Table of Contents
When Claude Dispatch fails, do not assume the cause from a generic symptom or keep resubmitting the same task. First confirm that the desktop is available, then test the Dispatch thread with a harmless read, isolate the failing resource, and verify that any previous write did not partially succeed.
Published claims that Dispatch has a universal success rate or that a fixed percentage of failures comes from sleep, networks, or task complexity are not reliable. Results depend on the current app version, plan, operating system, permissions, connectors, network, prompt, and task. Troubleshoot the state you can observe.
Quick checks before diagnosing the task
- Confirm the requirement: Dispatch tasks that use local files or apps run through the desktop, so the computer must be awake and online and Claude Desktop must remain open.
- Check both apps: update Claude Desktop and the Claude mobile app, then confirm that both use the same account and show the same Dispatch thread.
- Check connectivity: verify that the desktop and phone each have a working internet connection. Temporarily rule out a VPN, firewall, or restrictive network if organizational policy permits.
- Check the folder or service: make sure the intended local folder is connected or the connector is still authenticated.
- Run a read-only test: ask Claude to confirm access to one approved folder and return one exact filename without changing anything.
- Capture the symptom: note the visible message, time, affected file or connector, and whether a permission prompt appeared.
Dispatch setup includes an option to keep the computer awake. If you did not enable it, use the operating system's supported power settings while the computer is plugged in. Avoid blindly running background commands copied from an old guide, especially on a managed device.
Use this diagnostic sequence
- Can the phone open the persistent Dispatch thread? If not, update or restart the mobile app, confirm the account, and check the service status.
- Is the desktop awake with Claude Desktop open? If not, Dispatch cannot reach local files or apps. Wake it and reopen Claude Desktop.
- Does a harmless read from an approved folder work? If not, inspect folder permissions, connectivity, and the exact path.
- Does the same task work without a connector or browser? If yes, diagnose that connector, site permission, or authentication separately.
- Did a narrow first step work? If yes, the original prompt may combine too many independent actions. Split it at natural verification points.
- Did the task previously contain a write? Inspect the destination before retrying; a timeout can occur after an action partially completed.
Common symptoms and safe fixes
| Error or symptom | Likely causes to check | Safe next step |
|---|---|---|
| Dispatch thread is unavailable on mobile | Outdated mobile app, wrong account, plan or workspace restriction, temporary service issue | Update the app, confirm the account and workspace, then check Anthropic's status page before changing the desktop setup. |
| Local file or app cannot be reached | Computer asleep or offline, Claude Desktop closed, folder not connected, app permission missing | Wake the computer, open Claude Desktop, confirm the approved folder or per-app permission, and retry a read-only test. |
| “Failed to access file” | Incorrect path, file moved or renamed, folder outside the permitted scope, insufficient filesystem permission | Verify the exact path on the desktop. Ask Claude to list the parent folder without making changes. |
| No response or a task appears stuck | Task still running, waiting for approval, network interruption, service load, or an overly broad request | Check the session details and phone notifications. Do not duplicate a write; inspect the target, then retry only the failed step. |
| Approval notification did not arrive | Mobile notifications disabled, Focus mode, app background restrictions, or the task is visible only in the session | Open the Dispatch thread directly, check notification settings, and answer the pending request there. |
| Connector authentication error | Expired login, revoked access, plan or account mismatch, organization restriction | Test the connector independently, review requested scopes, then disconnect and reconnect it from Customize > Connectors if appropriate. |
| Custom connector times out | Remote MCP server unreachable from Anthropic's cloud, firewall restriction, unstable service, or slow response | Confirm that the server is publicly reachable as required and review firewall and authentication settings. Do not expose an internal server merely to bypass policy. |
| Browser action fails | Claude in Chrome disabled for the task, site permission missing, unsupported page behavior, blocked category, or prompt-injection defense | Refresh the page, verify the extension and site permission, then perform a read-only test on the trusted page. |
| Output is incomplete or unsupported | Ambiguous scope, too much source material, missing file, model error, or an interrupted task | Ask for source filenames and a list of missing items. Split the task and verify each output against the originals. |
| Unexpected file or application changed | Overbroad permission, ambiguous destination, computer-use error, or malicious instructions encountered in content | Stop the session, preserve evidence, restore from version history or backup, revoke unnecessary access, and review the prompt and sources before continuing. |
| Usage or length limit message | Plan limit or request size | Read the reset time or limit shown in the product. Reduce unnecessary context; do not repeatedly resubmit the same large task. |
Split work at verification points
Breaking a broad task into stages does not guarantee a particular success rate, but it makes failures easier to locate and retry.
Instead of:
Review the latest pull request, find security problems, fix TypeScript errors, update tests, commit the changes, and post the review.
Use a sequence such as:
- “Identify the latest pull request in [exact repository] and list its changed files. Read only; do not modify the working tree.”
- “Review [specific changed file] for security risks. Cite line numbers and explain uncertainty. Do not edit.”
- “Run the approved type-check command and report the exact output. Do not fix anything yet.”
- After reviewing the findings, authorize a specific edit in a named branch or worktree.
- Run the relevant tests and show the results.
- Inspect the diff and save a review draft locally. Post or merge only after human approval.
One message may still contain several tightly related operations when they share a single outcome and can be verified together. The goal is not the fewest words; it is a clear boundary between reading, deciding, writing, and publishing.
Make file operations recoverable
- Use an exact source and destination, not “the config file” or “my report.”
- State whether overwriting is allowed. For early workflows, say “If the destination exists, stop.”
- Write drafts to a dedicated folder rather than editing the only copy.
- Ask Claude to read the saved file back and report its path.
- Use version control, document history, or backups for important material.
- Before retrying after a timeout, check whether the file was already created or changed.
Check permissions and connectors deliberately
Cowork permission modes affect when Claude asks before using connector tools. A blocked tool cannot be used, while broader approval modes can allow more actions. Choose the narrowest mode that lets you complete the task, especially when the browser, email, messaging, or write-capable connectors are involved.
When a connector fails:
- Check that the connected service account still works.
- Review whether the connector is enabled in the current conversation or workspace.
- Inspect the scopes and any organization restrictions.
- Disconnect and reconnect only if reauthentication is actually required.
- Test a harmless read before attempting a write.
Do not replace a failed connector with computer-use clicks unless you understand the additional risk. Direct computer use has no sandbox between Claude and desktop applications.
Know when to stop
Stop the task instead of troubleshooting through the agent if it involves an unexpected request for credentials, a new app permission, an unfamiliar download, a suspicious link, financial activity, sensitive personal data, or a change that is hard to reverse. Repeated attempts can turn a small ambiguity into multiple unintended actions.
Also stop after an unexplained destructive or external action. Revoke access if necessary, preserve logs or screenshots, inspect affected systems directly, and follow your organization's incident process.
Create a personal troubleshooting note
For each recurring workflow, record the approved resources, expected output, common visible failures, read-only test, recovery method, and point at which the task must wait for a person. Update the note after app upgrades or connector changes.
Anthropic's official Dispatch setup guide lists current requirements and limitations. For general messages, see its Claude error guide; for integrations, use the connector troubleshooting steps.
Developers diagnosing the separate command-line product can also use this Claude Code installation and diagnostic guide. If you are deciding whether Claude is the right tool for the task, compare ChatGPT, Claude, and Gemini.
Reader Comments 0
Sign in with email or Google to join the discussion.