Questions, answered.
What ContextStream is, how it fits next to your instruction files, how to connect it, and what happens to your code. Stated plainly, including the limits.
On this page
What is ContextStream?
ContextStream is shared project context for AI coding agents. It combines hosted code search with saved project knowledge—decisions, lessons, preferences, insights, and plans—and delivers relevant context to Cursor, Claude Code, Codex, Grok, and other MCP clients so agents build on past work across sessions.
How is ContextStream different from Agents.md or CLAUDE.md files?
Instruction files guide agents inside their tools and can be shared in a repository. ContextStream adds project-scoped code search and saved decisions, lessons, plans, and runbooks that supported agents can retrieve across tools and sessions via MCP. Teams can use instruction files and ContextStream together.
How do I connect ContextStream to Cursor, Claude Code, or Codex?
Start free at contextstream.io, then run curl -fsSL https://contextstream.io/scripts/mcp.sh | bash from your project folder, or add the hosted MCP URL https://mcp.contextstream.io/mcp?default_context_mode=fast in your MCP client. Setup walks you through signing in, choosing a workspace, linking the project, and selecting your supported editor, configures the connection and instruction files, then starts indexing in the background. Restart your editor afterward, then ask which project you’re in.
Where does the first context come from?
Your indexed project files provide the initial code context. Saved decisions, lessons, plans, and documents add the reasoning behind the work. Ask your agent to record important knowledge explicitly—indexing cannot recover every undocumented decision.
Is this another documentation job?
Your connected agent can save decisions, lessons, and plans while you work. Ask it to record important outcomes and update them when things change. Capture depends on your workflow; not every conversation becomes durable knowledge automatically.
Can I start on my own?
Yes. Start with one project and one supported agent. Your project knowledge can follow you across sessions and supported editors. Invite authorized collaborators when you need shared access.
Does my code leave my machine?
Yes. Code search sends eligible file contents to ContextStream for hosted indexing, even with a local MCP connection. Review .gitignore and .contextignore before indexing. Built-in exclusions help, but they do not guarantee that every sensitive file is excluded.
Who can see the context?
Access depends on the record’s scope, workspace and project permissions, and any explicit sharing. Personal records are not automatically team-visible, and connected services retain their own account permissions. Review membership and external shares before adding sensitive material.
Can I remove stored context?
Yes, but stored records, transcripts, and indexed source have separate removal controls. Turning transcript capture off stops future capture; it does not delete existing history. Review saved decisions and other records separately, and remember that removal cannot retract copies already shared elsewhere.
What do the ContextStream benchmarks show?
The published benchmarks report 90.0% on the full LongMemEval-S suite (450/500, official GPT-4o judge, with self-consistency; 89.6% single-shot), 99.38% useful-file Recall@10 on the judged code-search corpus, and 23/24 agentic wake-bench tasks passed with ContextStream versus 14/24 without. These are results on the published evaluations, not a guarantee for every project. See the benchmarks page for methodology and limits.