Wire an MCP math server into your agent

They let an agent route math work to MathKernel, which planned the request and delegated computation to several theorem and numeric engines.

Done with Custom agentยท Advanced

What they did
They ran MathKernel as an MCP server over stdio with 162 math_ tools. The agent first found out what the server could do, then turned the problem into a context and sent it to math_reason for solving and checking. For extended searches, the system used async jobs, and each answer came back with a derivation path and a trust label.
How it went
No performance figures were published. The design notes say the trust label is limited by the weakest supporting evidence, and any decimal input downgrades the output to numeric form, which shuts off Lean and SMT proof routes.
Tips
The first launch installs Lean 4 and Mathlib unless MATHKERNEL_SKIP_LEAN_INSTALL=1 is configured. For GPU use, install the extra nvidia pip packages; without them, MathKernel stays on CPU.

Try it yourself with Claude Code

Use the MCP math server to work on [problem]. Begin by listing the math tools that are available, then decide which engine or combination of engines fits the task. Build a context, solve the problem, and run a separate verification step when it makes sense. Return the derivation trail and trust label with the answer, and stop for my approval before any irreversible step such as submitting, sending, deleting, or paying. If the input contains decimals, tell me that the result will be numeric rather than proof-oriented.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on github.com โ†—

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.