Useful skill installer options:

The installer is idempotent. If an existing patchxnote-mcp directory differs and is not managed by this package, it refuses to overwrite it unless --force is supplied. Managed installs include .patchxnote-agent-skill.json with the package version and source hash.

For users of the standard skills CLI, this Codex example installs from GitHub into the current project. It installs the Skill, not the MCP connection. Choose either this route or the npm-bundled install above.

npx -y skills add ZsTs119/patchx-freenote-agent --skill patchxnote-mcp --agent codex --yes

Local setup supports these client IDs:

vscode, cursor, codex, claude-code, claude-desktop, windsurf, trae, qoder, workbuddy

vscode, cursor, codex, claude-desktop, and windsurf can write a local config file after confirmation. claude-code, trae, qoder, and workbuddy return manual commands or copyable config in V1. Platform clients such as Feishu Aily, Doubao Work Partner, Tencent Agent Development Platform, and enterprise WorkBuddy require the hosted remote MCP gateway and platform-console acceptance instead of local npx.

Useful setup flags:

patchxnote setup --client cursor --dry-run --print-config
patchxnote setup --client cursor --yes
patchxnote setup --client cursor --no-browser
patchxnote setup --all-local-supported --dry-run
patchxnote setup --client cursor --output json

Run setup in the same OS/runtime that will later launch MCP. For example, a Windows desktop editor needs Windows Credential Manager credentials, while a WSL or remote VS Code session needs credentials in that Linux runtime.

For generic local stdio MCP hosts, use the pure JSON printed by:

npx -y patchxnote-agent@latest mcp config

The default config looks like this:

{
  "mcpServers": {
    "patchxnote": {
      "command": "npx",
      "args": ["-y", "patchxnote-agent@latest", "mcp", "serve"]
    }
  }
}

Some clients may require a wrapper-specific field such as type: "stdio" or a different top-level key, but the command and args stay the same. If a client rejects npx, kills slow cold starts, or requires allowlisted absolute paths, use the fallback printed by:

npx -y patchxnote-agent@latest install --print-config

The fallback config uses the installed binary path:

{
  "mcpServers": {
    "patchxnote": {
      "command": "/absolute/path/to/patchxnote",
      "args": ["mcp", "serve"]
    }
  }
}

Account And Record Lookup

Webhook Configuration And Sending

AI Result Inspection

Record tools require an explicit platform argument: mobile or desktop. The record list now includes formal saved results plus readable model-generated outputs when the server has model IO data. patchxnote model-io list remains the lower-level AI processing list for finding request IDs and filtering by task or state.

Webhook MCP tools share the same local config, keychain, templates, and sender modules as the CLI. They do not return full webhook URLs or signing secrets, and send calls perform external network requests only when the MCP client explicitly invokes the send tool.

AI result tools are explicit inspection tools. They may expose source text or AI payloads for the logged-in user, so use them only from trusted local MCP hosts. Large fields should be written to an explicit local out file.

The examples below are independent commands. Choose the operation for your task; they are not one script to run from top to bottom.

Browser MCP login and local MCP service:

patchxnote version
patchxnote mcp login
patchxnote mcp status
patchxnote mcp config
patchxnote setup --client cursor
patchxnote mcp serve

Terminal CLI login:

patchxnote login
patchxnote auth status

List AI processing runs and export results:

patchxnote model-io list --platform mobile
patchxnote model-io source-text --request-id <request_id> --platform mobile --out ./source.txt
patchxnote model-io provider-response --request-id <request_id> --platform mobile --out ./provider-response.json
patchxnote model-io parsed-result --request-id <request_id> --platform mobile --out ./parsed-result.json
patchxnote model-io packaged-result --request-id <request_id> --platform mobile --out ./packaged-result.json
patchxnote model-io export --request-id <request_id> --platform mobile --out ./model-io.json

Get request_id from patchxnote model-io list --platform mobile|desktop when you need a lower-level AI processing run. MCP patchxnote_list_memories returns id and platform for record rendering, drafts, webhook workflows, and model IO field tools; for model-generated entries, that id can be the same value as request_id.

Configure and send webhooks:

patchxnote webhook set "Product Feishu" --type feishu --url-stdin
patchxnote webhook list
patchxnote webhook test "Product Feishu"
patchxnote webhook draft --memory-id <memory_id> --platform mobile --out ./patchxnote-drafts/example
patchxnote webhook send --target "Product Feishu" --file ./message.md
patchxnote webhook send --target "Product Feishu" --draft ./patchxnote-drafts/example
patchxnote webhook remove "Product Feishu"

Useful global flags:

--server-base-url <url>   PatchX Freenote API base URL
--profile <name>          local profile name
--output json             machine-readable output where supported
--config <path>           non-secret config file path

The npm package is a small installer/launcher wrapper:

npx -y patchxnote-agent@latest mcp login
npx -y patchxnote-agent@latest mcp status
npx -y patchxnote-agent@latest mcp config
npx -y patchxnote-agent@latest mcp serve
npx -y patchxnote-agent@latest skill install
npx -y patchxnote-agent@latest login
npx -y patchxnote-agent@latest setup --client cursor
npx -y patchxnote-agent@latest install
npx -y patchxnote-agent@latest update

Webhook URLs and optional Feishu/DingTalk signing secrets are stored in the local secure credential store, not in the non-secret config file. --url-stdin and --secret-stdin avoid shell history. CLI and MCP webhook sending is manual only, does not follow redirects, and surfaces provider errors directly.

patchxnote model-io export is the preferred complete AI processing export command. patchxnote webhook export-model-io remains available for compatibility.

Version 0.2.12 and later default to production. Existing explicit --server-base-url flags, PATCHXNOTE_SERVER_BASE_URL / legacy PATCHNOTE_SERVER_BASE_URL environment variables, and server.base_url config values override that default; update or remove test-server overrides in both the terminal and MCP host configuration.

Run mcp login in the same OS/runtime and profile that starts MCP, then check mcp status --verify. OAuth credentials are bound to the server address, so test login does not authenticate production. This update does not migrate test accounts or records. To keep both environments, use separate profiles and an explicit base URL for each.

0.2.15 Highlights

• Corrects the MCP Registry namespace to preserve GitHub owner casing: io.github.ZsTs119/patchxnote-agent.
• Adds the mcp serve launch arguments and validates release metadata before publication.
• Keeps the existing runtime behavior and Skill license.

0.2.13 Highlights

• Unifies the English and Chinese setup prompts across README, npm, the bundled Skill, and marketplace copy.
• Explicitly names the production service, Remote MCP endpoint, setup page, and App download URL; existing test-address overrides must be updated when switching.
• Points npm and plugin homepage metadata to the production setup page and retires obsolete domain proposals in current guidance.

0.2.12 Highlights

• Defaults to the production API at https://freenote.patch-x.cn; hosted MCP uses https://freenote.patch-x.cn/mcp.
• Keeps explicit server flags, environment variables, and config-file overrides available for test environments.
• Documents production re-login and migration of existing test-server configurations.

0.2.11 Highlights

• Bundles the canonical PatchX Freenote MCP Skill inside the npm package.
• Adds npx -y patchxnote-agent@latest skill install for npm-based skill installation without relying on a separate skills CLI or GitHub clone.
• Makes skill installation idempotent, records a managed marker, and protects manually edited skill folders unless --force is explicitly used.
• Extends skill package sync and validation so OpenAI, Claude, and npm copies stay byte-identical to skills/patchxnote-mcp/.
• Updates the one-line setup prompt and discovery metadata to prefer the npm-bundled skill installer while preserving MCP setup, browser OAuth, and tool verification.

0.2.10 Highlights

• Adds the reusable PatchX Freenote MCP Skill at skills/patchxnote-mcp/ so compatible AI clients can keep the setup and usage SOP across fresh or long sessions.
• Adds OpenAI/Codex, Claude Code, Agent Skills, MCP Registry, Smithery, and third-party directory draft packaging and listing materials.
• Adds MCP Registry metadata through server.json and package.json#mcpName, plus local validation and stdio smoke scripts for release evidence.
• Updates the one-line setup prompt so agents install the skill when supported, then run MCP setup, browser OAuth, and tool verification without asking users to paste codes or tokens into chat.

0.2.9 Highlights

• Polishes the browser OAuth loopback success and failure pages shown after patchxnote mcp login.
• Keeps the post-login result page focused on plain user guidance, without exposing OAuth codes, state values, or token-shaped details.
• Adds regression coverage for the browser callback pages so future changes keep those sensitive details out of the UI.

0.2.8 Highlights

• Adds patchxnote setup --client <id> and npm wrapper delegation with dry-run, JSON output, confirmation, config printing, force repair, and local MCP smoke hooks.
• Adds a client registry for VS Code, Cursor, Codex, Claude Code, Claude Desktop, Windsurf, Trae, Qoder, WorkBuddy, Feishu/Doubao, Tencent platform, and P1 follow-up clients.
• Adds JSON and TOML config merge adapters with backup, conflict detection, rollback, and manual JSONC mode.
• Adds patchxnote mcp login/status/logout, browser OAuth with PKCE, MCP OAuth secure storage, and remote /mcp stdio proxy mode with local fallback.
• Adds website page specs, detail-page copy, and remote platform gateway design for product-style onboarding.

0.2.6 Highlights

• MCP expands to 19 tools across account/record lookup, webhook delivery, and AI result inspection.
• Webhook workflows are available to MCP: configure named aliases and manually send to Feishu, DingTalk, or generic webhooks.
• AI processing runs can be listed first, then inspected by request_id for source text, AI response, parsed result, and final result.
• Record lists can now include readable model-generated outputs returned by the server, so users can find a record first and then inspect its source text, AI response, parsed result, or final result.
• Webhook aliases containing dots, Chinese text, or spaces now persist and reload correctly from the local config file.
• README, npm README, and public visual assets have been refreshed for the new user-facing positioning.