Discuss Baizhi Agent Toolkit plugin onboarding and curated distribution

What feature would you like to see?

Hi Kimi Code team,

I maintain Baizhi Cloud's Agent Toolkit integrations and coordinate product operations. I am submitting this request through ct-jaryn on behalf of Baizhi Cloud. We would like to discuss a maintainable third-party plugin and its credential-onboarding path before proposing any repository change.

The intended experience is to select Baizhi Agent Toolkit from the plugin manager, explicitly confirm installation, supply a personal API key through a supported hidden-input/credential mechanism, and use three scoped tools: web search, page reading, and structured extraction.

Kimi Code CLI already supports remote Streamable HTTP and bearerTokenEnvVar; we are not asking for basic MCP support. A small plugin could declare our fixed HTTPS endpoint and an enabledTools allowlist for websearch_search, web_scrape, and web_extract, without automatically exposing additional remote capabilities. Environment-based configuration would remain available for advanced users.

The remaining questions are:

1. Would maintainers consider this service-specific integration for the Curated catalog, and what review requirements should a third-party maintainer follow?
2. Is there a supported plugin credential-onboarding mechanism that we should use? In the current manifest parser and CLI installation path, we did not find a general declarative secret-input setting. We do not want to present manual environment-variable setup as an existing native password prompt.

If a core feature is needed, we will follow your issue-first policy and wait for the team's direction or an explicit invitation, rather than opening an uninvited feature PR.

Additional information

Baizhi Agent Toolkit is a hosted commercial tool service, not an LLM inference provider. The proposed adapter would be free and open source; users need their own Baizhi account and key, and service calls can consume credits. Our official integration repository contains configuration, documentation, and tests, not the hosted backend implementation.

• Product
• Console and personal keys
• Official integration documentation
• Remote endpoint: https://agent-toolkit.app.baizhi.cloud/mcp (Streamable HTTP, user-provided Bearer authentication).

Source review at 86e08208e3445171f6500c948234d7e3a0684324 found the existing remote MCP schema, Bearer environment-variable handling, and third-party Curated entries. These are implementation precedents, not evidence of approval or completed Baizhi compatibility testing.

We would maintain versioned plugin metadata, English/Chinese setup and troubleshooting documentation, credential-redaction safeguards, and scoped tests. Acceptance checks would include missing/canceled credentials, unchanged unrelated configuration, fixed-endpoint authentication, effective three-tool filtering, installation trust/cancel, reload and credential replacement, and separately authorized production discovery/tool testing. No real key would appear in a plugin package, repository, issue, logs, or CI.

Tool inputs, URLs, and extraction instructions are sent to Baizhi's hosted service; costs and the verified privacy/data-flow policy would be documented. No shell hooks, code execution, payment tools, arbitrary-tool forwarding, or automatic-approval rules are proposed. Users must still review Kimi's selected permission mode; we would not promise that every MCP invocation always receives a separate approval prompt.

This is a scope and contribution-route request. We have not completed a Kimi Code CLI plugin installation, native credential-input test, or production-service test. We understand that current plugin installation is user-scoped, not project-scoped, and that Curated inclusion is a maintainer decision rather than self-service publication. We are not proposing an entry in the Kimi-maintained Official tier.

Thank you for considering the proposal. We would be glad to cooperate on an integration that fits Kimi Code's architecture and review standards. This request does not imply Kimi endorsement or an established partnership.

Start by reading packages/agent-core-v2/src/mcpCore/config-schema.ts and client-remote.ts, then compare the existing entries in plugins/marketplace.json. The issue requests maintainer guidance on Curated inclusion and credential onboarding rather than naming a defined code change. Done requires an agreed contribution route and acceptance criteria before implementation.