Wolfram

Name: Wolfram MCP Server
Availability: InStock
Author: Wolfram

Data & Research3 toolsNo Auth Required

Wolfram's hosted MCP server exposes 3 tools with no sign-in: a Wolfram|Alpha query, a documentation search, and a Wolfram Language evaluator that runs code in a stateless cloud kernel. Verified live 2026-08-18.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on 18 August 2026.

Anthropic states this reflects the level of review a connector received, not a security audit.

Connect Wolfram via MCP

https://agenttools.wolfram.com/mcp

Works in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.

Wolfram Tools & Capabilities (3)

WolframContextRead-only

Uses semantic search to retrieve any relevant information from Wolfram. Always use this tool at the start of new conversations or if the topic changes to ensure you have up-to-date relevant information. This uses semantic search, so the context argument should be written in natural language (not a search query) and contain as much detail as possible (up to 250 words).

WolframLanguageEvaluatorRead-only

Evaluates Wolfram Language code for the user in a Wolfram Language kernel. If a formatted result is provided as a markdown link, use that in your response instead of typing out the output. Parse natural language input with `\[FreeformPrompt]["query"]`, which is analogous to ctrl+= input in notebooks. Natural language input is parsed before evaluation, so it works like macro expansion. You should ALWAYS use this natural language input to obtain things like `Quantity`, `DateObject`, `Entity`, etc. This is a stateless kernel, so you cannot reuse definitions from previous evaluations.

WolframAlphaRead-only

Use natural language queries with Wolfram|Alpha to get up-to-date computational results about entities in chemistry, physics, geography, history, art, astronomy, and more.

Read from the server on 2026-08-18, including each tool's own safety annotations.

Limits

One tool runs arbitrary code, and the annotations do not distinguish it. All three tools are readOnlyHint: true, destructiveHint: false. Only Wolfram's prose separates the evaluator from the two search tools, and it does so clearly. Grade the risk by the code argument, not the flag.
No published sandbox statement for this endpoint. Wolfram documents a kernel sandbox, but only in its Enterprise Private Cloud guide, which calls it "configurable, up to and including disabling it entirely". Nothing states which configuration agenttools.wolfram.com uses, and the AgentTools repository documents the *absence* of sandboxing for customer-deployed cloud servers. We tested nothing and claim nothing beyond that.
The terms restrict AI-product use. Wolfram's Terms of Use state the Services should not be used alongside your own AI-powered tools or services without a separate licensing agreement, and prohibit AI training on Services data and bulk downloading. Free of charge is not the same as free to embed.
Nothing persists. Stateless kernel, no sessions, no reusable definitions. Multi-step work has to be reassembled into a single code string on every call, which gets awkward fast.
No file access, in either direction. Wolfram states the hosted service supports neither upload nor download. Anything needing local files is directed to the Local MCP product instead.
60-second default time limit. Per Wolfram's tool documentation the default timeConstraint is 60 seconds. Long computations must be raised per call or broken into steps.
WolframContext is degraded without a subscription. Wolfram's repository states that without LLMKit it provides no reranking or filtering and returns no Wolfram|Alpha results — so on the free endpoint it is a documentation search, not a knowledge search.
No numeric quota is published. Wolfram says free for limited personal use and small-scale, casual use. It publishes no call count, rate limit or 429 behaviour for this endpoint. The Wolfram|Alpha developer API's published 2,000 non-commercial calls per month is a *different* product's limit and should not be read onto this one.
No prompts and no resources. The initialize capability block declares neither. Both list methods return empty arrays rather than an error, but the capability block is authoritative.
Protocol negotiates down to 2025-03-26. We requested 2025-11-25 and 2025-06-18; both returned 2025-03-26. Clients depending on newer protocol features will not get them here.
An unknown tool name returns a raw HTTP 500. Rather than a JSON-RPC error, the server returned a Wolfram Cloud internal-error body including an internal source-file reference. It is an ugly failure mode, not a dangerous one, but do not expect well-formed protocol errors.
We never called the evaluator with real code. Our check was initialize, tools/list, prompts/list, resources/list, resources/templates/list and three deliberately empty tools/call requests made to observe the auth gate. Every behavioural claim above comes from Wolfram's documentation or the tool schemas.

Frequently asked questions

Does the Wolfram MCP server require an account or API key?

No, not at this endpoint. We opened an anonymous handshake on 2026-08-18 and the server returned all three tool definitions without a credential. We then called each tool with empty arguments and got a parameter validation error rather than an authentication challenge, which means the request cleared any gate. Wolfram describes the hosted service as free for limited personal use.

Is WolframLanguageEvaluator safe to auto-approve?

Wolfram's own setup guide says no. It recommends always allowing the Wolfram|Alpha and Wolfram Context tools while setting the Wolfram Language Evaluator to require approval. All three tools carry identical readOnlyHint true annotations, so the annotation does not make that distinction and the vendor's written guidance does. Treat code evaluation as code evaluation.

What is the difference between WolframAlpha and WolframContext?

One answers questions, the other finds syntax. Wolfram's agent-facing documentation describes WolframAlpha as natural language queries for direct factual lookups, and WolframContext as semantic search of Wolfram documentation used to locate the right function before calling the evaluator. WolframContext returns reference material about Wolfram Language, not computed answers about the world.

Does the Wolfram MCP kernel remember anything between calls?

No. The tool description states this is a stateless kernel and that definitions from previous evaluations cannot be reused. Wolfram's own product comparison labels the hosted Cloud MCP as one-shot stateless and says it does not support sessions. We confirmed the transport behaves that way, answering tools/list with no session identifier and with an invalid one alike.

Can the Wolfram MCP server read or write files on my machine?

No. The connector is a remote HTTP service, so code runs in Wolfram's cloud rather than on your computer. Wolfram's comparison table states the hosted Cloud MCP does not support external files and supports neither upload nor download, and lists local file access as a Local MCP feature instead. Nothing in the three schemas accepts a file path.

Why are all three Wolfram tools marked idempotentHint false?

Because repeat calls are not guaranteed to match. Many Wolfram computations are deterministic, but the same tools also answer questions about current time, live financial and astronomical data, and random sampling. A query about today returns a different result tomorrow. The annotation is set for the tool as a whole, so the weakest case governs all three.

Does connecting Wolfram to Claude cost anything?

No, and no key is needed, but the licence matters more than the price. Wolfram calls the hosted service free for limited personal use and publishes no numeric call quota. Its general Terms of Use separately state the Services should not be used alongside your own AI-powered tools or services unless a separate licensing agreement is in place, which affects commercial embedding.