Saving your context window from death by a thousand tool calls

What it does

Context Mode sits between your AI coding agent and its MCP tools, catching raw data before it floods the context window. It replaces direct tool output with sandboxed summaries—turning a 315 KB dump into roughly 5.4 KB—and stores session events in a local SQLite database. When the LLM eventually compacts the conversation to save space, the agent retrieves only relevant prior context via BM25 search instead of re-injecting everything.

The interesting bit

The project enforces a “think in code” paradigm: rather than reading dozens of files into the prompt just to count functions, the agent writes a short script, executes it inside the sandbox, and logs only the result. That shifts computation from the LLM to an actual runtime, which is where the 100× savings come from.

Key highlights

• Claims a 98% reduction in context bloat by intercepting tool output before it reaches the model.
• Session continuity survives conversation compaction via an FTS5 index and BM25 retrieval; previous sessions are wiped immediately if you don’t --continue.
• Ships with 11 MCP tools, including sandbox executors and a meta-tool that opens a local analytics dashboard tracking 90 metrics.
• Enforces routing automatically on hook-capable platforms; others need a one-time routing file copy.

Caveats

• Licensed under ELv2, which carries restrictions on commercial use and redistribution that go beyond typical open-source terms.
• Auto-routing requires platform hooks; agents without hook support need a manual routing file copy to get full enforcement.
• It keeps raw data out of context but explicitly refuses to police the model’s output style, so token waste from verbose agent responses is still your problem.

Verdict

Worth a look if you run long agent sessions and watch your context window evaporate after thirty minutes. If you prefer lightweight, fully open-source tooling or only dabble with AI agents occasionally, the license overhead and platform-specific setup may not pay off.