mcp-server

Purpose

ContextStream MCP Server is used to provide AI coding agents with persistent memory, semantic search, and shared project context. Its official positioning is: agents understand the code itself, while this tool lets agents understand the decisions behind the code. It connects code, documentation, and conversations, enabling agents to find the right files, recall saved decisions, and continue building on past work across different sessions and tools.

Main Capabilities

• Persistent memory: carries decisions, lessons learned, and intent across sessions and tools.
• Semantic code search: in officially published benchmarks, the median successful code search response is 95 milliseconds, and it states that this is not a latency guarantee.
• Traceable sources and scoped access: the official materials mention traceable sources and scoped access.
• Shared project context: connects code, documentation, and conversations for use by agents.
• Agents can retrieve relevant files and cite sources, and can also save real decisions and their reasons for retrieval again in new sessions.
• Data control: indexing sends eligible source code content to ContextStream for hosted search; transcript saving and local Git metadata collection are enabled by default and can be turned off.

How to Use

Supports macOS, Linux, and Windows PowerShell. Installation methods include:

• npm (requires Node.js 20+): the npm launcher downloads the native Rust binary for the corresponding platform and verifies the SHA-256 checksum; for production automation, it is recommended to pin an exact package version.
• Hosted MCP (no local process): for clients that support Streamable HTTP and OAuth.
• Script and CI setup: when automating, it is recommended to wrap the installation in pipefail to avoid a download failure being reported as success;setup --yes reads the key from CONTEXTSTREAM_API_KEY or saved credentials; when the account has multiple workspaces, --workspace-id.
• After installation, follow the guidance to connect the project and supported editors, and restart the editor.
• The official support list provides one-line add commands for Claude Code, Codex, Gemini CLI, Qwen Code, Copilot CLI, and Droid, as well as one-click installation for VS Code.

Applicable Scenarios

Development work that needs to preserve architectural decisions, project intent, and experience across multiple AI coding sessions; teams using multiple MCP clients or models; scenarios where agents are expected to automatically locate relevant code files and cite sources.

Limitations and Notes

• Indexing provides code context, but cannot recover all undocumented decisions.
• Official benchmark numbers are based on specific configurations and do not constitute a latency guarantee.
• Establishing only a hosted connection does not sync the local checkout.
• The repository contains only the MIT-licensed Rust MCP server, client, editor settings, and release tools; the hosted backend is not included in this repository.
• Deep project learning for new accounts is disabled by default;setup --yes preserves the current selection; the agent cannot consent on the user's behalf.
• Before use, review the data processing and control documentation, and be cautious about connecting sensitive projects.
• Specific pricing details, broader platform support scope, and other details are not fully provided in the given materials.