What is AgentOps
AgentOps is a policy firewall and tamper-proof audit anchor for AI agents that move money. When your agent wants to pay for something or take an on-chain action, it asks AgentOps first. AgentOps checks the rules you set, signs only if they pass, records the decision, and anchors a cryptographic proof of it on the Casper blockchain — permanently.
The two halves of the product
- The console — where you fund a Casper float, set spend policy, issue agent keys, watch decisions land live, and export judge-verifiable audit evidence. No code required.
- The integration surface — how your agent asks for permission at runtime. Two doorways into the same guard: an MCP server (for Claude Code and other LLM agent frameworks) and the
@agops-labs/sdknpm package (for TypeScript backends). Same firewall, same signing, same anchor.
NOTE
The SDK is a client, not the product. Everything it does is also doable from the console or over plain HTTP — it exists so you don’t hand-write requests.
How a request flows
- Your agent hits a paid API and gets a
402 Payment Required— or wants to swap on CSPR.trade, or submit a deploy. - Instead of paying, it asks AgentOps to authorize the action.
- AgentOps runs the policy firewall: kill-switch, rail, service allowlist, per-action max, spend cap, velocity limit.
- On ALLOW it signs and returns a payment header; on DENY it returns a plain-English reason and nothing is signed.
- Either way the decision is recorded, and its SHA-256 fingerprint is anchored on-chain in the
GuardRegistrycontract.
Your agent never holds the signing key. That is what makes the caps binding — there is nothing for the agent to route around.
Who this is for
- Teams running agents that spend — you need a hard ceiling on autonomous spend and a record you can defend.
- Developers integrating an agent — you want one call between your agent and any payment, with typed results.
- Operators and finance — you need reconciliation, statements, and evidence that survives a dispute.
Where to go next
- New here? Start with the Quickstart — account to first enforced decision.
- Want the vocabulary first? Read Core Concepts.
- Ready to wire up an agent? See Choosing an Integration, then Connect Claude Code or the SDK Quickstart.
- Looking for exact package names and versions? Packages & Versions.
