": "https://yourfirm.app.n8n.cloud/mcp-server/http", "auth": {"mode": "oauth"}}, after an owner or admin turns on Settings > Instance-level MCP in n8n.
n8n exposes a workflow's details only when that workflow is enabled for MCP, one by one or by project, while search_workflows returns a preview of every workflow the signed-in person can view. The agent starts there, so it can list the workflows it cannot read instead of silently skipping them. For each enabled workflow it calls get_workflow_history, which lists saved versions newest first (50 a page, with offset for more), and keeps the versions saved since last Friday at 15:00 in the firm's timezone.
For each workflow with new versions it calls get_workflow_versions_diff with the last version before the window as fromVersionId and the newest as toVersionId, and get_workflow_version on the newest when it needs to read a node in full. n8n reduces credentials in a version to their id and name, so the agent sees which account a node uses, never a password. It then calls search_workflow_executions for each changed workflow with startedAfter set to the change and status of error, to say whether the new version has failed since.
From that it writes changes.md: one section per changed workflow with its name, when it changed, what changed in plain English ("the reminder now waits 5 days instead of 7", "a new branch sends a Gmail message when the portal upload has no tax year"), and whether it has failed since. A change that adds, removes or edits a node that sends mail (n8n-nodes-base.gmail, n8n-nodes-base.microsoftOutlook, n8n-nodes-base.emailSend) or that changes which credential such a node uses goes to the top, because that is the change that reaches clients. It runs, publishes and unpublishes nothing.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Write this week's change log for our n8n instance, covering changes saved since 2026-11-06 15:00 America/New_York. Use the n8n connection. Call search_workflows to list every workflow I can view; list by name any that are not enabled for MCP and so cannot be read. For each enabled workflow call get_workflow_history and keep versions saved in the window. For each workflow with new versions call get_workflow_versions_diff from the last version before the window to the newest, and get_workflow_version on the newest if you need a node in full. Then call search_workflow_executions with startedAfter set to the change time and status error. Write changes.md: one section per changed workflow with name, change time, what changed in plain English, and failures since. Put first any change that adds, removes or edits a node of type n8n-nodes-base.gmail, n8n-nodes-base.microsoftOutlook or n8n-nodes-base.emailSend, or changes the credential one uses. Name credentials by their name only. Do not call execute_workflow, test_workflow, publish_workflow, unpublish_workflow or any tool that creates or updates anything.",
"connections": [
"n8n"
]
}- Get an API key. In the console, open Settings, then API keys, and create a key for your org. Send it in the Authorization header as a Bearer token.
- Send the request. POST the body in request.json to https://api.phoenix.so/o/{org}/sessions, with your repository as the resources URL. The run starts in its own cloud sandbox.
- Read what came back. Follow the run on its live stream or wait for the run.completed webhook, then read the result and its trajectory.
What comes back
The session returns changes.md as an artifact and a short answer naming how many workflows changed and which ones touch client email. The partner reads the top section on Friday afternoon, before the firm's weekend reminders go out, and asks whoever made a change about anything they did not expect.
What the partner checks: that every workflow they care about is enabled for MCP (the log lists those it could not read), that a change described as harmless really is, by opening the diff in n8n's own version history, and any change with failures since it was saved. The log describes what the workflow now does; whether a client should get that email is the partner's decision. The transcript has every MCP call and what it returned, so a line that looks wrong can be traced to the version it came from.
Run it on every Friday afternoon
In the console, a Project automation with a schedule trigger (0 15 * * 5 with timezone set to the firm's) runs this prompt every Friday. A schedule trigger runs the saved prompt as written and fills in no dates, so the saved prompt asks for "changes saved in the seven days before this run" in place of the fixed time above. Its notify_each_run set to true sends each Friday's result to the person who set it up on its own, instead of folding it into one daily message.
Each week's run is its own session, so the firm keeps a dated record of what its automations did each week. When a client asks in March why they got a reminder in November, the November log names the change and links back to the version.
The use case this job follows, with its automation and trigger: Produce a report or file, no repo.
Why run it on Rush?
n8n keeps every version, but nobody at a small firm reads the version history of eight workflows every week. The agent reads all of it and writes only what changed, in the words the partner uses, with the changes that reach clients first. The n8n sign-in stays in Rush's connection, limited to the workflows the firm enabled for MCP, not a personal API token in a script on the operations manager's laptop.
Each run has its own computer and a record of every call it made and what its model calls cost, which matters here because the agent is reading workflows that send client email. The prompt forbids running or publishing anything, and the transcript shows that no such call was made.
Frequently asked questions
Which n8n versions does this need?
n8n's tool reference lists version history from 2.29.0 and the version diff from 2.36.0. On an older instance the agent can still read history and compare two versions itself with get_workflow_version, which the prompt can ask for instead.
Will it see workflows we have not enabled for MCP?
Only their names and previews, from search_workflows. Their history and contents stay closed to it, so the log lists them by name as not read. Enable the ones the partner cares about, one by one or for a whole project or folder.
Can the agent see client data or passwords in the workflows?
It reads workflow definitions, not the documents that pass through them, and n8n reduces credentials in a version to an id and a name. The prompt also keeps it off execution data: it reads only whether runs failed.
Our instance is self-hosted. Does that matter?
The instance must be reachable from the internet for a cloud client to connect, and any proxy in front of it must pass n8n's MCP-Protocol-Version, Mcp-Method and Mcp-Name headers. Connect it with the instance's own address in place of the n8n Cloud one.