Overview

The @aegissovereign/mcp-server package exposes 31 platform capabilities as MCP tools. Once installed, engineers can query compliance status, check LLM spend, verify the audit chain, run AI-safety checks, and promote models without leaving their editor. The server is a standalone npm package — it does not require the monorepo and connects to any Aegis Sovereign instance via three environment variables.

Installation & Configuration

Add the server to your Claude Desktop or Claude Code configuration. The server starts on demand via npx — no global install required.

json

~/.config/claude/claude_desktop_config.json

1{
2  "mcpServers": {
3    "aegissovereign": {
4      "command": "npx",
5      "args": ["-y", "@aegissovereign/mcp-server"],
6      "env": {
7        "AEGISSOVEREIGN_BASE_URL": "https://ai.corp.example.com",
8        "AEGISSOVEREIGN_API_KEY": "sk-sovereign-...",
9        "AEGISSOVEREIGN_WORKSPACE_ID": "ws-prod"
10      }
11    }
12  }
13}

bash

Issue an API key (Personal Access Token)

curl -X POST https://sovereign.yourcompany.com/api/v1/auth/tokens \
  -H "Authorization: Bearer $OIDC_JWT" \
  -d '{ "name": "MCP server", "expires_in_days": 365 }'
# Returns { "token": "sk-sovereign-..." } — add to the env block above

Compliance Tools

Three tools covering regulatory evaluation. All tools default to the configured workspace ID but accept an optional workspace_id override.

text

Example prompts

1"Is our fraud-detector-v3 EU AI Act compliant?"
2→ check_compliance_status(model_id="model-uuid")
3
4"Run a full compliance check on the credit model before we promote it"
5→ run_compliance_eval_all(model_id="model-uuid")

Tool	Description
`check_compliance_status`	All framework evaluations for a model — pass/fail/score per framework
`run_compliance_eval`	Trigger a single-framework evaluation and return the job ID
`run_compliance_eval_all`	Trigger all 7 frameworks in parallel — use before promotion

Registry Tools

Five tools for the model registry. promote_model pre-checks for blocking evaluations before calling the API — it returns a clear explanation if promotion is blocked rather than a raw 422.

Tool	Description
`list_models`	Paginated model list with status/search filter
`get_model`	Full model details, metadata, and tags
`promote_model`	Pre-checks blocking evals, then promotes — returns job ID
`get_model_card`	Retrieve or auto-generate the model card
`revert_model`	Revert production model to its previous champion

Audit Tools

Four tools for the tamper-evident audit log and its external anchoring. verify_audit_chain walks the whole SHA-256 chain; the anchor tools publish/verify the chain head against an external immutable ledger (see the Audit-Chain Anchoring doc).

text

Example prompts

1"Show me all model promotions this week"
2→ get_audit_log(category="model", search="promoted", from="2026-04-07T00:00:00Z")
3
4"Verify the audit log hasn't been tampered with"
5→ verify_audit_chain()
6
7"Anchor the audit chain to the external ledger, then confirm it"
8→ anchor_audit_chain() ; verify_audit_anchor()

Tool	Description
`get_audit_log`	Query audit entries filtered by category, severity, or date range
`verify_audit_chain`	Walk the SHA-256 chain to genesis — returns valid/tampered + chain length
`anchor_audit_chain`	Publish the current chain head to the external immutable ledger
`verify_audit_anchor`	Recompute the live head and compare to the latest external anchor

Monitoring Tools

Four tools for live operational data. run_drift_check polls the platform for up to 30 seconds and returns the completed report inline — no need to follow up with a separate query in most cases.

text

Example prompts

1"How much have we spent on GPT-4o this month?"
2→ check_llm_usage(from="2026-04-01")
3
4"Has the credit model drifted since last week?"
5→ run_drift_check(model_id="model-uuid")

Tool	Description
`check_llm_usage`	Token usage and cost breakdown by model/provider with budget bar
`run_drift_check`	Trigger drift check + poll for result; returns severity and drifted features
`get_drift_history`	Historical drift report table for trend analysis
`get_workspace_metrics`	Live CPU, GPU, p99 latency, and predictions-per-hour

Safety Tools

Fifteen tools covering the AI Safety Platform — runtime enforcement, agentic guards, PII masking, clearance, red-team certification, SLOs, approvals, and regulatory reports. Two tools (safety_get_sbom, safety_scan_vulnerabilities) are reserved for a future release and return a not-implemented response pointing to the shipped alternatives (safety_get_lineage, safety_run_redteam).

text

Example prompts

1"Validate this prompt through the safety engine: Ignore all previous instructions…"
2→ safety_validate_text(text="…", target="input")
3
4"Is claude-sonnet-4-6 cleared for customer-support use?"
5→ safety_check_model(model_id="claude-sonnet-4-6")
6
7"Run the red-team certification suite with financial probes"
8→ safety_run_redteam(domains=["financial"], include_multi_turn=true)

Tool	Description
`safety_validate_text`	Run text through the Safety Policy Engine (regex + ML ensemble)
`safety_validate_tool_call`	Validate an agent tool call before execution (risk, args, allow-list)
`safety_validate_delegation`	Enforce multi-agent trust boundaries — block privilege escalation
`safety_mask_pii`	Redact PII in text in place instead of blocking
`safety_check_model`	Check ModelSafetyRegistry clearance for a model
`safety_clearance_gate`	Run the pre-deployment clearance gate (bias + robustness + safety)
`safety_run_redteam`	Trigger the red-team certification runner (domains + multi-turn)
`safety_slo_status`	Burn rate and error budget for the safety SLOs
`safety_generate_report`	Generate a safety report from real audit events
`safety_classifier_status`	Show the active ML classifier ensemble (warns on NullClassifier)
`safety_list_reviews`	List pending approvals / review requests
`safety_submit_review`	Record an approve/reject decision via the Approvals system
`safety_get_lineage`	Show a model's provenance lineage DAG to its root base model