Context Mode
Keep raw tool output out of your coding agent's context window
About Context Mode
Context Mode is an MCP plugin that sits between a coding agent and its tools and keeps large tool output out of the model's context window. Every grep, file read or gh issue list normally lands in the conversation and is re-sent on every subsequent turn; fifty turns in, one command can have cost hundreds of thousands of input tokens. Context Mode intercepts that output before it reaches the context, stores it in a local SQLite full-text index, and lets the agent search it when it actually needs a piece. Their homepage claims up to 98 percent of context saved per session across 17 agent adapters, including Claude Code, Cursor, Copilot, Codex, Gemini CLI, JetBrains Copilot, Antigravity CLI and Kiro.
The plugin is free, installs with npm i -g context-mode, runs entirely on your machine, and needs no account and sends no telemetry. The licence is the Elastic License 2.0, which is source-available rather than open source in the OSI sense: you can read, modify and run it, but you cannot offer it as a hosted service. That is why this entry is filed as freemium. The paid layer is Context Mode Platform at $20 per seat per month, an organisation dashboard fed by the same plugin through an opt-in switch; the page states it forwards only structural events - tool names, file paths, error counts - and never source, prompts or file contents. The traction is verified rather than claimed: 22,134 GitHub stars, 110 contributors, 71,902 npm downloads in the thirty days before this was written, and a Hacker News front-page post at 570 points in February 2026. The "553,300+ developers" figure on the homepage is their own cumulative install count and could not be checked independently.
Install the plugin globally and register it with your agent as an MCP server. From then on, when a tool call returns something large, the plugin writes the raw result into a local FTS5 store and returns a compact reference to the agent instead. The agent has search tools it can call against that store, so it pulls the three matching lines from a 4,000-line log rather than reading the whole log into the conversation. Session memory persists across runs, so the store survives a restart. Nothing leaves the machine on the free plugin. If your organisation buys the Platform, you flip one opt-in switch and the same plugin forwards structural events over an authenticated channel to a private workspace, where the first Solution, Insight, reports productive session rate per engineer, retry waste per team and named patterns, with views narrowed for CTO, EM, IC, CISO, FinOps and DevOps roles. Memory, Audit and Cost Solutions are listed as coming in 2026.
- •Tool Output Sandboxing - Large results go to a local index and the agent searches them instead of carrying them in context
- •17 Agent Adapters - Claude Code, Cursor, Copilot, Codex, Gemini CLI, JetBrains Copilot, Antigravity CLI, Kiro and more
- •Local and Account-Free - The plugin runs on your machine with no cloud, no telemetry and no sign-up
- •Persistent Session Memory - The store survives restarts, so prior context is searchable in later sessions
- •Platform Insight ($20/seat) - Opt-in org dashboard built from structural events only, per their privacy statement
- •Elastic License 2.0 - Source-available; free to run and modify, not free to resell as a service
Anyone running a coding agent on a large repository who watches the context fill with tool output and the bill follow it. That is most readers of this directory, and the plugin is free, so the cost of trying it is an npm install. The Platform is for engineering leadership who want a view of how AI-assisted work is going across a team; it is also, whatever the privacy framing, a dashboard that shows an engineering manager who is stuck, and some teams will not want that. It is not for anyone who needs an OSI-approved licence, and not for anyone building a hosted product on top of it. Note that three of the four Platform Solutions are roadmap, so today the $20 buys Insight and a promise.
