Claude Research Not Working: Access and Run Checks

When Claude Research is not working, first separate a missing Research control from a failed run or an unavailable source. Check your account access and the current Research instructions before changing settings. fenbs can retain the question, checked links, and unfinished steps while you diagnose the problem, so a retry has a clear starting point.

Which part of the workflow has stopped?

1. Name the symptom: Research is absent, the request does not start research, the run errors, or expected sources are missing.
2. Record the account and workspace you are using, the app or browser, and the exact visible message. Remove private information from screenshots you plan to share.
3. Check the current feature instructions and any workspace restrictions. A control shown in an old tutorial may have moved.
4. Try a small question using public sources, then add the original sources back individually. Keep the test separate from the full assignment.

For example, “the button is absent in my work workspace” is a different problem from “research starts but cannot read our document.” Changing a connector will not establish account eligibility. A good diagnosis changes one condition at a time and records the result instead of repeating the same large prompt.

Why is the Research control missing?

Use Anthropic's Research documentation↗ (opens in a new tab) to check current account eligibility and where to enable it. The guide also identifies model restrictions a workspace administrator may need to review. Sign in to the intended account and select the intended workspace before comparing your screen with the instructions.

Research requires web search. Anthropic's web-search guide↗ (opens in a new tab) explains both workspace enablement and interface differences, including experiences without a separate search toggle. A missing toggle alone is therefore not enough to conclude that search is unavailable. Ask your workspace owner to check a restriction rather than attempting to bypass it.

If access appears correct but the screen differs, compare a current supported app or browser while preserving the original conversation. Record whether the difference follows the account or the surface. Avoid repeated sign-outs or broad browser-data deletion as your first step; those actions remove useful state without necessarily addressing a feature-access problem.

What if Claude answers without doing research?

In a new, short conversation, explicitly ask Claude to use the research tool for one bounded question. Choose something that genuinely needs source gathering, such as comparing the documented accessibility features of two named products. Require original-source links and a list of anything it could not verify.

Inspect the result rather than accepting the word “researched.” Can you open the cited pages? Do those pages support the statements? A polished answer without evidence is an output to review, not proof that the requested research workflow ran. Save confirmed links in fenbs Links and listings so a later retry can begin with evidence already checked.

If an unrelated Claude feature is failing, use its own troubleshooting path. Claude Code checks address the coding tool, while Cowork checks address that separate workspace. Instructions for one product should not be applied automatically to another.

Why are connected documents missing from the answer?

Check where you started the run. Anthropic documents that Research started on mobile searches the web without using connectors; use the web or Desktop for research that needs them. The same guide says Research cannot use local MCP server tools configured in Desktop. MCP is a standard way for an AI app to use an outside tool.

For a supported connector, confirm it is enabled for the chat and ask for the specific source you need. Open that source yourself using the same organizational identity where possible. Access to the connector is not evidence that every document in the connected service is readable.

Test one nonconfidential document before retrying a collection. If the document cannot be found, record its title, expected location, and the failure without copying private content into a report. The Desktop MCP troubleshooting guide covers connection problems separately; fixing such a connection does not remove Research's documented tool restrictions.

What should you do when a run returns an error?

Copy the message and check the official service status↗ (opens in a new tab). Read the affected component and incident details rather than assuming every error has a shared cause. Follow Anthropic's error guidance↗ (opens in a new tab) for explicit usage or length messages. Do not use repeated retries as a substitute for the recovery the message describes.

If a small public-source test succeeds, retry only one part of the original question. Ask for a compact result with the claim, source link, and uncertainty. If that fails too, preserve the exact test and result. This gives you a reproducible example rather than a vague report involving a long private conversation.

Use Anthropic's support route↗ (opens in a new tab) if the problem persists. Include the time and time zone, app version, workspace context, error wording, and what the small test showed. Share only what support needs through the appropriate channel, without passwords, tokens, or confidential research attachments.

How to do this with fenbs

1. Create a free account and add a project for the research question, with the intended outcome and source boundaries.
2. Add checked references to Links and listings. Use context notes to distinguish confirmed claims from search leads you have not opened.
3. Add a recovery task with the symptom, safe reproduction steps, checks already tried, and unanswered questions.
4. Follow the integration guide for your assistant, ask it to read the project context and continue the remaining work, then save a precise handoff before stopping.

fenbs does not enable Research, repair Claude, or grant access to connected documents. Use it when your sources and recovery steps need to survive across sessions. It is unnecessary for a brief disposable test; a clear local note may be enough. The AI at work hub covers longer workflows where that shared record becomes useful.