Drop AgentCache in front of any LLM workload. Cut token spend, put a hard budget and kill-switch around autonomous agents, and let them resume reasoning across runs — with every dollar saved measured, not estimated.
Add one server to your MCP config. Your agent gains caching plus a pre-spend budget gate and cross-run memory — no code changes.
// claude_desktop_config.json / cursor mcp.json { "mcpServers": { "agentcache": { "command": "npx", "args": ["-y", "agentcache-mcp"], "env": { "AGENTCACHE_API_KEY": "ac_live_…" } } } }
Check the cache before you call a model; store the answer after. Namespaced per tenant, priced automatically.
# 1 — is it cached? (a miss is 200 hit:false, never an error) curl https://agentcache.ai/api/cache/get?key=$HASH \ -H "Authorization: Bearer ac_live_…" \ -H "X-Cache-Namespace: my-app" \ -H "X-Model: claude-opus-5" # 2 — store the model's answer for next time curl -X POST https://agentcache.ai/api/cache/set \ -H "Authorization: Bearer ac_live_…" \ -H "X-Cache-Namespace: my-app" \ -d '{"key":"'$HASH'","value":{…}}'
ac_live_ key from your dashboard.OpenRouter and Stripe tell you what you spent. agentcache_gate tells your agent whether a call is allowed before the money leaves — against a budget, a request quota, a spend-anomaly guard, and a kill-switch. That is the difference between billing an autonomous agent and governing one.
ac_live_ key and the control-plane endpoints deployed (/api/cache/*, /api/governance/*, /api/analytics/savings). The MCP package name agentcache-mcp is the intended public name for src/mcp/server.ts; publish it to npm to make the one-line install real.