IRMind: an AI assistant for incident response, in the terminal, with a signed record
For the past few months I have been working on a personal project that is finally at a point where I want to show it: IRMind, an open-source AI assistant for incident response and digital forensics that runs in your terminal.
The starting idea was simple: something like Claude Code, but built for forensics. A tool you can talk to while you investigate, which loads specialised methodologies when they are needed and can split work across parallel tasks. Above all, it must not compromise the things DFIR cannot compromise: the evidence and the traceability of what was done with it.
Repository: github.com/ns2switch/irmind
Why another tool
While taking SANS FOR509 (Enterprise Cloud Forensics and Incident Response), I was reminded of something I already knew from practice. A cloud investigation means jumping between CloudTrail, Entra ID, Google Cloud audit logs, Kubernetes, the EDR, server logs… Every source has its own format, its own fields and its own traps. A lot of time goes into normalising, searching and correlating rather than thinking.
Language models help a lot with that part, but putting an LLM inside an investigation raises serious problems:
- Evidence is hostile. A log can contain text crafted to manipulate the model (prompt injection). A file name, a user agent or an email subject is enough.
- The model should not decide what runs. It can propose a query or an external lookup, but something has to check that the action is within the case scope.
- You must be able to prove what happened. If a conclusion ends up in a report, you must be able to reconstruct which evidence supports it and what the tool did to get there.
IRMind exists to solve those three things.
The principle: the model proposes, policy decides, everything is signed
The basic flow is this:
Analyst → TUI → case + policy → authorised connector → local evidence
↑ ↓
model signed audit chain
- The model suggests a tool call: a timeline query, loading a skill, an external enrichment…
- Policy checks the tool, its arguments, the case scope, the execution mode and the limits. External calls need a signed analyst approval.
- Approved connectors return evidence that stays local. The model only sees a bounded copy, marked as untrusted and delimited so it cannot be mistaken for instructions.
- Every proposal, decision, result and case change goes into a hash-chained audit log signed with Ed25519. At the end, the case can be sealed, exported and independently verified.
No model is enabled by default. You can ingest, query and verify evidence without using AI at all.
What it can do today
Ingestion and normalisation
IRMind detects each source’s format automatically and activates the matching parser. Each event is normalised to a common schema based on a subset of ECS and stored in Parquet, alongside the hash of the original. Some of the supported formats:
- Cloud: AWS CloudTrail and GuardDuty, Microsoft Entra ID and Microsoft 365, Okta, Google Cloud Audit Logs.
- Kubernetes: EKS, GKE and AKS audit logs, and Falco alerts.
- Windows: EVTX, processed in an isolated, sandboxed helper process.
- Linux: auth, auditd, journald, syslog, nginx and Apache.
- DFIR tools: Velociraptor, Hayabusa, Chainsaw, plaso, EZTools, Zeek and Suricata.
- Generic: JSON, ECS, OCSF and a declarative parser for tabular formats.
There are also connectors for Elastic, Splunk, Google SecOps, CrowdStrike, SentinelOne, MISP and VirusTotal, and for pulling evidence from S3, GCS and Azure Blob. All of them are experimental and stay disabled until you configure them.
Forensic skills
Skills are packaged investigation methodologies: AWS credential compromise, S3 data exfiltration, Azure identity compromise, Kubernetes, Windows, Linux, network… Each one declares which sources it needs and where its coverage gaps are. The model loads them only when needed, so no context is spent on what is not used.
Skills never extend permissions: a skill can guide the investigation, but it cannot bypass policy. Its use is also recorded and signed in the case.
They live in a separate repository, ns2switch/irmind-skills, so anyone can propose improvements or new methodologies without writing a line of Rust.
Parallel subagents
For larger tasks, IRMind can split the work between subagents with bounded tasks. For example, one reviews identity activity while another analyses data access. Plans are visible before they run, and you follow the tasks with /subagents. You can cancel them without losing work that already finished. Subagents cannot spawn further subagents or hold more permissions than the main agent, and they share a single token and call budget that you define.
Models
It works with OpenAI-compatible providers (including local models with Ollama, LM Studio or vLLM), Anthropic, DeepSeek and OpenRouter, among others. You choose the model, and evidence that does not need to be sent is not sent.
Try it in two minutes
The quickest way to see it is the guided demo. It needs no API key, account or network:
irmind demo
The real tool runs a ten-step synthetic investigation with a scripted model: ingestion, local queries, a streaming response, a signed approval and custody verification. The temporary case and keys are deleted when you exit.
On macOS, install it with Homebrew:
brew install ns2switch/irmind/irmind
For Linux and Windows there are binaries in the releases, signed with Sigstore and shipped with provenance attestations, an SBOM and checksums. The platform guide explains how to verify them before installing. Please do not run a script downloaded from the internet without verifying it, mine included.
What it is not yet
To be clear: IRMind is pre-1.0 evaluation software and is not validated for real incidents. I have run a full investigation with a real model (DeepSeek) on Invictus IR’s public CloudTrail dataset, including skills and finding confirmation, and it works well. There are still checks pending, though, and until 1.0 every minor version may change the case format without migration.
Use it to learn, evaluate and experiment, not for your next production incident.
What is coming
- More sources: Defender for Endpoint, CrowdStrike FDR, macOS (Unified Log and eslogger), Google Workspace, the GitHub audit log, DNS, IIS and load balancers, and AWS, Azure and GCP flow logs. They come with their skills: Microsoft 365 BEC, phishing triage, Okta compromise, EDR, macOS…
- Source-side queries (v0.6): query CloudTrail in Athena, Cloud Audit Logs in BigQuery and Entra ID in Log Analytics without downloading every log. The model never writes SQL or KQL: it proposes a structured query that policy validates, with a separate scan budget. The result is fetched once, signed and analysed locally.
How to contribute
The project is free software (MIT or Apache-2.0) and is built for the community. There are several ways to help:
- Try it and open issues for whatever breaks or is missing.
- Write skills in irmind-skills. If you have an investigation methodology that works for you, that is where it makes the biggest difference.
- Add parsers or mappings. There are issues labelled
good first issuefor firewalls, proxies, DNS and other sources.
If you try it, I would love to hear what you think. And if you know a public dataset you think it should be tested against, please tell me too.
Repository: github.com/ns2switch/irmind