Ownhand

io.github.jordan-gibbsv0.3.1更新于 Oct 2, 2026

Write in the user's own voice, fitted to the occasion. Learns from the text they actually send.

概览

AI 生成的概览

Ownhand 是一个远程 MCP 服务器,能按用户自己的写作风格、结合具体场合改写草稿,并从反馈中学习。

功能
Ownhand 为助手提供工具,把草稿改写成读起来像用户本人所写,并适配 Slack 私信、冷邮件、PR 描述等场合。它从 5 到 10 篇用户自己写的内容中建立 Hand,然后持续从用户实际发送的内容中学习。工具包括 write、send_feedback、get_hand、create_hand、update_hand 和 get_account,另有提示词和资源,如指南和场合表。
适用场景
当助手需要起草消息、邮件、提案、PR 描述、评审评论、文档或状态更新,而用户希望这些内容听起来像自己写的而不是通用 AI 文本时,值得添加。它面向需要大量书面沟通、并愿意对草稿给出反馈的人。
运行要求
远程 streamable HTTP 端点 OAuth 2.1 和动态客户端注册,在浏览器中完成;大多数客户端无需复制 API 密钥,脚本和 CI 另有 API 密钥选项。需要账号,新账号有 1 美元免费额度。
安装前请注意
付费服务:先有 1 美元免费额度,之后需充值并按 token 付费,每次改写约半美分,无订阅。写作和学习通过 OpenRouter 的 contributor 层级运行在 Muse Spark 上,因此该提供商可能使用 Ownhand 处理的一切内容(草稿、样本、线程上下文、反馈)来改进其模型;不要用它处理你希望排除在外的文本。存储的数据包括账号信息、Hand、请求和反馈;作为上下文传入的线程消息不会被存储。可随时删除 Hand 或吊销密钥。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Ownhand,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Web executable,通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "ownhand": {
      "type": "http",
      "url": "https://ownhand.dev/mcp"
    }
  }
}

README

Ownhand

Agents, writing like you.

Ownhand is a remote MCP server that writes text in your own voice, fitted to where it is going: a Slack DM, a cold email, a PR description. It learns your voice from 5 to 10 things you wrote, then keeps learning from what you actually send.

[Add to Cursor] [Add to VS Code]

  • Endpoint: https://ownhand.dev/mcp (streamable HTTP)
  • Sign-in: OAuth 2.1 with dynamic client registration. Add the URL, sign in in the browser, pick your Hand, click Allow. No API key to copy.
  • New accounts start with $1 of free credit.

Set up with one sentence

Paste this into any agent (Claude Code, Codex, Cursor, Gemini CLI, and others). It adds the server, starts sign-in and sets up your Hand.

text
Fetch and execute the instructions to set me up for Ownhand from https://ownhand.dev/agent-setup/prompt.md. If your web tool can't open it, download it with curl or Invoke-WebRequest.

Install by client

ClientInstallThen
Claude Codeclaude mcp add --transport http --scope user ownhand https://ownhand.dev/mcpRun /mcp, pick ownhand, sign in.
Claude Code plugin/plugin marketplace add jordan-gibbs/ownhand then /plugin install ownhand@ownhandAdds the server and an ownhand skill. Run /mcp to sign in.
Claude (claude.ai, Desktop, mobile)Settings, Connectors, Add custom connector, paste https://ownhand.dev/mcpClick Connect and sign in.
ChatGPTSettings, turn on Developer mode, add a connector with https://ownhand.dev/mcpSign in when asked.
CursorAdd to Cursor above, or examples/cursor.mcp.json in ~/.cursor/mcp.jsonClick Connect next to ownhand.
VS Code (Copilot)Add to VS Code above, or code --add-mcp '{"name":"ownhand","type":"http","url":"https://ownhand.dev/mcp"}'Sign in when asked.
Codex CLIcodex mcp add ownhand --url https://ownhand.dev/mcpcodex mcp login ownhand if no browser opened. Restart Codex if it was running (codex resume --last).
Gemini CLIgemini mcp add --transport http --scope user ownhand https://ownhand.dev/mcp, or gemini extensions install https://github.com/jordan-gibbs/ownhand/mcp auth ownhand
Windsurfexamples/windsurf.mcp_config.json in mcp_config.json (uses serverUrl)Sign in when asked.
Zedexamples/zed.settings.json in your Zed settingsSign in when asked.
Goosegoose configure, add a remote extension (streamable HTTP) with the URLSign in when asked.
Clineexamples/cline_mcp_settings.json (uses "type": "streamableHttp")Click to sign in.
Anything elseA remote streamable HTTP server named ownhand at https://ownhand.dev/mcpSign in when asked.

Raw install deeplinks, if you want to open them yourself:

text
cursor://anysphere.cursor-deeplink/mcp/install?name=ownhand&config=eyJ1cmwiOiJodHRwczovL293bmhhbmQuZGV2L21jcCJ9vscode:mcp/install?%7B%22name%22%3A%22ownhand%22%2C%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fownhand.dev%2Fmcp%22%7Dvscode-insiders:mcp/install?%7B%22name%22%3A%22ownhand%22%2C%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fownhand.dev%2Fmcp%22%7Dgoose://extension?url=https%3A%2F%2Fownhand.dev%2Fmcp&type=streamable_http&id=ownhand&name=Ownhand&description=Write+in+your+own+voice

More configs, including VS Code (.vscode/mcp.json) and Codex (config.toml), are in examples/. The live version of this table is at https://ownhand.dev/connect.

Scripts and CI

For clients without browser sign-in, an API key option is described in the docs: https://ownhand.dev/docs#connect

First run

Once connected, say: "Set up my Ownhand Hand." The agent asks for 5 to 10 things you wrote yourself (Slack messages, emails, PR descriptions) and how you'd describe your writing, then calls create_hand. Clients that show MCP prompts also list set_up_my_hand.

To make your agent use Ownhand without being asked, paste AGENTS.md into your CLAUDE.md, AGENTS.md, .cursorrules or custom instructions.

Tools

ToolWhat it does
writeRewrites a draft so it reads like you wrote it, fitted to the occasion. Keeps every fact. Returns the text and a request_id.
send_feedbackReports what you did with a draft (approved, edited with the exact sent text, or rejected) so your Hand learns.
get_handLists your Hands, or shows what one has learned: style card, rules, sample counts.
create_handCreates a Hand from texts you wrote yourself.
update_handAdds samples, sets how you describe your writing, or pins, rejects or activates a learned rule.
get_accountShows your account, credit balance, recent charges and usage.

Prompts: set_up_my_hand, teach_my_voice, reply_to_thread, draft_email. Resources: ownhand://guide (the agent playbook), ownhand://occasions (the occasion table).

Occasions: chat_dm, chat_channel, email_internal, email_formal, email_cold, email_warm, proposal_cold, proposal_warm, pr_description, review_comment, docs, status_update.

How learning works

  1. Your agent calls write and shows you the draft. Nothing is sent without your OK.
  2. You send it as is, edit it, or drop it.
  3. Your agent calls send_feedback with exactly the text you sent.
  4. Ownhand compares the draft with what you sent and updates your Hand. A reason you state becomes a rule right away. A pattern it only observes has to show up twice. Learning runs in the background, and the Hand's version goes up when it finishes.

You can pin or reject any learned rule with update_hand. Learning from your feedback changes only your own Hand.

Pricing

$1 free to start. Then you add credit and pay for the tokens you use, about half a cent per rewrite. No subscription and no monthly fee. Details: https://ownhand.dev/pricing.

Privacy

  • Stored: your account details, your Hands (samples, style card, learned rules), your requests and the feedback you send.
  • Not stored: thread messages your agent passes as context. They are used for that one request.
  • Not sold, not shared for advertising, and not used to train our own models.
  • Writing and learning run on Muse Spark through OpenRouter's contributor tier, so the provider may use everything Ownhand processes (drafts, samples, thread context, feedback) to improve its models. Don't use Ownhand for text you want kept out of that.
  • You can delete a Hand or revoke a key at any time.

Read the full policy: https://ownhand.dev/privacy.

Docs

What's in this repo

  • server.json: the entry for the official MCP Registry
  • .claude-plugin/, .mcp.json, skills/ownhand/: a Claude Code plugin and marketplace
  • gemini-extension.json, GEMINI.md: a Gemini CLI extension
  • AGENTS.md: the instruction block for any agent
  • examples/: client configs

License

The configs and docs in this repo are MIT licensed (see LICENSE). The Ownhand service itself is proprietary and run by Thalient Labs.

Contact: [email protected]

来源:README.md,提交 b6a8204

工具

0
工具元数据尚未被收录。

版本历史

1
  1. v0.3.1最新Oct 2, 2026