Blog ·
Query your audit log from Claude (MCP)
AuditKit's MCP server exposes seven tools (search_events, get_event, get_proof, verify_range, list_tenants, export_evidence, log_event) over Streamable HTTP from the same process as REST. Add it as a custom connector in Claude.ai, sign in with your email, pick the project, and ask "who changed billing settings last week, and is the chain intact?"
Install
Claude.ai: Settings → Connectors → Add custom connector → URL https://auditkit.dev/mcp → sign in with your email → pick the project → done. The server is an OAuth 2.1 authorization server (discovery at /.well-known/oauth-authorization-server, dynamic client registration, PKCE), so Claude never sees an API key; the token it gets is scoped to the project you chose.
Claude Code:
claude mcp add --transport http auditkit https://auditkit.dev/mcpAny MCP client that reads a JSON config and cannot do OAuth:
{ "mcpServers": { "auditkit": { "type": "http", "url": "https://auditkit.dev/mcp",
"headers": { "Authorization": "Bearer ak_live_…" } } } }ak_test_ and ak_live_ keys behave identically today; the prefix is a label for your own routing. Every new account gets read access to a shared Sample project with seeded events, so every read tool can be exercised before you log anything real.
The tools
| tool | title | annotations |
|---|---|---|
search_events |
Search audit events | read-only |
get_event |
Get one audit event | read-only |
get_proof |
Get tamper-evidence proof for an event | read-only |
verify_range |
Verify chain integrity for a tenant | read-only |
list_tenants |
List tenants | read-only |
export_evidence |
Export verifier-ready evidence | read-only |
log_event |
Record an audit event | writes; destructiveHint false |
Every tool declares readOnlyHint, destructiveHint, idempotentHint and openWorldHint, so an agent (or the person approving its tool calls) can tell a search from a write. There is no erase tool over MCP, on purpose. Scopes apply to OAuth tokens and API keys alike: a read-only principal gets key lacks write scope from log_event.
Same process, not a wrapper
The MCP handler is a Hono route in the same server as /v1. A fresh stateless server and transport are built per request, bound to the principal’s project, whether that is an OAuth token or an API key. Same data, same auth, same rate limits.
Example conversations
“List tenants and tell me which has the most events.” list_tenants.
“Show me every user.role.change for tenant acme in September.” search_events with action and a time range.
“Is acme’s chain intact, and how far is it anchored?” verify_range returns valid: true, the head hash, rooted_through and anchored_through.
“Give me the proof for event 01J… so I can send it to their auditor.” get_proof returns the three Merkle paths and the Rekor and OpenTimestamps receipts.
Limits
export_evidence truncates at 2000 lines over MCP; use GET /v1/export for full files. Search returns at most 200 events per call over MCP.