Context Bridge
Enables repository context packing, file/symbol ranking, task-relevant selection, and applying structured patches via MCP.
README
Context Bridge
Context Bridge is a local-first pipeline between a repository and external LLMs or MCP clients. It is intentionally narrow: it packs bounded repository context, ranks files and symbols with a tree-sitter/PageRank graph, applies structured SemPatch YAML through AST matching, and writes changes through a transaction layer with verification hooks.
It is not a full coding agent. It does not plan large refactors end to end. It focuses on the I/O boundary: what leaves the repo, and what comes back in.
What It Does
- pack repository context with manifests, byte limits, secret scanning, and optional visual snapshots,
- build repository maps from tree-sitter tags and reference scores,
- select task-relevant files within a token budget,
- apply SemPatch YAML with transactional prepare/commit semantics,
- expose the same primitives through CLI, MCP, and an agent skill,
- pair with a browser companion so web LLMs can operate on local code without copy/paste loops.
Install
python -m pip install -e .
python -m pip install -e ".[mcp,visual]"
Quick Start
Generate a repository map:
context-bridge map . --query "fix login token refresh" -o repo-map.md
Select a bounded bundle:
context-bridge select . --task "fix login token refresh" -o selected-context.md
Create a staged plan:
context-bridge playbook . --task "fix login token refresh" -o playbook.md
Apply a SemPatch response:
context-bridge apply answer.md --root .
context-bridge apply answer.md --root . --write
Migrate a legacy patch file once:
context-bridge migrate-old-patches legacy-answer.md -o answer.sempatch.md
SemPatch
SemPatch is the required patch format. Each block is explicit YAML:
--- sempatch ---
file: src/example.py
language: python
op: replace
selector:
rule:
pattern: |
def hello():
$$$BODY
replacement: |
def hello():
return "updated"
--- end sempatch ---
Use op: create_file with content: for new files. The tool validates the structure, matches
exactly one AST node, and runs inside a transaction when writing.
MCP
context-bridge mcp
The MCP server exposes pack_context, repo_map, context_playbook, select_context_for_task,
scan_text_for_secrets, inspect_sempatches, apply_patches, and sempatch_schema.
Browser Companion
The browser companion follows the same shape as existing Chrome/MCP bridge projects: prefer Chrome Native Messaging for the local host boundary, then fall back to the loopback HTTP bridge when native host registration is not available.
context-bridge ext native-manifest --extension-id <installed-extension-id> --output native-host.json
context-bridge ext serve
The extension/ directory is a Chrome MV3 companion. It calls the native host
com.context_bridge.native_host first and uses context-bridge ext serve only as a fallback.
Verification
Configure verify hooks in .context-bridge.toml:
[verify]
typescript = ["pnpm", "exec", "tsc", "--noEmit"]
python = ["python", "-m", "ruff", "check", "--quiet"]
rust = ["cargo", "check", "--quiet"]
apply --write runs verification by default. Use --no-verify only when you knowingly want
to skip the hooks.
Visual Snapshots
bundle --visual does not ship a custom component renderer. It reuses an existing
Storybook, Ladle, or Histoire setup in the target project and captures it with Playwright:
[visual]
explorer = "storybook"
command = ["npm", "run", "storybook", "--", "--ci", "--port", "6006"]
url = "http://127.0.0.1:6006"
timeout_seconds = 45
If no story exists for a component, Context Bridge reports that clearly instead of emitting a fake snapshot.
Development
python -m unittest discover -s tests -v
python -m compileall -q src tests
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。