RemoraFish MCP
Exposes authorized ChatGPT, DeepSeek, Kimi, and Grok web sessions to agents as MCP tools, including chat and local file attachments over HTTP or stdio.
README
RemoraFish
Agent-first framework for turning your authorized ChatGPT, DeepSeek, Kimi, Grok, and external-provider web sessions into OpenAI-compatible APIs, Anthropic Messages, and MCP tools. It ships a headless runtime, provider SDK, account control plane, built-in reference providers, and a reference web UI.
[!IMPORTANT] You are solely responsible for your accounts and for every action performed with RemoraFish. Use it only with accounts and browser profiles that you own or are explicitly authorized to operate. You are responsible for credential security, compliance with provider terms and applicable law, and consequences including account restrictions, suspension, or loss. The project authors do not operate your accounts or guarantee their availability.
RemoraFish v0.2 is a framework preview. It uses web sessions rather than official developer API keys. Providers may change their private protocols or restrict accounts; do not use an account you cannot afford to lose. The public repository contains no user credentials. Browser sessions and token caches are local runtime data stored outside the project tree by default.
Three ways to use RemoraFish
| Direction | What you get |
|---|---|
| Learn | A tested reference architecture for building Web2API adapters and MCP tools over LLM web services |
| Run | A ready local gateway with account pools, routing, OpenAI/Anthropic surfaces, inference MCP, CLI, and a reference UI |
| Extend | A versioned provider API, external plugin discovery, scaffold, and conformance suite for building your own infrastructure or fork |
The bundled UI is intentionally a reference application, not the framework boundary. Agents, API clients, custom frontends, and product forks all use the same headless runtime.
Quick start
1. Install
git clone https://github.com/howdeploy/remorafish.git
cd remorafish
uv sync --extra dev
2. Connect a coding agent
Coding agents should start with AGENTS.md. If an agent needs the RemoraFish project context delivered persistently across sessions, install and configure choirboy-prompt as an optional instruction transport. Replace its bundled demonstration lore with the RemoraFish project guide; do not use the demo biography unchanged.
3. Start the agent control plane
The recommended workflow is to let a local coding agent orchestrate account setup through the separate Admin MCP. RemoraFish itself scans the user's browser, keeps raw credentials out of the agent context, returns masked candidates, and imports only explicitly confirmed selections:
uv run remorafish admin-mcp \
--allow-browser-scan \
--allow-account-mutations
Direct CLI and reference UI
The CLI and reference UI remain available for direct operation:
# Browsers may stay open. Scan live snapshots, validate, and save valid accounts.
uv run remorafish accounts scan --browser all --check --save --yes
# Or choose several browsers and get agent-friendly JSON.
uv run remorafish accounts scan -b brave -b firefox --check --json
uv run remorafish serve
The server listens on http://127.0.0.1:8787 by default.
Open Settings → Accounts to add any number of provider accounts with a
per-account proxy. RemoraFish can open an isolated local browser, submit a
stored login/password, wait for MFA or captcha, capture the resulting runtime
token, and add the account to the live routing pool without a restart.
curl http://127.0.0.1:8787/v1/responses \
-H 'Content-Type: application/json' \
-d '{"model":"deepseek/chat","input":"Hello"}'
# MCP over stdio
uv run remorafish mcp
Agents can attach local files through web_chat. The call must include
attachments: [{"path":"./report.pdf"}] and explicit
confirm_file_access: true; RemoraFish validates the path and limits, uploads
the bytes through the selected provider's native ChatGPT, DeepSeek, Kimi, or
Grok web flow, and keeps local paths out of provider adapters. HTTP clients can
use inline base64 image/file parts. See HTTP APIs and MCP.
What is available
| Surface | Entry point |
|---|---|
| OpenAI Chat Completions | POST /v1/chat/completions |
| OpenAI Responses subset | POST /v1/responses |
| Anthropic Messages subset | POST /v1/messages |
| Model catalog | GET /v1/models |
| Inference MCP with local file attachments over HTTP / stdio | POST /mcp / remorafish mcp |
| Local agent account control | remorafish admin-mcp |
| Terminal / reference browser UI | remorafish chat / / |
| Provider framework | remorafish providers init / providers check |
Browser import supports Chrome, Chromium, Edge, Brave, Vivaldi, Opera,
Firefox, and LibreWolf on Linux, macOS, and Windows. Raw secrets are masked in
terminal and JSON output; saved credentials go to
~/.remorafish/accounts.json with mode 0600.
Login/password and authenticated proxy URLs are kept separately in the
encrypted ~/.remorafish/account_vault.json; its local key and persistent
login profiles also use owner-only permissions.
ChatGPT is exposed as chatgpt/auto; chatgpt/auto-search enables the web
search feature. Connect it from Settings → Accounts → ChatGPT with an
access token, an existing browser session import, or interactive browser
sign-in. The ChatGPT web protocol includes a changing anti-bot layer, so a
live account smoke test is still required after upstream frontend changes.
Documentation
- How the system works
- Provider framework and formal contract
- Agent-first account control plane
- Reference UI boundary
- Browser credential extraction
- Multi-account management and routing
- OpenAI, Anthropic, and MCP interfaces
- Testing real web subscriptions
- Original design contract and protocol research
Development
uv run pytest -q
uv run remorafish providers check
The official Agentation toolbar is bundled as a localhost-only React island so the main web UI can remain framework-free. Rebuild it after dependency changes:
npm install
npm run build:agentation
Agent Sync uses http://localhost:4747. Register its MCP server with Codex:
codex mcp add agentation -- npx -y agentation-mcp server
Restart Codex after changing MCP configuration. The RemoraFish source is MIT.
The bundled Agentation component remains under PolyForm Shield 1.0.0; its
required notice is shipped as remorafish/webui/agentation.LICENSE.txt. See
third-party notices for protocol implementation
attributions and all license boundaries.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。