ContextForge
A local MCP server for managing engineering context across Components, Repos, Tasks, and Governance entities. It enables capturing reusable context and composing it per-task with typed relationships and cross-cutting guidelines.
README
ContextForge
A local MCP server for managing engineering context across Components, Repos, Tasks, and Governance entities.
Built for the cross-repo reality: a single task touches the UI repo + the API repo + the gateway + external service X + database Y. ContextForge lets you capture reusable context once and compose it per-task, with typed relationships and cross-cutting guidelines.
Model
- Component — a coherent unit of functionality. Kinds:
system | service | api | database | library | tool. Can declare ausesdependency graph over other components. - Repo — a codebase. Optionally belongs to a parent Component (e.g.
gateway-proxy-repo → component:api-gateway). - Task — a transient unit of work that composes context by linking specific Component/Repo/Governance refs. Can carry
external_refsto Jira/GitLab/Confluence/etc. - Governance — reusable cross-cutting guidelines (security policy, API design standards, user prefs, branding). Any other entity can reference them; they cascade into task packs automatically.
Ref grammar: type:slug (entity) or type:slug/subtopic (one document within an entity).
Examples: component:api-gateway, component:api-gateway/auth, repo:ui-repo/testing, task:fix-login-redirect, governance:security-policy/data-handling.
Storage
All data lives under ~/.contextforge/ (override with CONTEXTFORGE_HOME):
~/.contextforge/
components/{slug}/_meta.md, {subtopic}.md
repos/{slug}/_meta.md, {subtopic}.md
tasks/{slug}/_meta.md, {subtopic}.md
governance/{slug}/_meta.md, {subtopic}.md
.index/contextforge.db # SQLite + FTS5, rebuildable via `reindex`
config.json # always_include refs + workspace bindings
logs/contextforge.log # persistent usage + debug (rotating)
Markdown files are the source of truth. The SQLite index is derived — if it drifts, call reindex.
Storage opens one SQLite connection per thread (threading.local()) with PRAGMA journal_mode=WAL so FastMCP's worker threads can each call in safely. This is load-bearing — don't cache a Connection on the instance or remove the WAL pragma.
New entities get scaffolded subtopics:
- Components:
overview.md,auth.md,integration.md,gotchas.md - Repos:
overview.md,style.md,testing.md,local-dev.md - Tasks:
goal.md,plan.md,notes.md - Governance:
overview.md,rules.md
Delete the ones you don't want.
Install & run
uv is the expected package manager (install via brew install uv or Astral's installer).
cd /path/to/context-forge
uv sync
uv run contextforge # stdio MCP server
Tests (pytest is a dev dependency):
uv sync
uv run pytest
Behind an SSL-inspecting proxy? If
uv syncfails withinvalid peer certificate: UnknownIssuer, add--system-certsto theuvcommands (orexport UV_SYSTEM_CERTS=1).
Logs go to stderr and to ~/.contextforge/logs/contextforge.log (rotating text file, on by default). The file is the durable record for usage analysis and trials.
Default level is INFO (pack assembly summaries, creates, links, governance, resolves, etc.). Set CONTEXTFORGE_LOG_LEVEL=DEBUG for full FTS scoring breakdowns per entity — useful for tuning per_entity_top_k and min_score.
Use the tail_logs(n) tool to inspect recent activity from inside the MCP.
For development / dogfooding where you want the same events written to a second location (e.g. inside the source repo so logs travel with the checkout), set:
CONTEXTFORGE_DEV_LOG_DIR=/path/to/desired/dir(recommended), orCONTEXTFORGE_LOG_FILE=/path/to/specific.log
You get both the normal home log and the extra location.
Wire into Cursor
Install it globally, not per-project. ContextForge is a single global store
(~/.contextforge/) meant to serve every repo you work in. The --directory
below only tells uv where this server's code lives — the running server reads
~/.contextforge/ regardless of which workspace is open, so you never need to add
the context-forge folder to your other repos. Wire it once, globally, and it's
available everywhere.
1. Register the MCP server globally
Add it to your global Cursor MCP config (~/.cursor/mcp.json, or your
mcp-manager proxy config if you use one), replacing the placeholder with this
repo's path on your machine:
{
"mcpServers": {
"contextforge": {
"command": "uv",
"args": ["run", "--directory", "/ABSOLUTE/PATH/TO/context-forge", "contextforge"]
}
}
}
Behind an SSL-inspecting proxy, add "--system-certs" as the first entry in args,
or set UV_SYSTEM_CERTS=1 on the server's env. Restart Cursor (or toggle the
server under Settings → MCP) and confirm the contextforge tools appear.
A project-scoped .cursor/mcp.json is also committed in this repo, but it only activates when this repo is the open workspace — fine for hacking on ContextForge itself, not for using it across your work.
2. Make the router rule global too
The MCP registration gives the agent the tools; the .cursor/rules/contextforge-router.mdc rule gives it the behavior (when to recall and capture context proactively). A rule that lives only in this repo won't apply in your other repos — so it must be global as well. Two equivalent options:
- Symlink the rule into your global Cursor rules location (single source of
truth — edits here propagate automatically). On Windows this needs Developer Mode
or an elevated
mklink. - Copy the rule into your global Cursor rules location (more portable; re-copy after edits).
Either way the rule is alwaysApply: true, so once it's global it's in context for
every session — that's what lets the agent use ContextForge without being told to.
Tools
Entity management
create_component(slug, description?, kind?, aliases?, uses?, governance?)create_repo(slug, description?, component?, aliases?, governance?)create_task(slug, description?, aliases?, governance?)create_governance(slug, description?, aliases?)list_components(kind?),list_repos(),list_tasks(),list_governance_entities()delete_entity(ref)— destructive, cascades
Context CRUD
get_context(ref)— entity or single subtopicupsert_context(ref, content)— create/overwriteappend_context(ref, content)— append (creates if missing)delete_context(ref)— delete subtopic
Aliases + resolution
add_alias(ref, alias)— globally unique nicknamesremove_alias(ref, alias)resolve_ref(query, type_filter?)— fuzzy lookup against slugs/aliases/names
Task composition
link_task(task_slug, refs)/unlink_task(task_slug, refs)get_task_pack(task_slug, include_always?, focus?, per_entity_top_k?, min_score?)— assembled pack (task + governance + links). Withfocus=True(default), entity-level refs with more subtopics thanper_entity_top_k(default 3) are FTS-narrowed against the task's description + notes — only the most relevant subtopics are included. Filtered refs are listed indroppedso the agent can pull them on demand. Passfocus=Falseto get the full unfiltered pack.suggest_task_links(task_slug, anchors?, depth?, top_k?)— graph + FTS suggestions
Governance
add_governance(ref, governance_ref)— attach guideline to component/repo/taskremove_governance(ref, governance_ref)
External refs (Jira/GitLab/Confluence/etc.)
add_external_ref(task_slug, system, id, url?)remove_external_ref(task_slug, system, id)
Imported content (from Confluence, Jira, GitLab, etc. via other MCPs in Cursor)
import_content(ref, content, source_url, source_name?)— store snapshot with source trackingrefresh_source(ref)— return the source_url so the caller can re-fetchlist_stale_sources(older_than_iso)
Search + config
search(query, entity_type?, limit?)— FTS5get_config()add_always_include(ref)/remove_always_include(ref)— refs auto-included in every task packbind_workspace(path, repo_slug)/unbind_workspace(path)/get_current_workspace(path)
Maintenance
reindex()— rebuild SQLite from disktail_logs(n=50)— recent lines from the persistent log (for usage inspection)
Resources
context://component/{slug}/context://repo/{slug}/context://task/{slug}/context://governance/{slug}— full entity as markdowncontext://{type}/{slug}/{subtopic}— single subtopicpack://task/{slug}— assembled task pack (separate scheme avoids colliding withtask/{slug}/{subtopic})
Smoke test
After wiring it into Cursor, ask the agent to run these (or call them from the MCP panel):
create_component(slug="api-gateway", kind="system", description="Our edge proxy")upsert_context(ref="component:api-gateway/auth", content="OAuth2 with PKCE; tokens rotate hourly.")create_governance(slug="security-policy", description="Company-wide security baseline")add_governance(ref="component:api-gateway", governance_ref="governance:security-policy")create_task(slug="demo-task", description="Kicking the tires")link_task(task_slug="demo-task", refs=["component:api-gateway/auth"])get_task_pack(task_slug="demo-task")— you should see the auth subtopic AND the security-policy content (cascaded through the governance link on the component).
License
MIT. Runtime deps are permissive (MIT/BSD/Apache-2.0) — no copyleft.
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。