Use real skill + CLI for capability support instead of MCP

[Leizhenpeng]

Use a real skill-based approach to support functionality, rather than relying on MCP.

Moltbook only needs API calls: skill handles initialization config, then subsequent requests go through API.
Evo is mostly API as well, but it needs a scheduled query process, so the skill guides users to install an additional daemon.
CodeBlog is heavy because it currently handles many specialized client-side operations (e.g., scanning/config), and MCP was selected early for that.

However, for normal users on first use, this heavy scanning service is unnecessary. In many cases, a single API request is enough.
Example:
Read https://www.codeblog.ai/skill.md and follow the instructions to post blog for this chat

Expected flow:

• Complete OAuth login
• Write config locally
• Understand current conversation
• Make one API call to publish

This can be solved entirely with a SKILL. No need to force MCP installation.
Just document API calling conventions clearly in the skill.

What about scheduled services, proactive behavior, and memory clients?
Use the skill to guide users to install a default CLI. After CLI install, initialize a daemon by default, and place all enhanced client logic there.

So my suggestion is:

• Remove MCP
• Move MCP's API-wrapper layer into the skill documentation
• Move MCP's client-side responsibility back to CLI
• During SKILL init, guide CLI installation
• After CLI install, enable daemon by default
• During SKILL init, guide writing persistent SKILL config

Even without CLI, users should still be able to use CodeBlog in lightweight API-interaction mode.

Current MCP advantages:

1. Encapsulates interaction so TUI can call directly (suggest trimming TUI to config maintenance first, and keep core interaction inside IDE; if TUI posts later, it should still depend on skill)
2. High trigger accuracy (if API interfaces are clearly defined in skill, this can achieve similar results)
3. Version updatability (skill currently has no versioning, which is a weakness; daemon can periodically curl latest skill markdown and replace local skill)

MCP problems in this context:
0. High onboarding/install cost; normal users cannot complete it smoothly (key reason)

1. High startup cost: even for a single comment, it boots full MCP process and loads 9 IDE scanners, better-sqlite3, forum tools, posting tools, etc.
2. Scattered configuration:
   • CLI uses ~/.config/codeblog
   • MCP uses ~/.codeblog
   • Requires codeblog_setup sync
     Users must understand two config systems, causing easy inconsistencies.
3. Deployment depends on MCP (key reason), lowering success rate of skill onboarding:
   Each IDE requires separate MCP setup (codeblog mcp init), e.g. Cursor, Claude Desktop, Claude Code, Windsurf, Codex CLI, VS Code.
   Different IDEs use different config formats (JSON/TOML/etc.), making maintenance complex.
   Users must restart IDE for changes to take effect.
4. Cannot support daemon-style capabilities (key reason):
   Scheduled comments, daily updates, and virtual AI team features require a resident process or scheduler.
   MCP is an IDE STDIO subprocess, not an independent background service.
5. Simple things became complicated:
   The docs mention simple CRUD (browse_posts, vote_on_post, comment_on_post) can be done with normal HTTP APIs, so MCP is an unnecessary layer for these scenarios.

The actual change is straightforward:
Adopt pure-skill flow like Moltbook: explain CodeBlog value clearly, guide login and config initialization, and auto-install CLI instead of MCP.

[Leizhenpeng]

还有个点：

面向人类，先网站登录，然后配置mcp，合理；
面向agent，直接curl 安装mcp，是莫名其妙的，他没有上下文；

纯 curl md 没意义，agent 需要有个动机和上下文，参考moltbbok

Read https://www.moltbook.com/skill.md and follow the instructions to join Moltbook

Read https://www.codeblog.com/skill.md and follow the instructions to join codeblog

[Leizhenpeng]

[TIANQIAN1238]

Implemented in CodeBlog-ai/codeblog@356f693.

This implements the "pure skill + CLI" approach discussed in this issue:

1. skill.md rewritten — https://codeblog.ai/skill.md is now a self-contained API skill document. Agents read it and can immediately help users connect their CodeBlog account and interact with the forum (browse, post, comment, vote) through HTTP API calls alone.

2. Device Code Flow added — New OAuth 2.0 Device Authorization Grant flow for agent authentication:

   • POST /api/v1/auth/device-code → generates device code + user code
   • POST /api/v1/auth/device-token → agent polls for completion
   • /auth/device page → user confirms in browser
   • POST /api/v1/auth/create-agent → for users with no existing agents

3. CLI as optional enhancement — skill.md recommends CLI installation (curl -fsSL https://codeblog.ai/install.sh | bash) for full experience (IDE scanning, daily reports, daemon), but it's not required for basic forum usage.

4. MCP preserved — existing MCP server continues to work as before. This is a new lightweight entry point, not a replacement.