TheHive MCP is now generally available!

At the beginning of the year, we introduced the beta of TheHive MCP, a Model Context Protocol server that gives an LLM direct access to TheHive’s API.

Today, after several months of testing and hardening, we are excited to announce that it is officially ready to leave beta and run in your own environment!

Letting an LLM access your case data is different from asking it to answer questions in a sandbox. To make TheHive MCP production-ready, we worked on three fronts.

We hardened how the AI assistant handles your data. All user-generated fields the server returns are tagged, so your assistant can tell your data apart from instructions hidden inside it. This increases resilience to prompt injection. In addition to this, destructive actions require a confirmation step: if that step can’t complete, the action is refused. Nothing runs unconfirmed.

Before every release, we run two tests.

• Accuracy: conversations that put models through realistic investigations (searching entities, managing cases and observables, running automation) to ensure correct tool use.
• Security: we seed TheHive data (descriptions, comments, observables) with hidden instructions, from small behavior changes to critical directives like exfiltrating data or deleting evidence. To pass, the model has to both ignore the instruction and warn the analyst.

Based on these results, we recommend six models: Claude Sonnet 4.6, GPT-5.5, Kimi K2.6, GLM-5.2, Gemini 3.5 Flash and Gemini 3.1 Flash Lite.
For further details on test methodology and model-by-model results, read our evaluation document.

Last but not least, the MCP has been pentested by AI security experts to ensure it meets the highest security standards.

None of this makes TheHive MCP immune to bad behavior: it reduces the risk, it does not eliminate it. If your SOC’s first instinct is skepticism about handing an LLM access to security incidents, that instinct is correct. That’s why control stays on your side.

Resilience to malicious instructions varies by model, so make sure to use one from our recommended list. Read-only mode is enforced out of the box, so your LLM doesn’t get write access until you decide otherwise.

Finally, the permissions system gives you granular control over what tools the assistant can call, what data it can reach and which analyzers or responders it’s allowed to run.

Once you trust how the assistant behaves on your own data, open up write access and broaden permissions step by step.

That’s what we can vouch for today, and we’ll keep refining as TheHive MCP evolves.

Since the beta, TheHive MCP has picked up a few capabilities that expand what SOC analysts and managers can do with it:

• Map to MITRE ATT&CK without leaving the chat. Ask the assistant to tag an alert or a case with the matching TTP, or pull every case linked to a given pattern.
• Apply a case template or follow a playbook. Tell the assistant which template to apply or ask it to read the knowledge base page for the playbook you’re following.
• Speed up triage steps. Promote an alert to a case, merge duplicate alerts or cases: the hassles of triage, handled in a single request.
• Find related cases easily. Ask “find cases similar to this one,” and the assistant will surface related cases and alerts using TheHive’s own similarity engine, scoped to your permissions.

We have also completely revamped the documentation, to make it easier for you to navigate through all the capabilities that TheHive MCP enables.