Chroma Memory MCP
An MCP server that gives Claude (and any MCP client) a shared team knowledge base with semantic search. Backed by ChromaDB and Google Gemini embeddings — works across languages (RU + EN).
README
Chroma Memory MCP
An MCP server that gives Claude (and any MCP client) a shared team knowledge base with semantic search. Backed by ChromaDB and Google Gemini embeddings — works across languages (RU + EN).
How It Works
The server stores team knowledge as entries in ChromaDB. Each entry has a project, slug, title, content (Markdown), tags, type, and author. When someone asks a question, Claude runs a semantic search and answers with full context. When someone makes a decision worth sharing, they tell Claude to save it — and it's available to every team member.
Claude ──MCP──▶ chroma-memory-mcp ──▶ ChromaDB
│
▼
Gemini Embeddings
(multilingual RU+EN)
Quick Start
1. Start the server
git clone https://github.com/sfrangulov/chroma-memory-mcp
cd chroma-memory-mcp
export GOOGLE_API_KEY=your-gemini-api-key
docker compose up -d
The server is now available at http://localhost:3000/mcp.
2. Connect Claude Code
Add .mcp.json to your project root:
{
"mcpServers": {
"chroma-memory": {
"type": "http",
"url": "http://localhost:3000/mcp"
}
}
}
3. Connect Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"chroma-memory": {
"type": "http",
"url": "http://localhost:3000/mcp"
}
}
}
MCP Tools
| Tool | Description | Key params |
|---|---|---|
write_entry |
Create a new memory entry | project, slug, title, content, tags?, type? |
read_entry |
Read entry by project + slug | project, slug |
update_entry |
Update an existing entry | project, slug, + fields to change |
delete_entry |
Delete an entry | project, slug |
search |
Semantic search across entries | query, project?, author?, n_results? |
list_entries |
List entries with filters | project?, author?, type? |
list_projects |
List all project names | — |
Entry types
note(default) — general knowledgedecision— architectural or process decisionssnippet— reusable code fragmentsdoc— documentation and reference materiallog— event logs and session records
Entry ID format
{project}:{slug} — for example mobile-app:auth-decision.
Configuration
| Variable | Required | Default | Description |
|---|---|---|---|
GOOGLE_API_KEY |
Yes | — | Gemini API key for embeddings |
CHROMA_URL |
No | http://localhost:8000 |
ChromaDB connection URL |
CHROMA_COLLECTION |
No | memories |
ChromaDB collection name |
MCP_PORT |
No | 3000 |
Server port |
MCP_HOST |
No | 0.0.0.0 |
Server bind address |
MCP_BASE_URL |
No* | — | Public HTTPS URL (required for OAuth) |
GOOGLE_CLIENT_ID |
No* | — | Google OAuth client ID |
GOOGLE_CLIENT_SECRET |
No* | — | Google OAuth client secret |
REDIS_URL |
No | — | Redis URL for session storage (enables replicas > 1) |
*OAuth is optional. Without it, the server runs in dev mode (no auth).
Without GOOGLE_API_KEY, semantic search is disabled (CRUD still works).
Deployment
Local (Docker Compose)
export GOOGLE_API_KEY=your-gemini-api-key
docker compose up -d
This starts ChromaDB + Redis + MCP server. Data persists in Docker volumes.
Production (Docker Compose + OAuth)
export GOOGLE_API_KEY=your-gemini-api-key
export GOOGLE_CLIENT_ID=your-client-id
export GOOGLE_CLIENT_SECRET=your-client-secret
export MCP_BASE_URL=https://memory.example.com
docker compose up -d
With OAuth enabled, each user authenticates via Google — their email becomes the author field on entries.
Kubernetes (Helm)
A Helm chart is included in helm/chroma-memory-mcp/. It deploys:
- MCP server (Node.js) — Deployment + Service
- ChromaDB — Deployment + PVC + Service
- Ingress with TLS (cert-manager + Let's Encrypt)
- Secrets for API keys
Prerequisites
- Kubernetes cluster with nginx ingress controller
- cert-manager with a ClusterIssuer (for TLS)
- DNS record pointing your domain to the cluster
Install
helm install chroma-memory ./helm/chroma-memory-mcp \
--set secrets.googleApiKey=YOUR_GEMINI_API_KEY \
--set secrets.googleClientId=YOUR_GOOGLE_CLIENT_ID \
--set secrets.googleClientSecret=YOUR_GOOGLE_CLIENT_SECRET \
--set ingress.host=memory.example.com
Key Helm values
# Namespace for all resources
namespace: chroma-memory
# MCP server
mcp:
image:
repository: sfrangulov/chroma-memory-mcp
tag: "0.1.2"
replicas: 1
port: 3000
# ChromaDB
chromadb:
image:
repository: chromadb/chroma
tag: "1.5.2"
persistence:
size: 5Gi
storageClass: microk8s-hostpath
# Ingress
ingress:
enabled: true
className: public
host: memory.example.com
tls:
enabled: true
clusterIssuer: lets-encrypt
Upgrade
# Build new Docker image
docker build --platform linux/amd64 -t sfrangulov/chroma-memory-mcp:0.x.x .
docker push sfrangulov/chroma-memory-mcp:0.x.x
# Update Helm release
helm upgrade chroma-memory ./helm/chroma-memory-mcp \
--set mcp.image.tag=0.x.x \
--reuse-values
Nginx ingress notes
The Helm chart configures nginx annotations for MCP streaming:
proxy-buffering: off— required for Streamable HTTP transportproxy-read-timeout: 3600— long-lived connectionsproxy-body-size: 16m— large entries
Building the Docker image
# For local use
docker build -t chroma-memory-mcp .
# For Kubernetes (must be linux/amd64)
docker build --platform linux/amd64 -t sfrangulov/chroma-memory-mcp:0.x.x .
docker push sfrangulov/chroma-memory-mcp:0.x.x
Authentication
For production deployments, enable Google OAuth2:
- Go to Google Cloud Console → APIs & Services → Credentials
- Create an OAuth 2.0 Client ID (Web application type)
- Add authorized redirect URI:
https://your-domain.com/oauth/google/callback - Enable the Generative Language API (for Gemini embeddings)
- Create an API Key (restrict to Generative Language API)
- Set
MCP_BASE_URL,GOOGLE_CLIENT_ID,GOOGLE_CLIENT_SECRET, andGOOGLE_API_KEY
The server automatically exposes OAuth discovery endpoints:
/.well-known/oauth-authorization-server/.well-known/oauth-protected-resource
Without OAuth (dev mode), the server accepts all requests and sets author to anonymous.
OAuth Flow Details
- Tokens are opaque UUIDs issued by this server (not Google JWTs)
- Token TTL: 24 hours — users re-authenticate daily
- Refresh tokens are not supported — sessions expire after 24h
- Google ID tokens are cryptographically verified via JWKS
- Server restart invalidates all sessions (unless Redis is configured)
HTTP Endpoints
| Endpoint | Method | Description |
|---|---|---|
/mcp |
POST |
Main MCP endpoint (tool calls) |
/mcp |
GET |
SSE stream (server notifications) |
/mcp |
DELETE |
Session cleanup |
/health |
GET |
Health check ({ status: "ok" }) |
Usage Examples
Save a decision
"Remember for the team: we chose PostgreSQL over MongoDB for ACID compliance."
Claude calls write_entry with project, slug, title, content, and tags.
Search memory
"What did we decide about authentication?"
Claude calls search with the query, reviews results, and answers with context.
Browse a project
"What's in our project memory?"
Claude calls list_projects, then list_entries for the relevant project.
Tech Stack
| Dependency | Purpose |
|---|---|
@modelcontextprotocol/sdk |
MCP server framework (Streamable HTTP) |
chromadb v3 |
Vector database client |
@chroma-core/google-gemini |
Gemini embedding function |
express v5 |
HTTP server |
jose |
Google JWT verification (JWKS) |
helmet |
Security headers |
express-rate-limit |
Rate limiting |
ioredis |
Redis client (optional, for scaling) |
zod |
Input schema validation |
Project Structure
├── server.js # MCP server + Express app
├── lib/
│ ├── memory-store.js # ChromaDB wrapper (CRUD + search)
│ ├── auth.js # Email extraction from auth info
│ ├── oauth-provider.js # Google OAuth2 provider
│ └── session-store.js # TTL session store (Memory/Redis)
├── test/ # Unit + integration tests (Vitest)
├── Dockerfile # Production image (node:20-slim)
├── docker-compose.yml # Local development
├── docker-compose.test.yml # Integration test environment
├── helm/ # Kubernetes Helm chart
│ └── chroma-memory-mcp/
│ ├── Chart.yaml
│ ├── values.yaml
│ └── templates/
└── SKILL.md # Claude skill for using the MCP tools
Development
npm install
# Run unit tests
npm test
# Run integration tests (requires Docker)
docker compose -f docker-compose.test.yml up -d --wait
npm run test:integration
docker compose -f docker-compose.test.yml down
Set TEST_CHROMA_URL to override the default http://localhost:8100 for integration tests.
License
MIT
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。
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。
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。
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。