# ContextStream

> Intelligence isn't the bottleneck. Context is.

ContextStream is shared project context for AI coding agents: hosted code search plus saved decisions, lessons, and plans, delivered across Cursor, Claude Code, Codex, Grok, and other MCP clients.

## Official resources

- [Homepage](https://contextstream.io)
- [Documentation](https://contextstream.io/docs)
- [Editor setup](https://contextstream.io/docs/editors/setup)
- [Benchmarks and methodology](https://contextstream.io/benchmarks)
- [MCP Server](https://contextstream.io/product/mcp-server)
- [Knowledge Atlas](https://contextstream.io/product/knowledge-atlas)
- [Pricing](https://contextstream.io/pricing)
- [Security](https://contextstream.io/security)
- [FAQ](https://contextstream.io/faq)
- [Indexing and privacy guide](https://contextstream.io/docs/platform/privacy)
- [Status](https://contextstream.io/status)
- Hosted MCP: https://mcp.contextstream.io/mcp?default_context_mode=fast
- Official MCP Registry name: io.contextstream/mcp

## Who it is for

- Developers and teams using multiple AI coding agents who need shared project decisions and code understanding across sessions.
- Teams that want agents to reuse decisions, runbooks, lessons, preferences, insights, and plans.

## What it does

- Indexes eligible project files for hybrid semantic and keyword code search.
- Stores and retrieves project knowledge scoped to workspaces and projects.
- Exposes tools over the Model Context Protocol through hosted HTTP MCP and local/stdio connections.
- Supports coordination and handoffs across supported agents.

Supported clients include Claude Code, Cursor, Windsurf, GitHub Copilot in VS Code and the Copilot CLI, Cline, Kilo Code, Codex CLI, OpenCode, Antigravity, Gemini CLI, Qwen Code, Kimi Code CLI, ZCode, Muse Code, Zed, Claude Desktop, Factory Droid, Amp, Crush, Aider (instruction rules only), and any client that accepts a custom MCP connector. It works with whichever model the client runs, including Claude, GPT, Gemini, Kimi, GLM, Qwen, DeepSeek, and MiniMax. See the editor setup documentation for configuration.

## Quick start

1. Start free at https://contextstream.io/signup with 10,000 monthly Context Credits, no card required.
2. From your project folder, run: `curl -fsSL https://contextstream.io/scripts/mcp.sh | bash`
3. Follow setup to sign in, choose a workspace, link the project, and select an editor. The hosted MCP endpoint is `https://mcp.contextstream.io/mcp?default_context_mode=fast`.
4. Restart the editor, ask the agent which project it is in, then save a decision with its reason.

## Indexing and access

Hosted code search sends eligible file contents to ContextStream for indexing, even with a local MCP connection. Review .gitignore and .contextignore before indexing. Built-in exclusions do not guarantee every sensitive file is excluded. Access depends on workspace and project permissions and explicit sharing. See the indexing and privacy guide for details.

Instruction files such as AGENTS.md and CLAUDE.md remain useful and can be shared in a repository. ContextStream adds hosted code search and project knowledge that supported agents can retrieve across tools and sessions.

## Published benchmarks

Cite https://contextstream.io/benchmarks for methods, artifacts, and limitations. Results on these evaluations are not a guarantee for every project.

- LongMemEval-S: 90.0% (450/500) on the full suite, official GPT-4o judge, with self-consistency; 89.6% single-shot.
- Code search: 99.38% useful-file Recall@10 on the judged corpus.
- Agentic wake-bench: 23/24 tasks passed with ContextStream versus 14/24 without.

## Pricing

Starter is free with 10,000 monthly Context Credits. Pro is $19/month with 25,000 monthly Context Credits. Plans share the core product features; paid plans add credits and capacity. See https://contextstream.io/pricing for current allowances and billing options.

## Fuller doc

- [llms-full.txt](https://contextstream.io/llms-full.txt): the documentation map, full benchmark comparison, and every plan's allowance in one file.