---
title: "When a Marketing MCP Call Fails: Claude’s Stop, Check, and Retry Checklist"
description: "Stop Claude from guessing when a marketing MCP call fails. Diagnose access errors, retry safely, and verify budget changes through groas."
url: "https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do"
image: "https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/4f6c718e-9258-42c6-bd71-9ce6efa9bc53.png"
published: "2026-11-20T16:27:22.960Z"
modified: "2026-11-20T16:27:23.020Z"
---

November 20, 2026 · 9 min read

# When a Marketing MCP Call Fails: Claude’s Stop, Check, and Retry Checklist

[](https://groas.com/author/alexander-perelman)

[Alexander Perelman](https://groas.com/author/alexander-perelman)Personally Tested

Head Of Product @ groas

[LinkedIn](https://www.linkedin.com/in/alexander-433793253/)

![A man's finger hovers over a vending machine keypad while a snack dangles on its coil, unsure if it dropped: like an MCP change that may or may not have applied.](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/4f6c718e-9258-42c6-bd71-9ce6efa9bc53.png)

In this article

1. [Set the rule before troubleshooting the connector](#set-the-rule-before-troubleshooting-the-connector)
2. [Match the error before changing settings](#match-the-error-before-changing-settings)
3. [Disconnected server: refresh, reconnect, start fresh](#disconnected-server-refresh-reconnect-start-fresh)
4. [Expired access: sign in before retrying](#expired-access-sign-in-before-retrying)
5. [Rate limits: wait, then make one small read](#rate-limits-wait-then-make-one-small-read)
6. [Refused changes: check scope, not persistence](#refused-changes-check-scope-not-persistence)
7. [Partial results: narrow the request, label the gap](#partial-results-narrow-the-request-label-the-gap)
8. [Make Claude’s failure report useful](#make-claudes-failure-report-useful)
9. [Retry reads once; verify writes first](#retry-reads-once-verify-writes-first)
10. [Confirm the change before calling it done](#confirm-the-change-before-calling-it-done)
11. [Put the rule in the project, then prove the connection](#put-the-rule-in-the-project-then-prove-the-connection)
12. [Run the failure drill everyone skips](#run-the-failure-drill-everyone-skips)

A failed MCP call should produce an error, not an invented spend number. The expensive part is the assistant’s next sentence: “The budget change applied,” when it has no idea whether it did. I run paid search from Claude and ChatGPT through [groas](https://groas.com/mcp), and my rule is simple: **stop, name the failure, check what happened, then act once.** Use this checklist to distinguish expired access from a rate limit or a refused change, give the assistant a safe response pattern, and verify changes before approving a resend.

## Set the rule before troubleshooting the connector

1. **Require a failed call to produce a failure report, not an answer from memory.** Spend, CPA, account status, and confirmation of a change all need tool evidence.
2. Tell Claude to name the tool, quote the error, and state what remains unknown before proposing a fix.
3. Make the retry limit explicit: one corrected or safely repeated call, then stop and report. Do not leave “try again” open-ended.
4. Treat error wording as diagnostic information, not something the assistant should smooth into a reassuring sentence. The [MCP error guide](https://trustedmcp.io/guides/mcp-tool-errors) explains why clear error text matters.
5. Separate a failed read from an uncertain write: missing spend data means the number is unknown; a disconnected budget change means its outcome is unknown.
6. Use the [verification reporting pattern](https://onesignaldesk.com/guides/stop-ai-agents-trusting-wrong-tool-outputs): what tool ran, what was expected, and what came back.
7. **Make “I do not have this” an acceptable answer.** A plausible range is still a guess when the account read failed.

![Clay marketer facing a laptop tool error while another figure shrugs](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/f6c9628b-33e3-47b6-88a0-de81c9203946.png)

## Match the error before changing settings

1. Look for `Couldn't reach the MCP server`: start with the connection and handshake, not account permissions.
2. Look for `Authorization failed`: check sign-in before rebuilding the connector.
3. Look for `Unexpected error while invoking tool`: the connection may exist while the individual call fails; inspect that call’s error.
4. Keep the exact message in the failure report so the next step follows the evidence, not the assistant’s interpretation. Use [Claude connector troubleshooting](https://claude.com/docs/connectors/building/troubleshooting) to match the message.
5. **Do not change the URL, permissions, and request at the same time.** Match the failure first, then apply the relevant checklist below.

## Disconnected server: refresh, reconnect, start fresh

1. Confirm the message concerns reaching the MCP server rather than an expired sign-in or a refused action.
2. Hard-refresh Claude, then check that groas still appears under Connectors before removing anything.
3. Check the handoff: stale sign-in, an old tool list, or a moved menu can look like a server problem. Follow the [connector handoff checks](https://agentmsgr.com/blog/custom-connector-not-working/).
4. Reconnect once; if the connection still fails, remove groas and add it back with `https://mcp.groas.ai/mcp` for a fresh sign-in and tool list.
5. **Start a new conversation after reconnecting.** Do not assume the old chat has picked up the refreshed tools.
6. Prove the connection with one narrow, read-only request before returning to the original task; leave writes alone until that read succeeds.

## Expired access: sign in before retrying

1. Look for `asks to sign in again`, `401`, or `Needs authentication`; start with access rather than assuming the configuration is broken. See the [expired-access signs](https://prompt-architects.com/blog/151-mcp-not-working-15-fixes).
2. Reconnect groas under Connectors, finish sign-in, then retry the same read once.
3. If sign-in loops daily, check for two signed-in accounts, a password change, or blocked third-party cookies before changing connector settings.
4. **If Claude Code flags Re-authenticate in `/mcp`, log out and log in again.** Do not edit the connector URL to solve an authentication prompt.
5. If the read still fails after authentication, stop and report the new error rather than treating another sign-in attempt as progress.
6. Keep the original request marked unresolved until a successful read supplies the data; a completed login is not a completed account check.

## Rate limits: wait, then make one small read

1. Treat `429` as a signal to stop calling, not a reason to press retry harder.
2. Honor the wait specified by `Retry-After` before one retry; the [429 handling guidance](https://docs.scopestack.io/help/integrations/mcp-connecting) makes that the next step.
3. Report which call hit the limit and what wait the response specifies, rather than saying only that the service is busy.
4. **After the wait, run the single cheapest read that can confirm access.** Do not immediately repeat the oversized request that left the task unresolved.
5. If the retry also fails, report it and stop; another loop does not make the account data more available.
6. Keep the missing number missing while waiting. Yesterday’s remembered spend is not a substitute for the requested read, even if it looks reasonable.

## Refused changes: check scope, not persistence

1. If reads work but create, send, or post fails with `Unauthorized` or a role error, check permissions before retrying. Use the [role-check guidance](https://docs.scopestack.io/help/integrations/mcp-connecting).
2. **Read the tool list before trusting that a change is possible.** Confirm the write tool exists, what scope it needs, and what it logs.
3. Use [the groas tool-description checklist](https://groas.com/post/tool-descriptions-decide-what-claude-does-how-to-read-a) to separate an available action from one the assistant merely describes confidently.
4. If reconnecting is necessary, review the requested permissions and approve only the access needed for the intended action; do not grant everything just to clear an error.
5. Run a successful read after reconnecting, then resolve the write permission separately. Read access does not prove the change is authorized.
6. Keep approval separate from capability: a tool being available is not permission for the assistant to move budget without your go-ahead.
7. **Do not repeat a refused write unchanged.** Fix the permission problem, confirm the intended action, and only then consider another attempt.

## Partial results: narrow the request, label the gap

1. If the result stops early or says it is truncated, request one account, one date range, and one metric set instead of asking for the whole business again.
2. Watch for response-size warnings; Claude Code can warn or truncate large outputs, as covered in the [truncation troubleshooting notes](https://prompt-architects.com/blog/151-mcp-not-working-15-fixes).
3. **Label partial results as partial and stale data as stale.** Do not present either as a complete, live account view.
4. Treat a successful empty result as “ran with no results,” not as a failed call; do not confuse an error or an incomplete response with a valid empty answer.
5. If the tool list shows zero tools or old tools, refresh or reconnect, then start a new conversation before narrowing the request.
6. Follow the [narrow-ask method for big accounts](https://groas.com/post/why-marketing-mcp-answers-go-wrong-on-big-accounts-context): keep the scope small enough to inspect what actually came back.
7. Report the scope with the answer so one account’s result does not quietly become an account-wide conclusion.

## Make Claude’s failure report useful

1. Name the tool that failed and the task it was supposed to complete; “something went wrong” leaves the next person starting from scratch.
2. Quote the error exactly, including `401`, `429`, or `Retry-After` where shown; keep the diagnostic detail intact.
3. State what is unknown: the spend number, the account status, or whether the requested change applied.
4. Give one next step: reconnect, wait, correct the request, narrow the scope, or check the change status.
5. **Never fill the gap with a remembered number or plausible range.** Say “I do not have this” and leave the decision unsupported until a read succeeds.
6. Before proposing a resend that could move budget, state the account, the exact change, and how the result will be confirmed.
7. Keep the report short enough to act on, but do not remove uncertainty to make it sound cleaner. “Unknown” is useful; false reassurance is not.

## Retry reads once; verify writes first

1. Classify the error before retrying: transient connection failures, timeouts, rate limits, and some server errors differ from malformed requests and permission refusals. Use the [retry classification guidance](https://www.wati.io/en/blog/ai-agent-mcp-server-outage).
2. Retry a safe-to-repeat read once after a transient failure, honoring `Retry-After` when it is provided.
3. Reconnect first for `401` or `Needs authentication`, then retry the read once after sign-in completes.
4. Correct a malformed request before trying again; sending the same invalid request is repetition, not troubleshooting.
5. Narrow a truncated request before retrying: one account and one date range, not the same oversized answer with “please” added.
6. **Never auto-retry a send, budget change, or delete.** Check whether it applied, then ask for approval before any resend.
7. Stop after the retry fails and report both attempts; do not let “one more try” become an unattended loop.
8. Keep retries bounded even when the error appears temporary. The task is to get a verified result, not to keep the assistant busy.

## Confirm the change before calling it done

1. Treat a disconnect during a write as an unknown outcome, not proof of failure; the action may have completed before the connection broke. See the [disconnect-handling guidance](https://www.wati.io/en/blog/ai-agent-mcp-server-outage).
2. Run a read-only identity check, then read the object you tried to change so the verification concerns the right account and action.
3. Open the groas change log and match the timestamp, account, and reason to the attempt; [groas](https://groas.com/mcp) logs every change with its reason and supports one-step undo.
4. **If the log and platform agree the change applied, report the confirmed value and do not resend.** The missing response is not a reason to make the change twice.
5. If the log and platform agree it did not apply, fix the cause and ask for approval before trying once more.
6. If the log and platform disagree, stop, report both values, and wait for a human decision rather than choosing the more convenient answer.
7. Before any approved resend, repeat the account, exact change, and verification plan. Approval should attach to a specific action, not an open-ended instruction to fix things.

## Put the rule in the project, then prove the connection

1. **Add an explicit failure rule to the project instructions.** A working connector does not, by itself, tell the assistant how to handle an unsuccessful call.

2. Use this block in the Claude Project or ChatGPT Project that handles the account:

   ```text
   When a groas tool call fails, stop. Name the tool, quote the error, and say what is unknown. Never answer spend, CPA, or status from memory.

   Retry a safe-to-repeat read once for a transient failure, honoring Retry-After when provided. For 401 or Needs authentication, reconnect first, then retry the read once. Correct malformed requests and resolve permission errors before another attempt. Narrow truncated requests.

   Never resend a write without checking its status and getting approval. If its outcome is unclear, say so. Label stale data as stale and successful empty results as no results. If the retry fails, stop and report.
   ```

3. Connect groas in [Claude](https://groas.com/claude) under Connectors with `https://mcp.groas.ai/mcp`, sign in, and request yesterday’s spend by account before allowing a write.

4. If using [ChatGPT](https://groas.com/chatgpt), connect the custom MCP server with OAuth at the same URL and request the same live read.

5. Set approvals to hold every write until you say go, then repeat a narrow read after reconnecting to confirm the connection works.

6. For teams and agencies, put the rule in the shared project rather than relying on each teammate to remember it in each chat.

7. **Require a tool timestamp behind every spend number.** A fresh result in one conversation does not make a remembered number in another conversation fresh.

![Two ad-budget dials, one verified in the change log and one with an unknown status](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/fa4fb12b-afb3-41ae-84f6-2da5db70738c.png)

## Run the failure drill everyone skips

1. Disconnect groas, ask Claude for yesterday’s spend by account, and confirm it says it does not have the number instead of guessing.
2. Check that the response names the failed tool or connection, preserves the error information available, and states what remains unknown.
3. Reconnect, narrow the request to one account and one date range, and allow one read retry; confirm the answer comes from the returned data.
4. Before approving the next change, check the account and change log together: match the timestamp, account, reason, and resulting value.
5. **Finish the failure drill with the log check.** This is the item most often skipped: confirming what actually happened before acting again. A retry costs seconds; a duplicate budget change costs margin, and a guessed CPA costs trust. Stop, name it, check the log, then act once.

## Frequently Asked Questions

### What should Claude do when a groas MCP call fails instead of giving me a spend number?

Claude should stop, name the tool, quote the exact error, and state what remains unknown. Never let it fill the gap with a remembered number or plausible range, because that guess can be presented as a real result. Make "I do not have this" an acceptable answer.

### How do I tell the difference between a connection problem, expired access, and a rate limit?

Match the error message first. \`Couldn't reach the MCP server\` points to the connection and handshake, \`Authorization failed\` or \`401\` points to sign-in, \`429\` is a rate limit to wait on, and \`Unexpected error while invoking tool\` means the connection exists but the individual call failed. Keep the exact message in the failure report and fix one thing at a time.

### I reconnected groas in Claude but it still seems stuck. What should I do after reconnecting?

Start a new conversation, because the old chat may not have picked up the refreshed tools. Then prove the connection with one narrow, read-only request, such as asking for yesterday's spend by account. Leave writes alone until that read succeeds.

### Claude keeps hitting a 429 error when pulling campaign data. Should I just keep retrying?

No. Treat \`429\` as a signal to stop calling, honor the wait specified by \`Retry-After\`, and then make one retry. After the wait, run the single cheapest read that can confirm access rather than repeating the oversized request. If the retry also fails, report it and stop.

### Claude lost connection while changing an ad budget. Should I resend the change?

Not until you verify what happened. A disconnect during a write is an unknown outcome, not proof of failure, so the change may have applied before the connection broke. Run a read-only check on the object you tried to change, then compare the groas change log against the platform. If both agree the change applied, report the confirmed value and do not resend.

### The MCP response got truncated when I asked for too much data. How should I rephrase the request?

Narrow the request to one account, one date range, and one metric set instead of asking for the whole business again. Label the earlier partial result as partial and stale data as stale, and report the scope with the answer so a result for one account is not read as account-wide.

### How do I make Claude handle MCP failures consistently across my whole team?

Add an explicit failure rule to the shared Claude Project or ChatGPT Project instructions, since a working connector alone does not tell the assistant how to handle an unsuccessful call. The rule should require stopping, naming the tool, quoting the error, retrying safe reads once, and never resending a write without checking its status and getting approval.

### How can I test whether Claude handles a broken groas connection correctly?

Disconnect groas and ask Claude for yesterday's spend by account, then confirm it says it does not have the number instead of guessing. Reconnect, narrow the request to one account and one date range, and allow one read retry. Finish by checking the change log against the platform before approving the next change, since that log check is the step most often skipped.

## Run Your Marketing On groas

Plug ChatGPT, Claude or any AI agent into groas and let it do the work of a whole marketing team: $19 a month flat with every feature included, and autopilot whenever you’re ready.

[Start For $19](https://groas.com/pricing)

[Or have groas run it for you](https://groas.com/done-for-you)

## Structured data

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://graph.groas.com/entity/groas","name":"groas","url":"https://groas.com/","description":"groas is the marketing infrastructure for your own AI: it connects your AI to all your marketing and gives it proprietary marketing data collected 24/7. Plug Claude, ChatGPT or any AI agent into groas and run all your marketing from it, for a tiny fraction of the usual cost, with the groas app there whenever you want to see more. It is the layer ChatGPT, Claude and any AI assistant run a business’s online marketing on. Through groas an assistant can do anything you could do natively in Google Ads, Bing Ads, Meta Ads and ChatGPT ads, on your website and in AI search, and, where the owner switches it on, can run connected campaigns by itself, so AI agents do the work a whole marketing team used to. groas also gives those agents large-scale proprietary marketing data it collects around the clock, on any schedule the owner chooses: every ad competitors run on Google and Meta, their landing pages and full funnels with the offers and prices in them, Reddit customer research, AI search answers from ChatGPT, Google AI Overviews, Gemini, Claude and Perplexity, and SEO data from Semrush and Ahrefs, backlinks included ($19 a month, flat, with everything included, for unlimited teammates on one shared workspace and unlimited ad accounts and websites, delivered as an MCP server). Separately, its done-for-you service has senior people and proprietary quantitative systems run paid ads, SEO and AI search for businesses, and white label for agencies, from $5,000 a month.","logo":{"@type":"ImageObject","url":"https://groas.com/icon-512.png","width":512,"height":512},"sameAs":["https://www.linkedin.com/company/groas/"]},{"@type":"WebSite","@id":"https://groas.com/#website","name":"groas","url":"https://groas.com/","inLanguage":"en-US","publisher":{"@id":"https://graph.groas.com/entity/groas"}},{"@type":"WebPage","@id":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do#webpage","url":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do","name":"When a Marketing MCP Call Fails: Claude’s Stop, Check, and Retry Checklist","description":"Stop Claude from guessing when a marketing MCP call fails. Diagnose access errors, retry safely, and verify budget changes through groas.","inLanguage":"en-US","isPartOf":{"@id":"https://groas.com/#website"},"about":{"@id":"https://graph.groas.com/entity/groas"},"primaryImageOfPage":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/4f6c718e-9258-42c6-bd71-9ce6efa9bc53.png"},"breadcrumb":{"@id":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do#breadcrumb"},"dateModified":"2026-11-20T16:27:23.020Z","mainEntity":{"@id":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do#article"}},{"@type":"BreadcrumbList","@id":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://groas.com/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://groas.com/blog"},{"@type":"ListItem","position":3,"name":"When a Marketing MCP Call Fails: Claude’s Stop, Check, and Retry Checklist","item":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do"}]},{"@type":"BlogPosting","@id":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do#article","url":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do","mainEntityOfPage":{"@id":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do#webpage"},"isPartOf":{"@id":"https://groas.com/blog#blog"},"headline":"When a Marketing MCP Call Fails: Claude’s Stop, Check, and Retry Checklist","description":"Stop Claude from guessing when a marketing MCP call fails. Diagnose access errors, retry safely, and verify budget changes through groas.","datePublished":"2026-11-20T16:27:22.960Z","dateModified":"2026-11-20T16:27:23.020Z","image":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/4f6c718e-9258-42c6-bd71-9ce6efa9bc53.png"},"author":{"@id":"https://groas.com/author/alexander-perelman#person"},"publisher":{"@id":"https://graph.groas.com/entity/groas"},"keywords":"retry, before, mcp, stop, change, call, fails, check","wordCount":1957,"inLanguage":"en-US"},{"@type":"Person","@id":"https://groas.com/author/alexander-perelman#person","name":"Alexander Perelman","jobTitle":"Head Of Product @ groas","description":"My background includes Goldman Sachs and Stanford Computer Science.","image":"https://groas.com/media/blog/4f17cac3a81acc224cc1b2cfab8f94bbc295a3089d13e8421125771ef8d4064f.jpg","worksFor":{"@id":"https://graph.groas.com/entity/groas"},"sameAs":["https://www.linkedin.com/in/alexander-433793253/"]},{"@type":"FAQPage","@id":"https://groas.com/post/when-a-marketing-mcp-tool-call-fails-what-claude-should-do#faq","mainEntity":[{"@type":"Question","name":"What should Claude do when a groas MCP call fails instead of giving me a spend number?","acceptedAnswer":{"@type":"Answer","text":"Claude should stop, name the tool, quote the exact error, and state what remains unknown. Never let it fill the gap with a remembered number or plausible range, because that guess can be presented as a real result. Make \"I do not have this\" an acceptable answer."}},{"@type":"Question","name":"How do I tell the difference between a connection problem, expired access, and a rate limit?","acceptedAnswer":{"@type":"Answer","text":"Match the error message first. `Couldn't reach the MCP server` points to the connection and handshake, `Authorization failed` or `401` points to sign-in, `429` is a rate limit to wait on, and `Unexpected error while invoking tool` means the connection exists but the individual call failed. Keep the exact message in the failure report and fix one thing at a time."}},{"@type":"Question","name":"I reconnected groas in Claude but it still seems stuck. What should I do after reconnecting?","acceptedAnswer":{"@type":"Answer","text":"Start a new conversation, because the old chat may not have picked up the refreshed tools. Then prove the connection with one narrow, read-only request, such as asking for yesterday's spend by account. Leave writes alone until that read succeeds."}},{"@type":"Question","name":"Claude keeps hitting a 429 error when pulling campaign data. Should I just keep retrying?","acceptedAnswer":{"@type":"Answer","text":"No. Treat `429` as a signal to stop calling, honor the wait specified by `Retry-After`, and then make one retry. After the wait, run the single cheapest read that can confirm access rather than repeating the oversized request. If the retry also fails, report it and stop."}},{"@type":"Question","name":"Claude lost connection while changing an ad budget. Should I resend the change?","acceptedAnswer":{"@type":"Answer","text":"Not until you verify what happened. A disconnect during a write is an unknown outcome, not proof of failure, so the change may have applied before the connection broke. Run a read-only check on the object you tried to change, then compare the groas change log against the platform. If both agree the change applied, report the confirmed value and do not resend."}},{"@type":"Question","name":"The MCP response got truncated when I asked for too much data. How should I rephrase the request?","acceptedAnswer":{"@type":"Answer","text":"Narrow the request to one account, one date range, and one metric set instead of asking for the whole business again. Label the earlier partial result as partial and stale data as stale, and report the scope with the answer so a result for one account is not read as account-wide."}},{"@type":"Question","name":"How do I make Claude handle MCP failures consistently across my whole team?","acceptedAnswer":{"@type":"Answer","text":"Add an explicit failure rule to the shared Claude Project or ChatGPT Project instructions, since a working connector alone does not tell the assistant how to handle an unsuccessful call. The rule should require stopping, naming the tool, quoting the error, retrying safe reads once, and never resending a write without checking its status and getting approval."}},{"@type":"Question","name":"How can I test whether Claude handles a broken groas connection correctly?","acceptedAnswer":{"@type":"Answer","text":"Disconnect groas and ask Claude for yesterday's spend by account, then confirm it says it does not have the number instead of guessing. Reconnect, narrow the request to one account and one date range, and allow one read retry. Finish by checking the change log against the platform before approving the next change, since that log check is the step most often skipped."}}]}]}
```
