Community setup: use Baizhi Agent Toolkit through OpenHuman's user-configured MCP connection

[ct-jaryn]

What is it called?

Baizhi Agent Toolkit setup for OpenHuman

What does it do?

An optional, user-configured connection for web research tasks: discover the MCP tools available to your own Baizhi account, search for public pages, inspect selected URLs, and ask the agent for a summary with source links. The guide follows the current mcp.json workflow, preserves other installed servers, and explains write-only credential readback, rotation, explicit Header deletion, and uninstall. This is maintained as part of the Baizhi integration effort. Users supply their own key; service calls send tool inputs to the hosted service and may consume credits. The guide does not impose a hard spending limit. The hosted service backend is not included in the open-source guide, and this is not an OpenHuman endorsement or a request for a built-in preset.

Where can people get it?

Community guide and reproducible validation

Kind

A setup or configuration

Tested against

OpenHuman source version 0.63.29, commit eb4fdc0f4f4a036d8ab7c12bd52f16128e4a6b5d, with its recorded recursive submodules; macOS ARM64, Rust 1.96.1.

Six focused native integration tests pass. They use OpenHuman's config/store implementation, the real tinymcp HTTP client, the product tool factory, and TinyAgents' actual tool-batch scheduler. A synthetic loopback MCP peer verifies initialize/list/call/session DELETE, missing/wrong/literal-placeholder credentials, secret-free readback and re-save, an empty Header object, rotation/deletion, partial Header updates, and preservation of another server.

The server and scheduled tool calls are fixtures. Test execution is restricted to loopback. These results do not cover a live Baizhi call, model-generated tool selection, the full OpenHuman session/approval middleware or use_mcp_server handoff, desktop UI, or registry discovery; there were no production tool or model calls. The guide and scripts include the exact scope and source-bound receipt. The minimal mcp feature profile is not a full desktop test suite.

AI assistance: OpenAI Codex helped research the host implementation, write the guide and tests, and run the synthetic validation. No claim of manual UI acceptance is made.

[redbotster]

the write-only readback and secret-free re-save patterns you described are the right instincts, but storing the raw key in mcp.json still means it sits in plaintext on disk and can land in logs or model context if the server echoes tool inputs. one alternative worth knowing: tools like 1claw (disclosure: i work on it) store the key in an hsm-backed vault and give the agent a short-lived scoped token that only references a path, so the raw value never touches the config file or context window. that said, your current approach with explicit rotation and deletion docs is a solid baseline for a community guide, especially since you're testing the missing/placeholder/wrong-credential cases explicitly.

[ct-jaryn]

I checked the credential-storage path for the guide's pinned OpenHuman 0.63.29 snapshot and clarified the guide.

At that snapshot, the mcp.json tab edits the workspace's installed-server store. Header credentials go into the MCP SQLite database under the configured workspace; the source path stores strings without application-level encryption even though configuration readback omits them. The guide now explicitly covers protecting that workspace and backups, and explains that deleting or replacing a local credential is separate from service-side revocation or secure erasure.

The clarification links to the exact host and store revisions. This is a source-verified documentation update; it does not extend the guide's existing synthetic validation into a production or desktop-UI claim. Prepared with AI assistance.