Learn-KeeperHub

A working journal of building agent-driven workflows on K — the execution and reliability layer for AI agents operating onchain.

This repo is the result of an end-to-end engagement: read the docs, wire up an agent, deploy real workflows against a real Safe multisig on Ethereum mainnet, and write down what worked, what broke, and what the docs got wrong. Every claim has a literal MCP call or curl response behind it.

The build is deliberately thin on code and heavy on platform leverage. We connected Claude Code to KeeperHub's MCP server with a single command and authenticated via OAuth 2.1, requiring no API key for interactive use. From there, every claim in the documentation was verified live, including list_workflows, search_templates, list_action_schemas, and get_plugin calls. Where MCP was insufficient for probing auth-states or rate-limit headers, we utilized the REST API directly via curl, sourcing keys from a gitignored .env file. To handle large MCP responses that overflowed the tool-output buffer, we routed data through jq against auto-saved files. Notable technical recoveries included introspecting list_action_schemas to identify the undocumented reference syntax and correcting schema mismatches where the AI emitted network names instead of chain ID strings. The entire session was pushed to main via a self-pacing loop cron driving iterative commits every 30 minutes, utilizing a stack consisting of Claude Code, KeeperHub's API, and MPC for non-custodial wallet primitives.