SIEMen

Security-focused AI data layer on Cloudflare Workers. Purpose-built for AI security agents — combines fast KV scratch cache, semantic vector search over findings, per-engagement agent memory, and a similarity-based LLM triage cache. Exposes both a REST API and a built-in MCP server so Claude/AI agents can read, write, and search security data directly.

Stack

Related MCP server: Memory MCP Worker

What It Does

Fast KV Scratch Cache

Per-engagement key/value store for in-flight agent notes, intermediate results, and shared state across agent steps. Namespaced to prevent cross-engagement leakage. Optional TTL (max 24h).

Semantic Vector Search

Embed and store security findings, CVEs, and controls with finding_store. Retrieve the closest matches by semantic similarity with finding_search. Namespace-scoped to engagement — an agent searching for "SQL injection" can only surface findings from its own engagement.

Per-Engagement Agent Memory

Agents store notes with engagement_remember and recall them by recency or vector similarity with engagement_recall. Supports sub-scoping by session_id for multi-step workflows.

Semantic LLM Triage Cache

semantic_triage routes security prompts through a two-level cache before hitting an LLM:

1. Exact-hash short-circuit — identical prompts reuse prior analysis instantly

2. Cosine similarity check — prompts above the configured threshold (default 0.92) return a cached response without an LLM call

3. Miss — calls the configured LLM gateway, writes result to KV + Vectorize for future hits

Cuts LLM spend significantly when agents analyze structurally similar findings across engagements.

MCP Tools

REST Routes

All /v1/* routes require Authorization: Bearer <SIEMEN_API_KEY>.

D1 Schema

Apply migrations in order with wrangler d1 execute siemen-db --remote --file migrations/<file>.sql:

• engagements — engagement registry (id, name, client, status, timestamps)

• memory — agent memory per engagement (content, tags, vector_id, session_id)

• findings — security findings/CVEs/controls (kind, title, body, severity, asset, external_id, vector_id)

• semantic_cache_log — cache outcome log (prompt_hash, outcome, similarity, model, tokens, created_at)

Deploy

# 1. Create infrastructure
wrangler vectorize create siemen-vectors --dimensions=768 --metric=cosine
wrangler d1 create siemen-db
wrangler kv namespace create siemen-sc-cache
wrangler kv namespace create siemen-kv

# 2. Update wrangler.toml with the returned IDs

# 3. Apply D1 migrations (run in order)
wrangler d1 execute siemen-db --remote --file migrations/migration-001-engagements.sql
wrangler d1 execute siemen-db --remote --file migrations/migration-002-memory.sql
wrangler d1 execute siemen-db --remote --file migrations/migration-003-findings.sql
wrangler d1 execute siemen-db --remote --file migrations/migration-004-semantic-cache-log.sql

# 4. Set secrets
wrangler secret put SIEMEN_API_KEY
wrangler secret put OPENROUTER_API_KEY
wrangler secret put LLM_GATEWAY_KEY

# 5. Deploy
npm run deploy

MCP Configuration

Add to your Claude Desktop / Claude Code MCP config:

{
  "mcpServers": {
    "siemen": {
      "url": "https://your-worker.workers.dev/mcp",
      "headers": {
        "Authorization": "Bearer ${SIEMEN_API_KEY}"
      },
      "alwaysAllow": [
        "engagement_open",
        "sec_cache_get",
        "finding_search",
        "engagement_recall",
        "cache_stats"
      ]
    }
  }
}

Set SIEMEN_API_KEY in your shell environment. Write tools (finding_store, sec_cache_set, engagement_remember, semantic_triage) require explicit approval by default.

Security

• All /v1/* and /mcp routes require a Bearer token verified with Web Crypto SHA-256 (no Node.js crypto — edge-compatible)

• Findings are namespace-scoped to engagement_id in Vectorize — cross-engagement leakage is impossible at the query layer

• Secrets via wrangler secret put only — never in source or wrangler.toml

• Rate limited at 60 req/min per token identity (not per IP)

See SECURITY.md for vulnerability reporting.

License

MIT