Get Started with AI¶
GRC-ITSM is built to be operated by AI assistants. Setup runs in three steps, in order:
-
Add the OAuth connector to claude.ai and install the plugin marketplace.
-
Build the foundation your SSP or Security Decision Record renders from, through documents, interviews, or both.
-
Prepare and close the recurring Operations Meeting with the plugin's guided commands.
Quick Summary
- The platform exposes an MCP endpoint that AI assistants connect to. Each member signs in through OAuth, so AI actions run under their own agent identity and are attributed and auditable like any other agent activity.
- AI can read and operate across the platform: tickets and service desk work, assets and inventory, security documentation, compliance data, continuous monitoring, reporting, and the knowledge base.
- The Stratus GRCITSM marketplace adds guided workflows on top. They cover authoring and generating your SSP or Security Decision Record, closing its gaps, adopting the templated policy library, running projects and supporting the recurring Operations Meeting, answering FedRAMP and CMMC questions with citations, and applying bulk changes safely.
How AI interacts with the platform¶
Claude connects to your instance through the Model Context Protocol (MCP) -- an open standard that lets an AI assistant call a platform's functions the same way an integration would. It works inside the same governed workflows as any human agent: the ticket is still the audit trail, the approval is still the evidence, and every action is attributed to a named identity.
Your instance exposes its capabilities to AI through MCP. When Claude is connected, it sees a catalog of platform operations -- lookups, searches, reports, and record updates -- and calls them on your behalf as you work in a conversation.
Three properties keep this governed:
| Property | What it means |
|---|---|
| Your identity, not a shared one | Each member signs in to the instance through OAuth, so AI acts as that person's own agent account -- never anonymously, and never through a shared credential. The MCP Access role gates which agents can use the MCP tooling. |
| Same rules as any agent | AI calls go through the platform's service layer, so ticket workflows, field validation, approval processes, and SLAs apply exactly as they do in the UI. |
| Attributed audit trail | Every record an AI assistant touches shows the acting agent identity, so AI-assisted work is distinguishable and reviewable after the fact. |
Reads and writes are treated differently in practice: lookups and reports are safe to let AI run freely, while anything that changes a record requires your explicit approval before it executes. The recommended Claude configuration enforces exactly that split.
Related documentation
- Agent Roles -- the MCP Access role that grants agents service access for AI integration
- Reporting -- the MCP report variants that back the platform's AI integration
- Change Management Policy -- introducing a new AI capability that touches federal data is a transformative change
- Operations Meetings -- the recurring meeting that runs a certified system from its GRC-ITSM records, which AI helps prepare and close (how it works)
What AI can do¶
The MCP catalog spans the whole platform. Rather than a list of individual functions, think of it as capability areas:
| Area | What AI can do |
|---|---|
| Service desk & tickets | Search, summarize, create, and update tickets across every ticket type; triage queues; set priorities and SLAs; add notes; manage ticket relationships |
| Assets & inventory | Query the asset inventory, create and update assets, and maintain the relationships that define your authorization boundary |
| Security documentation | Read and author the OSCAL-aligned foundation the platform holds -- system information, components, control implementations, implemented requirements, and implementation narratives -- that your security documentation renders from |
| Controls & requirements | Look up control baselines, control details, parameters, and requirement mappings for the frameworks your system is certified against |
| Continuous monitoring | Query open and closed findings, vulnerability details and history, PAIN Events, deviations, accepted risks, significant changes, Transformative Change Notice Windows, Trust Center agency access, and Key Security Indicator status |
| Projects & operations | Create and structure Projects and Project Tasks, set and read budgets, log time, and post outcome notes to the tickets a decision concerns -- the records an Operations Meeting works from and writes back to |
| Reporting & knowledge base | Run platform reports, search and author knowledge base articles, and pull the data behind your Trust Center reporting |
Notices, evaluations, and the human step¶
AI can prepare incident and change records, but the FedRAMP notices themselves stay with a person.
- Send notice actions have no AI or MCP tool. The Send actions for incident reports (initial, ongoing, final) and significant change notices open a templated email in preview, and a person presses them in the UI. Nothing moves the ticket off a notice-due step except that press. See FedRAMP Notices.
- The record incident report tools fill fields only. Recording an initial, ongoing, or final incident report fills the report content and never moves the workflow step.
- Evaluations must leave a ledger row. An evaluation fills PAIN Events (on an Issue) or Incident Events (on an Incident) with a short plain-English note. For an Incident, the reportability reason opens with
Reportability (IEC-CSO-EFR):. - Tools exist for recording PAIN events, completing an After Action Report and requesting its approval, the Transformative Change Notice Windows report, and the Trust Center Agency Access report.
One foundation, two artifacts
Under the FedRAMP Consolidated Rules for 2026, the Security Decision Record replaced the traditional System Security Plan with a persistently maintained record of the security decisions made across a service's lifecycle. This holds on both tracks, not just 20x: a Rev5 certification under the 2026 rules produces an SDR too, carrying per-control summaries where a 20x SDR carries Key Security Indicator summaries.
Whichever your package requires, it renders from the same asset foundation the platform holds -- components, implemented requirements, and implementation statements, modeled on OSCAL -- so authoring that foundation serves both. Machine-readable output is produced as schema-valid FedRAMP JSON. See Certification Package & SDR Maintenance for how the SDR is produced and kept current.
On top of the raw catalog, the grcitsm-analyst plugin (from the Stratus GRCITSM marketplace; see Install the plugin marketplace in Claude) adds guided multi-step workflows over that foundation:
- Onboarding -- build the full system definition from scratch through documents, interviews, or both, assembled locally and written in one approved pass (how it works)
- Generation -- render the security documentation from current platform state
- Ingestion -- turn policies, procedures, and interview transcripts into documentation content, with row-level approval before anything is written
- Interviewing -- author content from scratch through conversation
- Gap remediation -- after onboarding, a gap register ranks what the documentation still lacks, per baseline. Gaps are grouped into named remediation actions, each a task on the onboarding Project, with the evidence that closes it named up front
- Policy library adoption -- fetch the current templated policies and procedures from this site and adapt the organization-defined values through a short interview. Then import each as a knowledge base article bound to the component it governs, so the documents also feed the SSP or SDR as evidence
- Project management -- create and structure Projects and Project Tasks, set budgets, attribute time, and read budget burn-down, following the platform's project conventions
- Ticket updates -- priority, SLA, dates, and custom fields, each change carrying a note, one ticket or a batch at a time
- Webhook integration help -- compose and troubleshoot the payloads your scanners and alert sources post to the platform's incoming webhooks. The assistant drafts them; it never posts one
- Disciplined bulk writes and reads -- plan any large change to a reviewable file first, then apply it step by step; sweep large read sets through a cache instead of one call at a time
Running a system week to week through the recurring Operations Meeting uses these same building blocks, and the plugin ships guided commands for each step of the cycle. See Run Your System for the command set.
Ask questions, not just operate¶
The marketplace is also a research assistant for the compliance and platform knowledge around your system:
- FedRAMP and CMMC questions, answered with citations. The plugin bundles the FedRAMP 2026 Consolidated Rules, the FedRAMP 2026 narrative corpus, the Rev5 baselines, and the CMMC reference corpus. Ask "what does AC-2 require?", "which controls map to KSI-IAM-AAM?", or "what does CMMC AC.L2-3.1.1 require?" and get answers grounded in the requirement text rather than the model's memory. Requirement answers are routed through an accuracy-review gate before they reach you.
- What your automated checks cover. The plugin carries the catalog of the Steampipe/Powerpipe compliance checks -- which Key Security Indicators each cloud provider's checks cover, which indicators have no automated check and need manual evidence, and how the compliance exemption tags work.
- Learning GRC-ITSM itself. This documentation site is indexed for AI lookup, so Claude can answer "how do I configure email rules?" or "what is the User Access Request workflow?" by reading these docs directly.
- Underlying Halo platform questions. For product details of the underlying Halo platform, we recommend the Context7 documentation index, which indexes the Halo guide library in AI-readable form.
Guardrails and best practices¶
AI with write access to a live system of record deserves the same discipline as any privileged integration:
- Attributable identity. Members authenticate as themselves, so every AI action carries a real name. Grant the MCP Access role deliberately -- it is the lever for who can use AI against the tenant, and removing it (or deactivating the Claude MCP OAuth application) revokes access.
- Writes get human review. Configure Claude so that record-changing operations always prompt for approval. Reads are safe to run freely; changes are not.
- Least privilege. Scope the OAuth application and agents' roles to the permissions the work actually needs, and widen them only when a workflow demands it.
- Non-production first. Trial new AI workflows against a UAT or development instance before pointing them at production.
- Protect credentials. Treat the client secret like any privileged credential: rotate it with Generate when in doubt (the old secret stops working immediately), and never put it in version control or chat messages.
- Keep sessions focused. Long AI sessions accumulate context and drift; start fresh sessions for distinct pieces of work.
Changes are yours to approve
An AI assistant proposes and executes -- it does not own the outcome. Keep write operations behind an approval prompt, review what is about to change before confirming, and treat anything AI wrote to a record as work product to verify, exactly as you would a new team member's.
Additional details
- Best Practices for Claude Code in Halo -- the platform vendor's risk and mitigation guidance for agentic AI tooling
- Claude Code security documentation -- Anthropic's reference on permissions and safeguards