Mistral's tool-call patent: agent plumbing is now IP

Most agent frameworks converged on the same trick in the last eighteen months: instead of having the model emit one tool call at a time, have it write a small code block that calls several tools, then run that block in a sandbox. It felt like a community pattern. Mistral AI holds a granted US patent on it. The Mistral tool-call patent — US 12,670,045, "Code Implemented Tool Calls" — issued on June 30, 2026, and developers only started circulating it this week.

What actually happened

The patent record is short and specific. Application 19/557,103 was filed on March 4, 2026 and granted on June 30, 2026 — 118 days from filing to issue, against a US average measured in years. The inventor is listed as Gabriel Vergnaud; the assignee is Mistral AI.

The described method is the one you have probably already built. An LLM generates a code block in a programming language that encapsulates one or more tool calls. A server executes that block in a sandbox. When execution hits a tool call that has to run on the client, it pauses, ships the call to the client, waits for the result, substitutes that result back into the running code block, and resumes. Only the final output goes back to the model — not the intermediate results.

If you have wired MCP servers behind a code interpreter, or let a coding agent orchestrate several API calls inside one Python block to keep tokens down, that is the shape.

We have not seen Mistral assert this against anyone, and a granted claim is not the same as an enforceable one — patents get narrowed and invalidated all the time. We are not your lawyers. The point is narrower than a threat: a pattern you assumed was public infrastructure now has an owner on file.

Why an agent patent matters for your business

You are not going to get sued for running Claude Code. Patent risk in this layer sits with the companies shipping agent runtimes, not the shops using them. What changes for you is the durability question we keep coming back to.

The agent plumbing under your automations — how tools get called, batched, sandboxed, resumed — is turning into contested territory. That means the runtime under your workflows can change for reasons that have nothing to do with your product: a licensing deal, a rewrite to design around a claim, a vendor deciding a pattern is too expensive to keep. You want to be able to survive that without rebuilding your business logic.

The defense is boring and it works. Keep your tool definitions in your own repo, described in plain schemas, not buried in a vendor's proprietary config. Keep the orchestration layer thin enough that swapping how tools get invoked is a day of work, not a quarter. We made the same argument when agent plugins got a portable spec — the value is in your tools and your data, not in whose sandbox executes them.

Sources: US Patent 12,670,045 record, Mistral AI tool-calling documentation.