SOMA MCP
A private, self-hosted MCP server that wraps a retrieval pipeline over your own data, enabling trusted AI agents to access and manage your personal memory through standard MCP tools.
README
SOMA MCP
One private memory. Every agent you trust — powered by the official Model Context Protocol.
SOMA MCP is a private, self-hosted reference implementation of the official Model Context Protocol (MCP). It puts a single person's memory behind a small set of MCP tools that any MCP-compatible agent — Claude, ChatGPT, Cursor, VS Code, or your own — can call over an authenticated connection.
Where most MCP servers wrap a public API or a shared SaaS backend, SOMA MCP wraps your own retrieval pipeline over your own data, on hardware you control. Same protocol, same clients, sovereign substrate.
This repository is the reference implementation of the SOMA MCP server: the FastMCP transport shell, the FastMCP-free tool kernel, and the rate-limit / audit guard — with the retrieval engine abstracted behind an
Engineinterface. A dependency-freeInMemoryEngineships so the server runs and its tests pass out of the box. Plug in your ownEngineto front your own memory.
Why it's different
| Typical hosted MCP server | SOMA MCP | |
|---|---|---|
| Backend | A public/SaaS API | Your retrieval pipeline over your data |
| Hosting | Vendor cloud | Self-hosted (you control the box and keys) |
| Tenancy | Often multi-tenant | One private memory per instance |
| Access | API key / shared auth | OAuth 2.1 plus a fail-closed subject allowlist |
| Data exposure | Provider sees your data | Data never leaves your infrastructure |
| Protocol | Official MCP | Official MCP (same clients, same wire) |
The differentiator is not the protocol — it's sovereignty over the substrate while staying fully compatible with the standard.
What it implements
- Transport: Streamable HTTP (the current MCP transport) at
/mcp. - Capability:
tools— nine tools (search, get, debug, context, whoami, write, update, delete, feedback). A clean, valid subset of the protocol. - Authorization: OAuth 2.1 via Auth0, plus a fail-closed subject allowlist (a valid token is necessary but not sufficient) and a stricter allowlist for writes.
- Safety: per-subject sliding-window rate limit, JSONL audit log on writes, a two-step confirm on delete, note size caps, masked internal errors, and a central prompt-injection data-boundary fence on every returned memory string.
See docs/compliance.md for an honest gap analysis against
the spec, and docs/ for architecture, security, tools, and auth.
Quick start (local, unauthenticated)
pip install -e .
soma-mcp # serves Streamable HTTP at http://localhost:8000/mcp
Without AUTH0_* env vars the server runs unauthenticated for local
experimentation (it warns loudly). Set MCP_SINGLE_USER=1 so the fail-closed
guards allow your calls. Point MCP Inspector
at http://localhost:8000/mcp to browse tools/list and fire calls.
Out of the box it uses the InMemoryEngine (substring search, in-memory notes) —
enough to see the protocol working end to end.
Plug in your own memory
Implement the Engine protocol (src/soma_mcp/engine.py) over your own
retrieval stack and inject it:
from soma_mcp import build_server
from my_stack import MyEngine # implements soma_mcp.engine.Engine
mcp = build_server(MyEngine())
mcp.run(transport="http", host="0.0.0.0", port=8000)
The Engine surface is exactly nine methods (search, filter, get, assemble
context, add/update/delete note, log feedback). The tool kernel, guards and
prompt-injection fence are reused unchanged.
Production (authenticated)
Set the Auth0 variables and run behind TLS (e.g. a tunnel). See
docs/auth.md and .env.example.
| Variable | Purpose |
|---|---|
AUTH0_CONFIG_URL, AUTH0_CLIENT_ID, AUTH0_CLIENT_SECRET, AUTH0_AUDIENCE, BASE_URL |
OAuth 2.1 provider. |
MCP_TOEGANG_SUBJECTS |
Comma-separated allowlist of subjects permitted to use the instance. |
MCP_SCHRIJF_SUBJECTS |
Allowlist of subjects permitted to write. |
MCP_SINGLE_USER |
1 to run intentionally without an allowlist. Otherwise guards fail closed. |
MCP_RATE_LIMIT |
Calls per minute per subject (default 30; 0 disables). |
MCP_JWT_SIGNING_KEY, MCP_STORAGE_ENCRYPTION_KEY |
Enable encrypted, persistent token storage (pip install -e .[auth]). |
SOMA_DATA |
Data path (token store, audit log). |
Tests
pip install -e .[dev]
pytest
The security-critical kernel (fail-closed guards, the data-boundary fence, the two-step delete, size caps) is covered without an auth environment — the shell is dumb, the kernel is covered.
Note on language
SOMA is a Dutch-language personal system. The live SOMA server uses Dutch
tool parameter names (vraag, tekst, bevestig, diep, oordeel), as
documented under docs/. This public reference anglicizes them
(question, text, confirm, deep, verdict) for accessibility; the
semantics are identical. A couple of result-dict keys retain their Dutch names
(tekst, bronnen, datum, herkomst) to match the documented surface.
About the standard
MCP is an open standard introduced by Anthropic in 2024. In December 2025 it was contributed to the Agentic AI Foundation (AAIF) under the Linux Foundation, placing it under vendor-neutral governance. This implementation tracks the 2025-11-25 stable specification revision.
- Specification & SDKs: https://github.com/modelcontextprotocol
- Site: https://modelcontextprotocol.io
License
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。