Frenchie

io.github.Lab94由 GitHub 账号 Lab94 发布v0.5.0更新于 Oct 24, 2026

OCR, transcription, file extraction, and image generation for AI agents via MCP.

已验证Streamable HTTP可网页运行文件与存储媒体与设计AI 与机器学习

概览

AI 生成的概览

Frenchie 通过付费的积分制云服务，把 PDF、图片、音视频和 Office 文件转成 Markdown，并能根据文字提示生成图片。

可以做什么
用 OCR 把 PDF 和图片转成 Markdown 文本，可指定语言提示以提高准确度，本地模式下结果自动保存。把音频和视频文件转写成 Markdown 文稿，可指定 ISO 语言代码或自动识别。把 DOCX、XLSX、CSV、TSV 和 PPTX 文档提取为 Markdown，保存到本地目录或直接内联返回。根据文字提示生成单张图片，可设置风格、尺寸、质量、格式和背景。在托管 HTTP 模式下上传文件获取预签名地址，并获取结果文件的临时下载链接。查询异步处理任务的最新结果。
适用场景
适合需要让助手把扫描版 PDF、图片、会议录音或 Office 文档读成 Markdown，或根据提示快速生成图片的场景。尤其适合已能处理文件、又不想自建 OCR 或转写流程的智能体。
运行要求
可作为托管端点 mcp.getfrenchie.dev 使用，也可作为本地 npm 包通过 npx 启动，后者需要 Node.js。需要 Frenchie 账号和 API 密钥，通过 FRENCHIE_API_KEY 或 Authorization bearer 请求头提供。可通过 FRENCHIE_DEFAULT_LANGUAGE 或 X-Frenchie-Default-Language 请求头设置默认语言。所有处理都需要网络访问。
安装前请注意
处理会消耗付费积分：OCR 每页 1 点，转写每分钟 2 点，提取每页或每张幻灯片 0.5 至 1 点，生成图片每张 20 点，首次注册赠送 100 点。文件会上传到第三方服务；结果在首次交付后约 30 分钟过期，需及时保存 Markdown。API 密钥属于机密，安装程序会将其写入智能体配置文件。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Frenchie，将其添加到工作区。
  2. 连接 Authorization token 凭据。它们会被私密存储，绝不会在公开页面上展示。
  3. 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。

交给智能体安装

把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体，或 SourceWeft 对话中。智能体会读取该服务器的安装指南，先向你展示发布者、端点和登录方式，经你同意后再把它添加到你的客户端。密钥不会写进任何命令。

阅读 https://sourceweft.com/mcp/io.github.Lab94%2Ffrenchie-skill/install.md，并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前，先告诉我它的发布者、端点和登录方式，等我确认。不要把 API 密钥写进命令或我会提交的文件；问我密钥应从哪里获取。

查看智能体所依据的安装指南

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "frenchie": {
      "type": "http",
      "url": "https://mcp.getfrenchie.dev/",
      "headers": {
        "Authorization": "Bearer <YOUR_TOKEN>"
      }
    }
  }
}

README

[Frenchie — your agent's best friend.]

[smithery badge]

@lab94/frenchie

Frenchie — your agent's best friend.

Install Frenchie in your coding agent with one command. Frenchie Kit gives agents file and multimodal capabilities: read PDFs and images, transcribe audio and video, extract Office/CSV files to Markdown, and generate images from text prompts. Frenchie Method gives teams a structured workflow from requirement to release.

Upgrading from 0.1.x or 0.2.x? See MIGRATION.md for the breaking changes in 0.3.0 (stdio metadata-only responses, absolute-path MCP configs, new mcp --help / --selftest flags).

This package ships:

  • Frenchie Kit (/ocr, /transcribe, /extract, /generate-image, /frenchie-status commands + HTTP/stdio MCP guidance)
  • Frenchie Method (/frenchie-method, /requirement, /qa-analysis, /implement, /run-test, /pr-review, /release-check, and related workflow commands)
  • The Frenchie stdio MCP server (lab94-frenchie mcp) bundled for npx
  • An installer that wires skills and MCP config into your agent

Quick start

1. Install the skills

From the root of your project:

bash
npx @lab94/frenchie install

This installs Frenchie Kit and Frenchie Method skills. Method is text-only and works without an API key.

Optional Method config:

bash
npx @lab94/frenchie install --init-method

This creates .frenchie/method.config.json and .frenchie/.gitignore; it does not edit your root .gitignore.

2. Log in for Kit MCP tools when you need file or image capabilities

Create an account at getfrenchie.dev. You get 100 free credits on your first signup, once per email. No card required.

Then connect your local agent:

bash
npx @lab94/frenchie login

The login command opens a browser device-code flow, then writes FRENCHIE_API_KEY into the MCP config the installer created. The full key is never printed.

Fallback if you already have a key from the dashboard:

bash
npx @lab94/frenchie install --api-key fr_your_key_here

To target a specific agent:

bash
npx @lab94/frenchie login --agent claudenpx @lab94/frenchie login --agent cursornpx @lab94/frenchie login --agent codexnpx @lab94/frenchie login --agent vscodenpx @lab94/frenchie login --agent gemini

User-level installs (Antigravity, Windsurf, Zed, Claude Desktop) need the --global flag — auto-detect never writes under $HOME:

bash
npx @lab94/frenchie login --agent antigravity   --globalnpx @lab94/frenchie login --agent windsurf       --globalnpx @lab94/frenchie login --agent zed            --globalnpx @lab94/frenchie login --agent claude-desktop --global

Fallback with an existing key:

bash
npx @lab94/frenchie install --agent claude        --api-key fr_…npx @lab94/frenchie install --agent codex         --api-key fr_…npx @lab94/frenchie install --agent antigravity   --global --api-key fr_…npx @lab94/frenchie install --agent claude-desktop --global --api-key fr_…

After login, your agent can call ocr_to_markdown, transcribe_to_markdown, extract_to_markdown, or generate_image. OCR/transcription/extraction results are saved to .frenchie/<name>/result.md automatically; generated images are saved to .frenchie/<slug>/generated.<ext>.

Upgrading later

When a new Frenchie version ships, the pinned package spec inside each MCP config (@lab94/[email protected]) needs to point at the new version so npx stops serving a cached older bundle. Two upgrade commands cover everything:

bash
# Project-scoped configs in this directory (.mcp.json, .cursor/mcp.json, .codex/config.toml, etc.)npx @lab94/frenchie@latest install
# Every $HOME config that already has a frenchie entry — Antigravity, Claude Desktop, Windsurf, Zed,# and any global Claude Code / Cursor / Codex / Gemini configs you set up. cwd doesn't matter.npx @lab94/frenchie@latest install --global

install --global is upgrade-mode: it only rewrites frenchie entries that already exist. It will not surprise-install anything new under $HOME. First-time global installs still go through install --agent <name> --global.

3. Restart your agent

The installer prints the restart hint for your agent. After that, ask:

OCR ./report.pdf with Frenchie

…and Frenchie takes it from there.

For workflow orchestration, ask:

txt
/frenchie-method

/frenchie remains the Kit entrypoint in agents that use a single Frenchie command. It is not an alias for Method.

Hosted agents (Lovable, Manus, Claude.ai, ChatGPT.com, Le Chat)

These agents can't run local npm binaries. Use the hosted MCP endpoint instead:

URL:    https://mcp.getfrenchie.devHeader: Authorization: Bearer fr_your_key_here

The same @lab94/frenchie skill files work in HTTP mode — install them once with install --agent <name> and the included SKILL.md will tell the agent to upload files via upload_file before calling OCR/transcription/extraction. Image generation does not need an upload step in HTTP mode; it returns a short-lived imageUrl that the agent should download for the user.

What you get

Frenchie Kit

CommandWhat it does
/ocr <file>Parse a PDF or image into Markdown
/transcribe <file>Parse audio or video into a Markdown transcript
/extract <file>Parse DOCX, XLSX, CSV, TSV, or PPTX into Markdown
/generate-image <prompt>Generate a single image from a text prompt
/frenchie-statusCheck credits and recent jobs

Under the hood, Frenchie exposes these MCP tools:

  • ocr_to_markdown
  • transcribe_to_markdown
  • extract_to_markdown
  • generate_image
  • get_job_result
  • upload_file (HTTP mode only)
  • fetch_result_file (HTTP mode only)

Frenchie Method

Slash notation is the canonical Method command name. Claude Code reads command files directly. Codex Desktop gets the same entries through wrapper skills such as requirement and qa-analysis; restart Codex after install before trying /requirement. Hosts without slash support can use natural language such as Frenchie Method requirement ....

CommandWhat it does
/frenchie-methodRoute a task through Method from requirement to release
/method-checkValidate Method config, artifacts, gates, and next action
/requirementTurn a request into scoped requirements
/qa-analysis / /test-caseDesign proof before implementation planning
/sa-analysis / /implementation-planAnalyze architecture and implementation approach
/implementExecute from Method artifacts using the current agent
/run-testRun repo-declared verification and write a test report
/pr-review / /review-prReview the result, tests, security, and artifact drift
/release-checkRead-only release readiness gate

Method stores optional workflow artifacts under .frenchie/docs/feature/. install --init-method also writes a visible skillTriggers skeleton in .frenchie/method.config.json; command keys are empty slots teams can fill with installed agent skills, for example "requirement": ["brainstorming"]. Teams decide whether to commit or ignore Method config and artifacts.

Invocation — how to call Frenchie in each agent

Every agent handles MCP a little differently. /ocr is a Claude Code-only slash command; Method slash commands are host-dependent and use Codex wrapper skills outside Claude Code. Other agents use natural language, @-mention, or a server-name slash command. All facts below are dogfood-verified.

AgentInvokeFull guide
Claude Code/ocr TOR.pdfdocs
CursorUse Frenchie to OCR TOR.pdfdocs
Codex (Desktop / CLI / IDE)Kit: /frenchie TOR.pdf · @frenchie ocr TOR.pdf; Method: /requirement after restart, or Frenchie Method requirement ...docs
Antigravity/frenchie TOR.pdf (invokes by server name)docs
VS Code Copilot/frenchie TOR.pdfdocs
Claude DesktopUse Frenchie to OCR TOR.pdfdocs
WindsurfOCR TOR.pdf via Frenchiedocs
Gemini CLIOCR TOR.pdf with Frenchiedocs
ZedOCR TOR.pdf via Frenchiedocs

Something not working? See the symptom-first troubleshooting guide — every error we've hit in dogfood has a canonical entry.

Pricing

Simple numbers. No subscriptions.

File extraction pricing: DOCX 0.5 credits per page, XLSX 0.5 credits per sheet, CSV/TSV 0.5 credits per file, PPTX 1 credit per slide.

ActionCost
OCR1 credit per page
Transcription2 credits per minute
DOCX extraction0.5 credits per page
XLSX extraction0.5 credits per sheet
CSV/TSV extraction0.5 credits per file
PPTX extraction1 credit per slide
Image generation20 credits per image

$1 = 100 credits. Credits don't expire.

Privacy

Files are processed and deleted. Results expire about 30 minutes after first delivery. If you need a durable copy, save the Markdown when it comes back.

Supported formats

OCR: PDF, PNG, JPG, JPEG, WebP

Transcription: MP3, M4A, WAV, MP4, MOV, WebM

File extraction: DOCX, XLSX, CSV, TSV, PPTX

Image generation: PNG, JPEG, WebP output from text prompts

Need help?

来源：README.md，提交 91ba927

工具

7

ocr_to_markdown

OCR to Markdown

Convert PDF/image files into Markdown through Frenchie

输入参数api_keyfile_pathuploaded_file_reference

transcribe_to_markdown

Transcribe to Markdown

Convert audio/video files into Markdown transcripts through Frenchie. Set language (ISO 639-1 code, e.g. 'th', 'en', 'ja') for better accuracy; omit for auto-detection.

输入参数api_keylanguagefile_pathuploaded_file_reference

generate_image

Generate Image

Generate a single image from a text prompt through Frenchie. Required: prompt. Optional: style (free-text style direction), size, quality, format, background. stdio mode auto-saves the image to .frenchie/<slug>/generated.<ext>; HTTP mode returns a presigned imageUrl that the agent should download for the user.

输入参数sizestyleformatprompt必填api_keyqualitybackgroundoutput_dir

extract_to_markdown

Extract to Markdown

Extract structured documents (.docx, .xlsx, .csv, .tsv, .pptx) into Markdown through Frenchie. stdio mode auto-saves the result to .frenchie/<name>/result.md; HTTP mode returns inline Markdown.

输入参数api_keyfile_pathuploaded_file_reference

get_job_result

Get Job Result

Fetch the latest async Frenchie job result

输入参数job_id必填api_key

upload_file

Upload File (HTTP)

Get a presigned upload URL for use with ocr_to_markdown, transcribe_to_markdown, or extract_to_markdown in HTTP mode. After calling this tool, PUT the file to upload_url (with the correct Content-Type header), then pass object_key as uploaded_file_reference to the processing tool.

输入参数api_keyfilename必填file_size必填mime_type必填

fetch_result_file

Fetch Result File (HTTP)

Get a temporary download URL for a result file from OCR/transcription output. Use this to download images referenced as frenchie-result: in the result markdown.

输入参数api_keyobject_key必填

版本历史

1
  1. v0.5.0最新Sep 16, 2026

评分

已满足适用检查项的 100%

该评分衡量此收录条目的完整程度，依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价，也不表示服务器安全或当前可用：可连通性是带日期的检查结果，评分也不包含安装量或用户评价。

  • 近 30 天内可连通必需25 分
  • 已声明认证方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分