Context Mode
Context window optimization for AI coding agents. Sandboxes tool output with 98% reduction across 15+ platforms
Overview
Context Mode is an MCP server that optimizes context window usage for AI coding agents by sandboxing tool outputs. Instead of flooding the conversation with large outputs, it stores them in a persistent SQLite database with full-text search (FTS5) and retrieves only relevant information via BM25 search. This allows agents to resume work without losing track of file edits, git operations, tasks, or previous decisions—even across conversation compaction.
The Verdict
Who Should Use Context Mode?
Best For
- Long-running AI coding sessions
- Teams hitting context limits regularly
- Multi-platform development (Claude Code, Cursor, Copilot, etc.)
- Projects requiring session continuity
- Workflows with large tool outputs (logs, search results)
Not Ideal For
- Short, single-query interactions
- Projects where context isn't a bottleneck
- Teams not using supported platforms
- Workflows requiring full output visibility
What's Great
- 98% context reduction on tool outputs (315KB → 5.4KB demonstrated)
- Persistent knowledge base survives conversation compaction
- FTS5 full-text search with BM25 ranking
- Automatic hook integration across 16 platforms
- Code-first analysis: scripts replace multiple tool calls
- Routes agents away from context-flooding operations
- Enterprise adoption: Microsoft, Google, Meta, Amazon, Stripe
Watch Out For
- Adds indirection layer between agent and raw output
- Learning curve for sandbox workflow
- Requires MCP-compatible platform
- May hide output details you want to see
Editor's Note
Context Mode takes a different approach than compression tools: instead of shrinking output, it sandboxes it entirely and uses search to retrieve only what's needed. The 98% figure is real—it's comparing raw output size to the search summary returned. Whether this works for your workflow depends on how often you need the full output versus a smart summary.
Pricing
View all features & details
Core Features
- SQLite persistent storage
- FTS5 full-text search
- BM25 ranking algorithm
- Automatic event indexing
- Session continuity across compaction
Supported Platforms
- Claude Code
- Gemini CLI
- VS Code Copilot
- JetBrains IDEs
- Cursor
- OpenCode, KiloCode
- Zed, Codex CLI
- And more (16 total)
Workflow Changes
- Tool outputs sandboxed, not shown directly
- Agents write analysis scripts instead of reading files
- Search retrieves relevant context on demand
- Hooks route away from context-flooding tools
Technical Stack
- TypeScript implementation
- MCP server architecture
- SQLite with FTS5 extension
- Cross-platform hooks system
How It Compares
| Category | Context Mode | Headroom | RTK |
|---|---|---|---|
| Approach | Sandbox + search | Compression | Compression |
| Context reduction | 98% (sandboxed) | 60-95% | 60-90% |
| Session persistence | SQLite database | In-memory | — |
| Platform support | 16 platforms | Python + proxy | macOS/Linux/WSL |
| Full-text search | FTS5 + BM25 | — | — |
| Survives compaction | Yes | Partial | No |
| Output visibility | Summarized | Compressed | Compressed |
Summary: Context Mode sandboxes outputs entirely rather than compressing them, trading direct visibility for dramatic context savings and persistent session memory.