Context Mode Engine: keep raw tool output out of your agent's context window.

One command's output, paid for on every turn.

Claude Code, Cursor, Copilot, Codex and Gemini CLI all work the same way. Every file read, command output and search result stays in the conversation, and the whole conversation goes back to the model on each turn. Run gh issue list once and 59 KB of JSON comes back. Fifty turns later you have paid for it fifty times.

When the window fills, the agent compacts: it summarises the conversation and drops the originals. The files it was editing and the errors it fixed go with them.

Raw data stays in a sandbox. The agent gets the answer.

context-mode is an MCP server that sits between your agent and its tools. When a call would dump large output into the conversation, it runs in a sandboxed subprocess instead, and only what the script prints comes back. The full output goes into a local SQLite FTS5 database with BM25 ranking, and the agent searches it with ctx_search when it needs a detail.

In tools with hook support, the plugin hooks into the session lifecycle:

Tools without hooks use a routing file (AGENTS.md or GEMINI.md) instead.

The agent writes a script. Only its output enters the context.

To count lines in 47 files, an agent would call Read 47 times and keep every file in the conversation. With Think in Code it writes one short script, the sandbox runs it, and only the answer comes back.

// Before: 47 Read() calls, and every file stays in context
// After:  one ctx_execute() call, and only the line counts enter context
ctx_execute("javascript", `
  const files = fs.readdirSync('src').filter(f => f.endsWith('.ts'));
  files.forEach(f => console.log(f + ': ' +
    fs.readFileSync('src/' + f, 'utf8').split('\n').length + ' lines'));
`);

The same idea runs on the gateway too: Thinking in Code →

Your agent does not start from scratch.

The plugin writes each meaningful event to a local database. When the conversation compacts, or you resume with --continue or /resume, it rebuilds your working state and the model continues from your last prompt.

Your code, prompts and sessions stay on your disk.

Thank you for building it with us.

context-mode reached #1 on Hacker News and has 24,230 stars on GitHub (read 2026-09-30). 111 people have contributed code, and many more have filed issues, tested adapters and shared fixes. Every adapter on this page exists because someone asked for it or wrote it.

Context Mode Engine stays free and runs on your machine. It keeps tool output out of one agent's context window, on one machine, in 18 tools. Context Mode Gateway, our main product, is a different layer: it sits on the request path for Claude Code and Codex, lowers the cost of each request, and adds Cage, Memory, Skills, Search and session resume across machines and both clients. For teams, Context Mode Insight reads the plugin's events and shows how the team works with coding agents.