ContextStream MCP Server

Persistent memory, semantic code search, and team context for Claude Code, Cursor, VS Code Copilot, Windsurf — and every MCP-compatible AI coding assistant.

🏆 90.0% on LongMemEval-S — the full 500-instance suite with the official judge.
Beats supermemory with statistical significance; matches Zep. See the benchmarks →

Documentation • Pricing • FAQ

Try it in 30 seconds

npx --prefer-online -y @contextstream/mcp-server@latest setup

That one command detects your AI editors, writes their MCP configs and rules, installs lifecycle hooks where supported, indexes your project in the background, and verifies everything it just did — then you restart your editor and your AI has memory. Free tier available.

Automating it? Zero-prompt mode takes every default:

npx --prefer-online -y @contextstream/mcp-server@latest setup --yes

Side-by-side comparison: an AI coding assistant with ContextStream memory and semantic search vs. without

We win on memory benchmarks — measured honestly

ContextStream scores 90.0% on the full LongMemEval-S benchmark, the standard test of conversational memory over ~115k-token multi-session histories. That's the complete 500-instance suite with the official GPT-4o judge — 450/500 correct (89.6% single-shot, 90.0% with self-consistency k=3; Wilson 95% CI [87.1%, 92.3%]). Published June 14, 2026.

On multi-session recall — the memory that actually matters for a coding agent working across days of sessions — ContextStream scores 81.2% vs Zep's published 57.9% in the per-family comparison. And on the agentic project-memory benchmark, the same memory raises agent task success from 58% to 96%.

Full methodology, per-family breakdowns, and judge-comparability notes (competitor numbers are cited from each vendor's own publications): contextstream.io/benchmarks

What is ContextStream?

ContextStream is a Model Context Protocol (MCP) server that gives AI coding assistants long-term memory and deep codebase understanding. It indexes your code for semantic search, records your decisions, lessons, and plans across sessions, maps your dependency graph, and pulls in team knowledge from GitHub, Slack, and Notion — then delivers exactly the right slice of all that to your AI on every message.

It works with any MCP client: Claude Code, Cursor, VS Code + GitHub Copilot, Windsurf, Cline, Roo Code, Kilo Code, Codex CLI, OpenCode, Aider, Antigravity, and Claude Desktop.

Why do AI coding assistants forget everything?

Because every conversation starts from zero. Your AI re-reads the same files, re-derives the same architecture, repeats last week's mistake, and loses the thread the moment the context window compacts. ContextStream fixes the whole class of problem:

What your AI can do after setup

🔍 Find code by meaning, not keywords

Ask "where do we handle authentication?" and get ranked, snippet-level answers instantly. Hybrid semantic + keyword search with exact-token fusion, so a symbol lookup like resolveWriteScope lands the definition — not lookalikes. Search works the moment setup finishes: keyword results come back immediately while the semantic index builds in the background.

🧠 Remember everything that matters

Decisions, lessons, preferences, plans, tasks, docs, runbooks — captured during work and surfaced automatically on later turns, in later sessions, even after context compaction. Every prior session's transcript is indexed and queryable: "what did we decide about the id format last week?" just works.

💬 Ask the workspace when stuck

The built-in Agent Q&A tool lets your AI ask your workspace's knowledge base — prior decisions, conventions, runbooks, guardrails — and get a grounded answer with citations for every claim.

🕸️ See the whole graph

"What depends on UserService?" "What breaks if I change this function?" Dependency mapping, impact analysis, circular-dependency and dead-code detection over your whole codebase.

📦 Hand off context between agents

ContextCapsule packages project state into a portable, shareable snapshot — bootstrap a fresh agent, hand off to a teammate, or share a token-gated link with an external agent.

🛡️ Survive long sessions

Token pressure is tracked continuously (with thresholds sized to your model's context window). Before compaction hits, critical state is checkpointed; after it, context restores.

Which AI editors and agents does it support?

Anything that speaks the Model Context Protocol can connect — the table just shows what the setup wizard configures automatically.

The tools your AI gets

36 tools in the default surface, organized as consolidated domains so they cost ~75% fewer tokens than individual registrations:

init / context   → workspace state + the right context on every message
search           → semantic, hybrid, keyword, pattern, exhaustive, refactor modes
memory           → events, decisions, docs, runbooks, tasks, todos, diagrams, transcripts
session          → capture decisions & lessons, recall past sessions, plans, retroactive capture
qa               → grounded Q&A over your workspace knowledge base, with citations
graph            → dependencies, impact analysis, circular deps, unused code
capsule          → portable, shareable context snapshots for agent handoffs
entity           → tickets, incidents, releases, sprints, OKRs, risks
project / workspace / skill / media / vcs / reminder / integration / help

Plus focused write tools (capture_plan, memory_create_doc, …) so agents that display tool names show what they're doing. Your AI uses all of this automatically — you just code.

CLI commands

contextstream-mcp setup            # interactive onboarding wizard
contextstream-mcp setup --yes      # zero-prompt setup with sane defaults (great for CI/dotfiles)
contextstream-mcp doctor           # ✓/✗ diagnostics: auth, scope, index health, rules, hooks
contextstream-mcp index [path]     # index a project folder on demand

setup --editors=claude,cursor limits configuration to specific editors; doctor tells you exactly what's misconfigured and how to fix it — it runs automatically at the end of every setup.

Manual configuration

Skip this if you ran the setup wizard.

claude mcp add contextstream -- npx --prefer-online -y @contextstream/mcp-server@latest
claude mcp update contextstream -e CONTEXTSTREAM_API_URL=https://api.contextstream.io -e CONTEXTSTREAM_API_KEY=your_key

{
  "mcpServers": {
    "contextstream": {
      "command": "npx",
      "args": ["--prefer-online", "-y", "@contextstream/mcp-server@latest"],
      "env": {
        "CONTEXTSTREAM_API_URL": "https://api.contextstream.io",
        "CONTEXTSTREAM_API_KEY": "your_key"
      }
    }
  }
}

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "contextstream": {
      "type": "local",
      "command": ["npx", "-y", "contextstream-mcp"],
      "environment": {
        "CONTEXTSTREAM_API_KEY": "{env:CONTEXTSTREAM_API_KEY}"
      },
      "enabled": true
    }
  }
}

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "contextstream": {
      "type": "remote",
      "url": "https://mcp.contextstream.com",
      "enabled": true
    }
  }
}

{
  "servers": {
    "contextstream": {
      "type": "http",
      "url": "https://mcp.contextstream.io/mcp?default_context_mode=fast"
    }
  }
}

{
  "servers": {
    "contextstream": {
      "type": "stdio",
      "command": "npx",
      "args": ["--prefer-online", "-y", "@contextstream/mcp-server@latest"],
      "env": {
        "CONTEXTSTREAM_API_URL": "https://api.contextstream.io",
        "CONTEXTSTREAM_API_KEY": "your_key",
        "CONTEXTSTREAM_TOOLSET": "complete"
      }
    }
  }
}

{
  "mcpServers": {
    "contextstream": {
      "command": "npx",
      "args": ["--prefer-online", "-y", "@contextstream/mcp-server@latest"],
      "env": {
        "CONTEXTSTREAM_API_URL": "https://api.contextstream.io",
        "CONTEXTSTREAM_API_KEY": "your_key",
        "CONTEXTSTREAM_TOOLSET": "complete"
      }
    }
  }
}

FAQ

What is an MCP server?

An MCP (Model Context Protocol) server exposes tools and context to AI assistants over a standard protocol. Claude Code, Cursor, VS Code Copilot, Windsurf, and most modern AI coding tools are MCP clients — install one server, and every client you use gets the same capabilities.

Does ContextStream work with Claude Code / Cursor / Copilot / Windsurf?

Yes — all of them, plus Cline, Roo Code, Kilo Code, Codex CLI, OpenCode, Aider, Antigravity, and Claude Desktop. The setup wizard configures each editor's MCP config, managed rules, and (where the editor supports them) lifecycle hooks automatically.

Is my code private?

Your code is private and securely stored, isolated per workspace with no cross-tenant access. You control exclusions with .contextstream/ignore (gitignore syntax), and project(action="purge") completely de-indexes a project on demand — without touching your captured memory.

How much does it cost?

There's a free tier to start with. Larger indexes, the full code graph, and team features are on paid plans: see pricing.

How does ContextStream score on memory benchmarks?

90.0% on the full 500-instance LongMemEval-S suite with the official GPT-4o judge (89.6% single-shot) — beating supermemory's published 85.4% with statistical significance and matching Zep's published 90.2% within the confidence interval. On multi-session recall specifically, ContextStream scores 81.2% vs Zep's published 57.9%. Methodology and per-family breakdowns: contextstream.io/benchmarks.

How is this different from other AI memory tools?

Most tools store notes. ContextStream combines memory (decisions, lessons, plans, session transcripts), semantic code search over your indexed codebase, a dependency/knowledge graph, grounded Q&A with citations, and team integrations (GitHub, Slack, Notion) behind one MCP server — and proactively delivers the relevant slice on every message instead of waiting to be asked.

Can I use it in CI or scripted environments?

Yes: setup --yes runs the entire wizard with zero prompts (set CONTEXTSTREAM_API_KEY in the environment), --editors=<list> scopes it, and contextstream-mcp doctor gives scriptable ✓/✗ diagnostics.

Do I have to wait for indexing?

No. Project indexing runs in the background — keyword search works immediately, and semantic results fill in as the index builds. Check progress anytime with contextstream-mcp doctor.

How do I uninstall or disable it?

Remove the contextstream entry from your editor's MCP config, and set CONTEXTSTREAM_HOOK_ENABLED=false (or re-run setup) to disable hooks. project(action="forget_local") unbinds a folder locally without touching server-side data.

Troubleshooting

• Start with contextstream-mcp doctor — it checks auth, API reachability, folder scope, index health, rule files, and hooks, with a fix hint per failure.
• Remove duplicate ContextStream entries across Workspace/User config scopes.
• Check CONTEXTSTREAM_API_URL and CONTEXTSTREAM_API_KEY are set; remove stale version pins like @contextstream/mcp-server@0.3.xx.
• Restart your editor after config changes.
• Hosted HTTP OAuth: the remote transport's OAuth flow routes through vscode.dev; where that's blocked (corporate networks), use stdio with an API key instead.
• SDK compatibility: @modelcontextprotocol/sdk 1.28.0+ introduces breaking changes; this package pins >=1.25.1 <1.28.0. If you see Zod schema errors on startup, check your SDK resolution.

Links

Website: https://contextstream.io • Docs: https://contextstream.io/docs • Changelog: CHANGELOG.md

Stop teaching your AI the same things over and over.
_{ContextStream makes it brilliant from the first message.}