Anthropic Tool Use, with scope and audit attached.
Wrap Claude tool definitions with the live OC Agent delegation. Tool execution is gated by the active scope grammar pre-call; every tool result emits a signed agent-action envelope. Anchored to Bitcoin via the OC Stamp pipeline.
What it does
Claude's tool-use API exposes a list of tools and lets the model produce a sequence of tool_use blocks. The Console adapter intercepts those blocks at the boundary between Claude and your real tool implementation:
- Validates the tool call against the agent's active OC Agent delegation. The exercised scope must be a sub-scope of a granted scope; if not, the call is refused before your handler runs.
- Canonicalizes the (tool_name, input) tuple, BIP-322 signs it via the agent's wallet adapter, and emits an agent-action envelope (kind 30084) sharing OC Stamp's structure.
- Returns the tool result to Claude alongside the envelope id, so the assistant's next turn carries provenance metadata if you choose to surface it.
Why this matters for compliance
Without the adapter, “the model called my tool” is the entire audit trail. With it, every tool execution carries: the agent address (BIP-322), the active delegation, the exercised scope, the content hash of the inputs, and the OC Stamp anchor — replayable offline forever.
Status
Adapter package: targeted Q3 2026 in oc-packages. Will be a thin wrapper around @orangecheck/agent-core + @orangecheck/agent-signer with a Claude-tool-use shape.
What design partners can have today: a hand- rolled wrapper using the same primitives the MCP adapter uses. We'll pair an OrangeCheck engineer for a half day to wire it into your existing Anthropic integration.
Want this adapter sooner?
Adapter prioritization is design-partner-driven. Tell us your stack, the volume, and the compliance bar — adapters that named partners need move from "planned" to "shipping" inside a quarter.