BLAZLE Documentation

MCP Copilot

BLAZLE MCP Copilot is a full Model Context Protocol server: your agents and IDEs (Cursor, Claude, custom swarms) connect to BLAZLE and operate its security surface — but every action they take is governed. State-changing operations are never executed directly; they are proposed and held for human approval.

All three MCP primitives

Tools — 11 tools covering rules, incidents, audit forensics, incident response, and policy control. Mutating tools require HITL approval.

Resources — read-only tenant data: blazle://rules, blazle://incidents, blazle://audit.

Prompts — curated security prompts: triage_incident, harden_rules, explain_finding.

Tool Reference

Tool	Risk	Permission	Description
list_interceptor_rules	ReadOnly	ViewAudit	List all custom interceptor rules for the org.
add_interceptor_rule	Mutating	ManageRules	Create a new block/redact/log rule (regex).
toggle_interceptor_rule	Mutating	ManageRules	Enable or disable a rule. Cache-busted immediately.
delete_interceptor_rule	Destructive	ManageRules	Permanently delete a rule. Requires four-eyes.
list_recent_incidents	ReadOnly	ManageIncidents	List recent security incidents.
verify_audit_chain	ReadOnly	ViewAudit	Validate cryptographic hash-chain integrity of audit logs.
export_cef_logs	ReadOnly	ViewAudit	Export audit events as ArcSight CEF for SIEM ingestion.
revoke_api_key	Destructive	ManageApiKeys	Immediately revoke a compromised API key. Four-eyes.
suspend_webhook	Mutating	ManageWebhooks	Pause webhook delivery by clearing the endpoint URL.
update_tenant_policy	Mutating	ManagePolicies	Change enforcement_mode or min_block_rating.
rotate_provider_credentials	Mutating	ManageOrg	Evict LLM provider key cache to force re-fetch from vault.

Governed by design (HITL)

Every tool call — from the in-app Copilot or an external MCP client — passes the same control chain:

1. RBAC scope — the caller must hold the tool's permission.

2. Dogfood interceptor — the call's arguments are scanned by BLAZLE's own in-flight interceptor; secrets/PII are blocked.

3. Human-in-the-loop — read-only runs immediately; mutating actions wait for approval; destructive actions require four-eyes (approver ≠ requester).

4. Instant effect — rule mutations (create/toggle/delete) bust the Redis interceptor cache immediately so changes apply to the very next request.

5. Audit — proposal, decision and execution are written to the tamper-evident hash-chain.

Connect an external agent

1. Create a scoped access token in the dashboard under MCP Access. Its role is bounded below your own and can be revoked anytime.

2. Point your MCP client at the streamable-HTTP endpoint with the token:

{
  "mcpServers": {
    "blazle": {
      "url": "https://api.blazle.io/api/v1/copilot/mcp",
      "headers": { "Authorization": "Bearer mcp_your_token_here" }
    }
  }
}

The endpoint speaks JSON-RPC 2.0 over Streamable HTTP: POST returns JSON or an SSE stream (per your Accept header), and GET opens the server→client SSE channel.

Example prompts

Audit integrity check

"Verify the audit chain integrity"

SIEM export

"Export the last 50 events in CEF format"

Security hardening

"Switch enforcement mode to strict"

Incident response

"Revoke API key #5 — it has been compromised"

Rule management

"Show me all interceptor rules"

Try it via curl

curl -X POST https://api.blazle.io/api/v1/copilot/mcp \
  -H "Authorization: Bearer mcp_your_token" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

All docs