Claude Ask Your Org: Claude enterprise search pricing, limits and connector setup

Claude Ask Your Org is a pre-configured project that Anthropic adds to the sidebar of Claude Team and Enterprise organizations, designed for searching across company knowledge sources. It is enabled by default in Admin settings, but an owner has to complete a setup flow before anyone else can use it. The project name follows the format "Ask [Name]", so whatever an owner types during setup is what everyone sees in the left panel.

Anthropic describes four things the project provides: guided setup for connecting work apps, organization-wide availability once an owner finishes the initial setup, "specialized system prompts maintained by Anthropic for effective searches", and unified access across multiple data sources. The distinction from an ordinary Claude project is worth understanding, because it is the whole of what you are getting. In a standard project the system prompt is customizable by users and connectors are configured manually each time. In the enterprise search project the system prompt is managed by Anthropic and optimized for search, the connector setup is a guided onboarding, the project is provisioned automatically for the organization, and it receives "automatic improvements from Anthropic" rather than staying static.

In other words, Ask Your Org is a very good prompt plus a guided connector flow, wrapped around the connector infrastructure Claude already had. That is a real convenience and it is free with the seat. It is not a new retrieval system.

No. Anthropic states in the privacy section of its enterprise search documentation: "No external indexing: Search results are generated by making MCP calls. No data from connected services are indexed in our systems for serving queries." Every question you ask triggers live calls out to Slack, SharePoint, Google Drive or whatever else is connected, in the moment, and Claude reasons over whatever comes back.

For a security review this is close to an ideal answer. Your documents are not copied into a vendor index, so there is no second corpus to worry about, no re-indexing lag, and nothing to purge if you leave. Several buyers will choose Ask Your Org for this reason alone, and that is a defensible decision.

The trade is throughput and completeness. A search index answers from a structure built in advance, so it can rank across millions of documents in one pass. Live calls answer from whatever a handful of API requests return inside a timeout. Anthropic documents the consequences honestly in the same article. There is a troubleshooting entry for "My search results seem incomplete" and another for "Claude is responding slowly in my search project", whose listed causes begin with "Searching across many large data sources takes time". The suggested fixes are to narrow the scope, search one tool at a time, and be more specific about what you are looking for.

The best-practices section goes further and tells you to work the same way by default: "Instead of Find information about the product launch, try Search Slack and Google Drive for discussions about the Q4 product launch", and "Break complex queries into steps: For thorough research, ask Claude to search one source at a time, then synthesize findings." That guidance is good advice for the architecture. It is also, read plainly, a statement that a single broad question across everything is not the reliable path, which is exactly the question most people want to ask.

An owner clicks "Ask Your Org" in the left menu, chooses "Set up for your org", and is then required to pick a connector for both Documents and Chat. Email is recommended but optional. Only enabled connectors are available inside the project.

The full owner flow runs like this. Open the project from the sidebar. Choose "Set up for your org" to continue, or "Disable" to turn the project off for the team. Connect the required categories, then use "+ Add more" for anything else the team needs, choosing either "Select another MCP" to open the Connectors directory or "+ Add custom connector". Name the project on the last screen, add a description, and finish setup. To reverse it later an owner goes to Organization settings, then Capabilities, finds "Ask Your Org" and clicks Disable, which removes it from every user account.

Then the part no administrator can do for anyone else begins. Each user opens the project, follows a guided onboarding, and authenticates individually with each service they want searched. Anthropic notes that "The more connectors you enable, the more comprehensive your search results will be", which is true and is precisely the problem: comprehensiveness is set by each individual, one OAuth flow at a time. In a 400 person company with five connected apps that is up to 2,000 separate authentications, and the ones nobody completes are invisible. The answer still arrives, confidently, drawn from a smaller corpus than the reader assumes.

Users can adjust their own coverage afterwards from "Search and tools" in the lower left of the project, enabling or disabling specific connectors, with changes applying to new conversations only.

The seat is $20 per user per month on both Team and Enterprise. On Enterprise, claude.com/pricing reads "Pricing is $20 per seat per month plus usage billed at API rates, so cost scales with the models and tasks your team runs", and Anthropic confirms that "Search queries within the project count toward your plan's standard usage limits, just like other projects and conversations." So on Team the questions come out of the plan allowance you already pay for, and on Enterprise they are a separate, metered line on the bill.

That distinction matters more for enterprise search than for ordinary chat, because retrieval is token-heavy by nature. A question that fans out to four connectors pulls the returned content into the context window before Claude writes anything, and on Enterprise those input tokens bill at API rates. Anthropic publishes those rates: at the time of writing, Sonnet 5 is $2 per million input tokens and $10 per million output, Opus 5 is $5 and $25, and Fable 5.1 is $10 and $50.

The arithmetic is worth doing once with your own numbers rather than taking anyone's published estimate, because the variable that dominates is how much retrieved content each question drags in, and that depends entirely on your corpus. As an illustration only, with assumptions stated so you can substitute your own: a question that returns 40,000 tokens of retrieved context and produces an 800 token answer on Sonnet 5 rates works out to roughly $0.088. Fifty people asking four such questions a working day would be in the region of $370 a month in usage on top of $1,000 in seats. Change the model, the corpus or the fan-out and that number moves a long way, which is the point. Anthropic does not publish a per-query price for enterprise search and neither should anyone else.

Two governance details belong in the same budget conversation. Owners can set spend limits at organization and individual level, but Anthropic is explicit that setting a limit to unlimited still bills consumption and that you cannot disable billing for usage. And on self-serve Enterprise, where usage is bought up front in credits, the balance hitting zero stops usage for the whole organization, not just the person who spent it.

No. Anthropic states that "The search project is fully functional on Claude Desktop, but not available on Claude Mobile (iOS/Android)." The web and desktop apps carry it; the phone apps do not.

Whether that matters depends on who the buyer is. For a company rolling search out to engineers and analysts at desks, it is a non-issue. For a distributed workforce, field teams, or anyone whose whole relationship with company systems runs through a phone, it removes the population that arguably needs cited answers most, because they are the ones who cannot go and read the source document themselves. It is worth checking against your actual headcount split before you assume it is a detail.

Yes, and this is where Ask Your Org is genuinely strong. Anthropic documents it as permission-aware: "You only see search results from data you have permission to access in the original systems", and separately, "Does everyone in my organization see the same search results? No."

The architecture is what makes this straightforward. Because there is no vendor-side index, there is no stored copy whose permissions could drift out of step with the source. Every call runs as the individual user, against their own credentials, so the source system itself enforces access on every request. That is the cleanest possible answer to the question a security team will ask, and it is worth conceding without qualification. Plenty of products in this category handle permissions far less carefully, which is why we treat it as its own evaluation criterion on permission-aware AI search.

Two smaller points complete the picture. Conversations inside the search project are private to the individual unless deliberately shared, the same as any other Claude conversation, though Enterprise organizations can apply custom data retention policies across all conversations including those in the project. And search results are retained with their associated chats, so deleting the conversation deletes the result.

It is the right answer when your organization already pays for Claude, your work lives in a handful of well-adopted apps, your people are at desks, and the questions you want answered are the ones an individual could have answered themselves with twenty minutes of clicking. You are buying a very good prompt over connectors you already had, at no extra cost. Set it up.

It stops being the right answer at the point where the organization needs to rely on the answer. Four documented properties combine into one problem: there is no index, so completeness depends on what live calls return inside a timeout; coverage depends on which individuals authenticated which connectors; the vendor's own best-practice guidance is to search one source at a time; and on Enterprise the cost of asking scales with how much content each question drags in. None of those are defects. They are the honest consequences of a live-call design, and Anthropic documents all four rather than hiding them.

But an organization that wants to ask "what did we agree with this customer about renewal terms" and get a complete answer, without knowing in advance which of nine systems holds the agreement and without depending on whether that particular employee once clicked through a Salesforce OAuth screen, needs a corpus that was built in advance and connected once for everybody. That is the job enterprise search software exists to do, and it is what InSearch is. The tool at the top of this page runs against a real index, connected centrally, with permissions re-checked per person at query time. You can put your own question into it now.