BaizhiCloud Agent Toolkit for DeepSeek Harness

An optional community configuration bundle for the hosted BaizhiCloud Agent Toolkit. It loads DSH's existing MCP client with the service's Streamable HTTP endpoint and an API-key bearer header. It adds no transport, backend implementation, credential store, or default model provider.

This contribution is part of integration and promotion work for BaizhiCloud Agent Toolkit; it is not an official DeepSeek integration.

Requirements

• DeepSeek Harness 0.1.5-rc.2 and pnpm on PATH. The MCP peer dependency is pinned to the tested release; compatibility with 0.1.6 alpha or later versions has not been established.
• Your own BaizhiCloud Agent Toolkit account and API key, available as BAIZHI_API_KEY in the DSH process environment. Review service pricing before calling tools. Keep the key out of Git, patches, and shared logs.
• An existing DSH profile with the standard tools runtime and network access to https://agent-toolkit.app.baizhi.cloud/mcp.

Install in your chosen profile

From a reviewed checkout of this bundle:

npm pack --ignore-scripts
npx @deepseek-ai/dsh@0.1.5-rc.2 plugin --profile web add /absolute/path/to/dsh-plugin-baizhi-agent-toolkit-0.1.0.tgz --ignore-scripts

Replace web with the profile you actually use. DSH installs the package in that profile and adds its declared bundle to the profile's layer list. There is no build or install hook. Package installation alone does not authenticate the hosted service.

Provide BAIZHI_API_KEY using your normal environment/secret mechanism, then restart the selected profile. Missing, empty, or whitespace-only values disable this bundle entry, so it makes no connection until a nonempty key is available. A supplied value is trimmed before use in Authorization: Bearer ….

The MCP client discovers the service's current tool catalog dynamically. DSH namespaces tools as mcp__baizhi__<remote-tool-name>. Review the actual offered tools and requested operation before approving calls; this bundle does not restrict the service to search/read or promise that every exposed tool is read-only.

A first task after connecting: Find the official Model Context Protocol transport documentation, read the relevant page, and summarize it with source links. Success means the tool returned useful content and the linked sources support the summary. A visible configuration entry, local test, or discovered catalog alone is not production acceptance.

Data, failures, and removal

Tool inputs, including search queries, URLs, and any other approved arguments, are sent to BaizhiCloud. Calls may incur charges. This is static API-key bearer authentication, not OAuth. Service access and tool availability depend on your account.

If tools are absent, check whether the selected profile loaded this bundle and whether its process has a nonempty key. Connection/authentication failures do not intentionally stop the rest of the profile (failOnStartupError: false); the existing MCP client's reconnect behavior still applies. A 401 requires correcting account/key access, not installing another login wrapper.

To remove it from a profile:

npx @deepseek-ai/dsh@0.1.5-rc.2 plugin --profile web remove dsh-plugin-baizhi-agent-toolkit

Restart the profile afterwards. Removing a local bundle does not revoke a remote API key; manage keys in your BaizhiCloud account.

Validation and scope

npm ci --ignore-scripts
npm test

Six tests use the actual published DSH profile/bundle loader, Cordis lifecycle, tools registry, and MCP client against a loopback HTTP fixture with synthetic credentials. They cover initialize, tool discovery, registry dispatch, bearer rejection, missing/empty/whitespace credentials, and disposal. The test runner isolates HOME and environment; it overrides only the endpoint and reconnect policy inside the test profile and never calls the production endpoint. Disposal checks stream closure, tool unregistration, and no reconnect; it does not claim server-side session deletion.

The tests do not exercise a live BaizhiCloud account, model, web UI, production billing, every operating system, or newer DSH releases. The public integration project and this MIT configuration bundle do not make the hosted backend open source. This is an independent community integration, not an official DeepSeek endorsement.