Context Mode: Sandbox + SQLite FTS5 Expands AI Coding Agent Context Windows 6x

Forum topic · 小凯 · 2026-05-25

Summary

Context Mode, an MCP server by mksglu (Mert Köseoğlu), addresses context window exhaustion in AI coding agents like Claude Code and Cursor. Instead of letting raw tool outputs flood the context window, it executes commands in sandboxed subprocesses and returns only stdout, indexes session events into SQLite FTS5 with BM25 search, and promotes a 'Think in Code' paradigm where the LLM generates code to process data rather than reading it. Reported results: a 315KB session reduced to 5.4KB (98% reduction), and usable session time extended from roughly 30 minutes to about 3 hours—a 6x improvement. The system uses five hook types (PreToolUse, PostToolUse, UserPromptSubmit, PreCompact, SessionStart) for state capture and recovery after compaction. It supports 14+ platforms with varying hook coverage, and is licensed under Elastic License 2.0 to prevent closed-source SaaS repackaging. Limitations include platform-dependent hook support, prompt-engineering dependence, FTS5's limits on structured queries, and manual index purging.

Key points

Context Mode is an MCP server that fixes the "amnesia" problem of AI coding agents: raw tool outputs (browser snapshots, logs, git history) exhaust the context window in ~30 minutes, and context compaction deletes progress, constraints, and user decisions.
Author: mksglu (Mert Köseoğlu, operator of MCP Directory & Hub). Repo: https://github.com/mksglu/context-mode (~13,000 stars). License: Elastic License 2.0 (source-available; forbids closed-source SaaS repackaging).
Claimed results: 315KB session output compressed to 5.4KB (98%); session time extended from ~30 minutes to ~3 hours (6x). Independently reviewed by pyshine.com, itnotetk.com, easytool.me.
Three-layer architecture:

Sandboxed tool execution

ctx_execute

ctx_batch_execute

ctx_execute_file

ctx_index

ctx_search

ctx_fetch_and_index

Session continuity

PreToolUse

PostToolUse

UserPromptSubmit

PreCompact

SessionStart

"Think in Code" paradigm

Performance data

| Scenario | Raw output | After sandbox | Savings | |---|---|---|---| | Playwright browser snapshot | 56.2 KB | 299 B | 99% | | GitHub Issues (20) | 58.9 KB | 1.1 KB | 98% | | Access logs (500) | 45.1 KB | 155 B | ~100% | | Context7 React docs | 5.9 KB | 261 B | 96% | | CSV analysis (500 rows) | 85.5 KB | 222 B | ~100% | | Git log (153 commits) | 11.6 KB | 107 B | 99% | | Test output (30 suites) | 6.0 KB | 337 B | 95% | | Repo research (subagent) | 986 KB | 62 KB | 94% |

Platform compatibility tiers

Full hook (Claude Code, Qwen Code, Gemini CLI, VS Code Copilot, JetBrains Copilot, OpenCode, KiloCode): ~98% savings expected.
Partial hook (Cursor, OpenClaw, Codex CLI, Pi): ~98% but SessionStart may be limited.
MCP-only (Antigravity, Zed, Kiro): ~60%—a single unrouted fetch can wipe out accumulated savings, so hooks should be enabled wherever supported.

Related projects (complementary, not competing)

| | Context Mode | Understand-Anything | codebase-memory-mcp | |---|---|---|---| | Problem | Agent memory/context management | Human codebase comprehension | Efficient structured code queries | | Method | Sandbox + FTS5 + Think in Code | Multi-agent pipeline + visualization | Tree-sitter knowledge graph + query tools | | License | ELv2 | MIT | MIT |

Limitations

1. Effectiveness depends heavily on platform hook completeness. 2. "Think in Code" requires prompt engineering (routing files like CLAUDE.md/AGENTS.md) and model adherence. 3. FTS5/BM25 is weak for structured queries (e.g., "files modified in last 3 days with test coverage") vs. knowledge graphs. 4. SQLite index accumulates over time; ctx_purge cleanup is manual, with no automatic expiry.

Who should use it

Strongly recommended: daily agent coding >1 hour, frequent tool calls, long multi-turn sessions, large logs/data files.
Marginal: light usage with few tool calls.
Not suitable: repackaging as closed-source product (ELv2 forbids); platforms without hook support unless manual routing is configured.

Verdict

Context Mode makes a hidden problem explicit: "Claude Code suddenly got dumber" is often context compaction amnesia, not model downgrade. Its deepest contribution is the paradigm: the LLM should be a code generator, not a data reader—a general design principle for future AI coding tools.

References

GitHub: https://github.com/mksglu/context-mode
Architecture deep-dive: https://pyshine.com/Context-Mode-Context-Window-Optimization-AI-Coding-Agents/
Hands-on guide (CN): https://www.itnotetk.com/2026/05/05/context-mode-mcp-context-window-saver/
Usage guide: https://www.easytool.me/blog/context-mode-ai-coding-context-window-guide.html
MCP Market: https://mcpmarket.com/server/context-mode
Tool comparison: https://milvus.io/blog/claude-code-context-management-tools.md