Context Mode MCP — Use Cases, Install & Live Demo

Why use it

Key features

98% context reduction via sandboxed tool execution
Per-session SQLite FTS5 index survives compaction
Automatic routing hooks (PreToolUse, PostToolUse, PreCompact, SessionStart)
12 platforms supported (Claude Code, Cursor, Codex, Gemini CLI, Copilot…)
'Think in code' paradigm — have the LLM write scripts, not process data itself

Live Demo

What it looks like in practice

context-mode-mcp.replay ▶ ready

0/0

Install

Pick your client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "context-mode-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "context-mode"
      ]
    }
  }
}

Open Claude Desktop → Settings → Developer → Edit Config. Restart after saving.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "context-mode-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "context-mode"
      ]
    }
  }
}

Cursor uses the same mcpServers schema as Claude Desktop. Project config wins over global.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "context-mode-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "context-mode"
      ]
    }
  }
}

Click the MCP Servers icon in the Cline sidebar, then "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "context-mode-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "context-mode"
      ]
    }
  }
}

Same shape as Claude Desktop. Restart Windsurf to pick up changes.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "context-mode-mcp",
      "command": "npx",
      "args": [
        "-y",
        "context-mode"
      ]
    }
  ]
}

Continue uses an array of server objects rather than a map.

~/.config/zed/settings.json

{
  "context_servers": {
    "context-mode-mcp": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "context-mode"
        ]
      }
    }
  }
}

Add to context_servers. Zed hot-reloads on save.

claude mcp add context-mode-mcp -- npx -y context-mode

One-liner. Verify with claude mcp list. Remove with claude mcp remove.

Use Cases

Real-world ways to use Context Mode

Stop wasting context on raw tool output

👤 Heavy Claude Code users who hit compaction mid-session ⏱ ~15 min intermediate

When to use: You feel your agent "forget" after 30 minutes and burn $$ on re-reads.

Prerequisites

Claude Code v1.0.33+ — brew upgrade claude-code or npm update -g @anthropic-ai/claude-code

Flow

Install

/plugin marketplace add mksglu/context-mode
/plugin install context-mode@context-mode✓ Copied

→ Plugin registers hooks + 6 sandbox tools
Verify

/context-mode:ctx-doctor✓ Copied

→ All checks [x]
Use normally

Now do a big task that would normally blow context — e.g. analyze all .log files under /var/log✓ Copied

→ ctx_execute used instead of raw Read; output indexed not dumped
Check savings

/context-mode:ctx-stats✓ Copied

→ Per-tool reduction ratio reported

Outcome: Same task, 10x longer before compaction, and the agent remembers.

Pitfalls

Other MCPs bypass the routing — Pass their tool calls through ctx_batch_execute for consistent accounting

Combine with: serena

Resume a long task seamlessly after the model compacts

👤 Devs on multi-hour agent sessions ⏱ ~5 min beginner

When to use: Your task is too big to fit in one context window.

Flow

Start with --continue

[CLI flag] claude --continue✓ Copied

→ Prior session's SQLite loaded
Observe

What did I last ask you about? What tasks are in progress?✓ Copied

→ Accurate recap from FTS5 index, not hallucinated

Outcome: A genuinely resumable workflow.

Pitfalls

Forgetting --continue starts fresh — Persistence requires --continue; absence means deliberate fresh start

Combinations

Pair with other MCPs for X10 leverage

context-mode-mcp + serena

Serena for semantic code nav, context-mode for keeping its output out of context

Use serena for the lookup, ctx_execute to post-process the results.✓ Copied

context-mode-mcp + playwright

Page snapshot output stays sandboxed

Snapshot the page via playwright, let context-mode hold it — search it instead of dumping.✓ Copied

Tools

What this MCP exposes

Tool	Inputs	When to call	Cost
ctx_execute	code: str (js/python)	Run a script to process data; output only what's needed	free
ctx_batch_execute	ops: [{tool, args}]	Wrap multiple tool calls through the sandbox	free
ctx_index	content: str, tags?: []	Stash data for later retrieval	free
ctx_search	query: str	Retrieve from session index	free
ctx_fetch_and_index	url: str	Pull a URL, keep it out of context but searchable	free
ctx_stats	none	See how much context was saved	free

Cost & Limits

What this costs to run

API quota: Local
Tokens per call: Dramatic reduction — quoted 98% on real sessions
Monetary: Free (ELv2 license)
Tip: The bigger the session, the bigger the win — use on long tasks

Security

Permissions, secrets, blast radius

Credential storage: Session SQLite in ~/.context-mode/

Data egress: None

SQLite persists across sessions only if you use --continue — otherwise it's wiped

Troubleshooting

Common errors and fixes

ctx-doctor shows missing hooks

Restart Claude Code fully after install; hooks register at startup

Verify: Re-run /context-mode:ctx-doctor

Sandbox won't execute code

Node/Python runtime missing. Doctor tells you which. Install with brew/asdf.

Search returns nothing

Verify FTS5 built: sqlite3 ~/.context-mode/session.db "SELECT name FROM sqlite_master". Reinstall if broken.

Alternatives

Context Mode vs others

Alternative	When to use it instead	Tradeoff
Raw Claude Code (no plugin)	Short sessions that never hit compaction	No context saving
claude-mem	You want the memory layer without the sandbox-tool paradigm	Different tradeoff — memory only, no tool output sandbox

More

Resources

📖 Read the official README on GitHub

🐙 Browse open issues

🔍 Browse all 400+ MCP servers and Skills