ContextStream
Gives coding agents persistent memory across sessions and tools: semantic code search plus saved decisions, lessons, and plans, so you stop re-explaining project context every new chat.
Overview
ContextStream connects your code, docs, and conversations so an agent can find the right files, recall saved decisions, and build on past work across sessions and tools instead of starting from zero each time. The MCP server/client in this repo is MIT-licensed; the hosted search backend is a separate ContextStream cloud service (free tier: 10,000 monthly credits, no card required).
How it works
- Indexing sends eligible project files to ContextStream's hosted search, so the agent runs semantic code search instead of grepping the repo blind.
- The agent can save a real decision, and the reason behind it, mid-session. A later session, even in a different editor, can then retrieve that decision without you re-explaining it.
- ContextStream can also capture local Git metadata (commit SHAs, branch names, aggregate diff stats) to enrich context. Transcript saving and hook-based capture are on by default and can be turned off in the data-handling settings.
- It ships as a native Rust binary through an npm launcher (
npx -y @contextstream/mcp-server@latest), or you can connect straight to the hosted MCP endpoint for clients that support Streamable HTTP and OAuth.
Examples
Use ContextStream to find the files relevant to my next change. Cite the
sources and retrieve any saved project decisions that should guide the work.
Installation
You need an MCP-capable client (Claude Code, Cursor, Copilot, Windsurf, etc.) and a free ContextStream account, created during onboarding with no separate website signup needed.
- Run the install script for your OS (macOS/Linux or Windows PowerShell, above) and follow the onboarding prompts to sign in and connect your editor.
- Restart your editor after setup.
- Ask your agent to find files relevant to a change, then ask it to save a decision. Start a new session and confirm it can recall that decision.
- Review the data handling and controls docs before connecting a sensitive or private project, since indexing sends file contents to the hosted service by default.
Related assets
Awesome MCP Servers
The largest community catalog of MCP servers: thousands of entries across dozens of categories (databases, browser automation, cloud, communication and more), tagged by language and scope.
Context7
Pulls updated, version-specific documentation and code examples straight from the source into your prompt, so agents stop citing outdated APIs or hallucinating ones that don't exist.
Playwright MCP
Microsoft's official MCP server for browser automation, driving Playwright through structured accessibility snapshots instead of screenshots, so any MCP client can browse, click and fill forms without a vision model.