bichon-mcp
MCP server that enables AI agents to search and read locally-archived emails via the Bichon email archiver API, without exposing Gmail or IMAP credentials.
README
bichon-mcp
MCP server that wraps the Bichon email archiver REST API, letting an AI agent (Claude Code, Claude Desktop) search and read locally-archived emails — without ever touching Gmail or IMAP credentials directly.
Architecture
Gmail / IMAP
└─► Bichon (Docker, localhost:15630)
local Tantivy FTS + compressed EML store
└─► bichon-mcp (this repo, stdio MCP server)
└─► Claude Code / Claude Desktop
After the initial IMAP sync, the AI only ever calls localhost:15630. No external auth is granted to the AI.
Requirements
- Python 3.11+
- uv (recommended) or pip
- Bichon running locally — see Bichon quickstart
Installation
1. Clone and install
git clone https://github.com/pras-labs/bichon-mcp
cd bichon-mcp
uv venv
uv pip install -e .
2. Get a Bichon access token
Start Bichon (Docker):
docker run -d --name bichon -p 15630:15630 rustmailer/bichon:latest
Then get a token:
curl -s -X POST http://localhost:15630/api/login \
-H "Content-Type: application/json" \
-d '{"username":"admin","password":"admin@bichon"}' \
| python3 -c "import sys,json; print(json.load(sys.stdin)['access_token'])"
For a long-lived token, create one in the Bichon WebUI under Settings → Access Tokens.
3. Configure your client
Claude Code
Copy .mcp.json.example to .mcp.json and fill in the token:
cp .mcp.json.example .mcp.json
{
"mcpServers": {
"bichon-email": {
"command": ".venv/bin/python",
"args": ["-m", "bichon_mcp"],
"env": {
"BICHON_BASE_URL": "http://localhost:15630",
"BICHON_ACCESS_TOKEN": "<your-token>"
}
}
}
}
.mcp.jsonis gitignored — your token will not be committed.
Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):
{
"mcpServers": {
"bichon-email": {
"command": "/path/to/bichon-mcp/.venv/bin/python",
"args": ["-m", "bichon_mcp"],
"env": {
"BICHON_BASE_URL": "http://localhost:15630",
"BICHON_ACCESS_TOKEN": "<your-token>"
}
}
}
}
Configuration
All configuration is via environment variables:
| Variable | Default | Description |
|---|---|---|
BICHON_BASE_URL |
http://localhost:15630 |
Bichon instance URL |
BICHON_ACCESS_TOKEN |
(required) | Bearer token for Bichon API |
BICHON_STATS_CACHE_TTL |
300 |
Mailbox stats cache TTL in seconds. 0 = disabled |
BICHON_REQUEST_TIMEOUT |
30 |
HTTP request timeout in seconds |
MCP Tools
list_accounts()
List all email accounts configured in Bichon.
Returns: [{id, email}]
Example prompt: "What email accounts are set up?"
list_mailboxes(account_id)
List mailbox folders for an account (INBOX, Sent, Drafts, etc.).
account_id* integer From list_accounts
Returns: [{id, name, total, unseen}]
Example prompt: "Show me all mailboxes for my account."
get_mailbox_stats(account_id, mailbox_id)
Per-mailbox statistics: total email count, top senders, top subjects by frequency. Results are cached (default 5 min, configurable via BICHON_STATS_CACHE_TTL).
account_id* integer From list_accounts
mailbox_id* integer From list_mailboxes
Returns:
total_emails integer Exact count from Bichon
sampled integer Emails analysed (up to 200 most recent)
top_senders [{sender, count}]
top_subjects [{subject, count}]
oldest_in_sample / newest_in_sample integer (unix ms)
cached boolean
Example prompt: "Give me stats for my INBOX — who emails me most?"
search_emails(query, ...)
Full-text search across archived emails. Returns summaries only — no body text.
query* string Full-text search term (max 500 chars)
date_from string ISO date YYYY-MM-DD (inclusive)
date_to string ISO date YYYY-MM-DD (inclusive)
sender string Filter by sender address (max 254 chars)
account_id integer Scope to one account
mailbox_id integer Scope to one mailbox folder
limit integer Max results, default 20, max 50
Returns: [{id, account_id, subject, from, to, date, preview, thread_id, has_attachments}]
Example prompts:
- "Search for emails about invoices from last month."
- "Find emails from notifications@github.com in my INBOX."
get_email(account_id, envelope_id)
Fetch the full plain-text content of one email. HTML is stripped. Body is truncated at ~4000 tokens.
account_id* integer From search_emails result
envelope_id* string UUID from search_emails result
Returns: {account_id, envelope_id, body, attachment_count}
Example prompt: "Show me the full content of that invoice email."
list_threads(subject_contains?, sender?, limit?)
Group search results by thread_id to show conversation threads.
subject_contains string Filter by subject text (max 500 chars)
sender string Filter by sender (max 254 chars)
limit integer Max threads to return, default 10
Returns: [{thread_id, account_id, subject, latest_from, latest_date, count}]
Example prompt: "Show me all threads from GitHub notifications."
get_sender_summary(email_address)
Aggregate stats for a sender: total email count, date range, top subjects.
email_address* string Valid email address (max 254 chars)
Returns: {email, count, first_seen, last_seen, top_subjects}
Example prompt: "How many emails have I received from packt@mail.packtpub.com and what are they about?"
Typical workflow
User: "Summarise my email activity for the past month."
Claude:
1. list_accounts() → finds account IDs
2. list_mailboxes(account_id) → finds INBOX mailbox_id
3. get_mailbox_stats(...) → top senders, subjects, total count
4. search_emails(query="", date_from="2026-04-20", ...) → recent emails
5. get_email(...) × N → reads relevant messages
→ synthesises a summary
Security notes
- The access token is read from
BICHON_ACCESS_TOKENat startup — never hardcode it in source files. .mcp.jsonis gitignored to prevent accidental token commits.envelope_idvalues are validated as UUIDs before use in API paths.- All string inputs have length limits; integer IDs must be positive.
- API error responses are not forwarded to the AI (only the HTTP status code is).
- Email bodies are passed directly to Claude. Malicious emails may contain text designed to manipulate AI behaviour — a known limitation of any email-reading AI tool.
Development
# Install in editable mode
uv pip install -e .
# Verify MCP tools are registered (no Bichon needed)
python3 -c "
import subprocess, json, os, threading, queue, time
proc = subprocess.Popen(['.venv/bin/python', '-m', 'bichon_mcp'],
stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL,
env={**os.environ, 'BICHON_BASE_URL':'http://localhost:15630','BICHON_ACCESS_TOKEN':'test'},
text=True, bufsize=1)
q = queue.Queue()
threading.Thread(target=lambda: [q.put(l.strip()) for l in proc.stdout], daemon=True).start()
for msg in [
{'jsonrpc':'2.0','id':1,'method':'initialize','params':{'protocolVersion':'2024-11-05','capabilities':{},'clientInfo':{'name':'t','version':'0'}}},
{'jsonrpc':'2.0','method':'notifications/initialized','params':{}},
{'jsonrpc':'2.0','id':2,'method':'tools/list','params':{}},
]:
proc.stdin.write(json.dumps(msg)+'\n')
proc.stdin.flush()
deadline = time.time()+10
while time.time()<deadline:
try:
line = q.get(timeout=1)
obj = json.loads(line)
if obj.get('id')==2:
print([t['name'] for t in obj['result']['tools']])
break
except: pass
proc.terminate()
"
# Expected: ['list_accounts', 'list_mailboxes', 'get_mailbox_stats', 'search_emails', 'get_email', 'list_threads', 'get_sender_summary']
License
AGPL-3.0. If you distribute this software, the source must remain open.
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。
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。
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。
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。
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。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。