Canon MCP
Serves as the serving layer for Canon governed context repositories, enabling AI agents to query domain knowledge through DAX and SQL queries with user identity via Entra OBO.
README
Canon MCP
The serving layer for Canon governed context repositories. Reads a canon-context-accelerator
repo at runtime and exposes its domain knowledge to AI agents over the MCP protocol, executing
DAX and SQL queries under the user's identity via Entra OBO.
canon-context-accelerator (context layer)
domains/<slug>/*.yaml,*.md ──git read (60s TTL)──► canon-mcp (this repo)
_compiled/_index.json single container
_compiled/bundles/<domain>.md 6 MCP tools
.canon-cache/<d>/<connector_id>/schema.json OBO → Fabric
Architecture invariant: the image carries code, not content. Anything a serving path reads must be readable from git at runtime. Content changes are a push, not a redeploy.
Tools
| Tool | Description |
|---|---|
list_domains |
List all Canon domains from the compiled index |
get_domain_context |
Full grounding bundle for a domain (call before any query) |
query_measures |
Compose and execute SUMMARIZECOLUMNS DAX server-side (governed) |
run_dax |
Raw DAX escape hatch — requires justification, logged |
execute_sql |
Governed SQL fallback against warehouse connectors |
whoami |
Diagnostics: UPN, OBO health, config completeness |
Environment variables
Repo pointer (required for remote mode)
| Variable | Description |
|---|---|
CANON_REPO_PROVIDER |
github or ado — enables remote git reads; omit for local file mode |
CANON_REPO_OWNER |
GitHub org / ADO org |
CANON_REPO_NAME |
Repository name |
CANON_REPO_TOKEN |
PAT with read:contents scope |
CANON_REPO_BRANCH |
Branch to read from (default: main) |
CANON_REPO_ADO_ORG |
ADO org name (ADO only) |
CANON_REPO_ADO_PROJECT |
ADO project name (ADO only) |
Auth (required for HTTP mode)
| Variable | Description |
|---|---|
CANON_AUTH_TENANT_ID |
Entra tenant ID |
CANON_AUTH_CLIENT_ID |
App registration client ID |
CANON_AUTH_CLIENT_SECRET |
Client secret (for OBO token acquisition) |
CANON_MCP_BASE_URL |
Public HTTPS URL of this server (used in OAuth metadata) |
Transport / runtime
| Variable | Description |
|---|---|
CANON_MCP_PORT |
HTTP port (default: 8000) |
CANON_MCP_REFRESH_SECRET |
Shared secret for POST /refresh. Required when CANON_REPO_PROVIDER is set; the server refuses to start without it. |
CANON_REPO_ROOT |
Local path to context repo (local mode only) |
Transport is streamable HTTP only (no stdio, no SSE).
Env vars carry secrets and repo-pointer config only — never topology.
Workspace IDs, dataset IDs, and connector topology come from scan-config.yaml
in the context repo, not from environment variables.
Local mode
CANON_REPO_ROOT=/path/to/canon-context-accelerator \
uv run python -m serving.server
HTTP / Container App mode
CANON_REPO_PROVIDER=github \
CANON_REPO_OWNER=YourOrg \
CANON_REPO_NAME=canon-context-accelerator \
CANON_REPO_TOKEN=ghp_... \
CANON_AUTH_TENANT_ID=... \
CANON_AUTH_CLIENT_ID=... \
CANON_AUTH_CLIENT_SECRET=... \
CANON_MCP_BASE_URL=https://your-container-app.azurecontainerapps.io \
CANON_MCP_REFRESH_SECRET=... \
uv run python -m serving.server
Development
uv sync --extra dev
uv run pytest
uv run ruff check .
CI runs ruff check + pytest on every PR via .github/workflows/test.yml.
Two-repo contract
canon-mcp consumes these artifacts from the context repo (never baked into the image):
| Artifact | Produced by | Used for |
|---|---|---|
_compiled/_index.json |
compile.yml |
list_domains, domain enum |
_compiled/bundles/<domain>.md |
compile.yml |
get_domain_context |
domains/<d>/models/<m>.yaml |
humans + scan | query_measures validation |
domains/<d>/examples.yaml |
humans | few-shot grounding in bundle |
connections/<c>.yaml |
humans | execute_sql allow-list |
scan-config.yaml → domains[].models[], connectors[] |
humans | connector resolution |
.canon-cache/<d>/<connector_id>/{schema,profiles}.json |
scan.yml |
model schema, dim profiles |
Changing any of these shapes is a coordinated two-repo migration.
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。