Nanosaur2-670M
well9472 · Kit
Makes illustrations from text prompts with a small 670M-parameter diffusion model, run through ComfyUI. Built as a low-budget research model for people who wan…

mksglu’s free, local plugin that keeps tool output out of your coding agent’s context window. Raw output stays in a sandbox and a SQLite index, and only the answer comes back. The setup notes cover Claude Code, Codex, Gemini CLI, Cursor, Copilot and more.
Context Mode is an open-source project on GitHub in the Developer category, listed by Mindket. It is free. This is version 2, last updated on . Install it from GitHub as its README says. Get Mindket’s setup notes on your Mindket account, as a zip.
/plugin marketplace add mksglu/context-mode /plugin install context-mode@context-mode
And 4 more in Where to use it
Context Mode is a free, local plugin that keeps tool output out of your coding agent’s context window. The README says every MCP tool call dumps raw data into that window, and compaction then makes the agent forget which files it was editing. Context Mode is an MCP server whose tools run code and read files in a sandbox on your machine. Only what the code prints reaches the model. A SQLite index with search lets the agent look up what it saved. Hooks track your edits, tasks and decisions, so a session can pick up after the conversation compacts.
Developers who work in long coding-agent sessions and watch logs, test output, web pages and API responses fill the context window. Skip it if your sessions are short. It adds hooks and a local database to your setup, and on some agents a routing file you copy, and that is a lot for a quick job.
| Industry | Who uses it | How they use it | Example |
|---|---|---|---|
| Software teams | Developers on large codebases | Let the agent run a script over many files and read only the result | A developer asks for a count of functions per file, and the agent prints one line per file instead of reading each file in. |
| Operations and support | Engineers who read logs and exports | Analyze a long log or a CSV in the sandbox and search it | An engineer points the agent at an access log and asks which paths fail, and only the matching lines come back. |
| Web and QA | Developers who use browser tools | Index a fetched page or a Playwright snapshot and search it | A tester has the agent fetch a docs page, index it and search for one pattern, so the page never enters the chat. |
| Long projects | Anyone who hits compaction | Resume with files, tasks and decisions rebuilt from the local database | After the conversation compacts, the agent still knows which files it was editing and what you last asked. |
The README documents these agents, each with its own install route. Routing is enforced only where an agent has hooks.
Get it saves a zip of Mindket’s setup notes in your downloads folder, with a page for each agent from the maker’s README and docs. Claude Code installs a plugin. Most other agents take a global package and a config entry.
| Need | Detail |
|---|---|
| Node.js | 22.5 or newer, or Bun. The README says Linux with Node below 22.5 is unsupported |
| An agent | One from the list above, already installed. Claude Code needs v1.0.33 or newer. Antigravity CLI needs 1.0.7, Copilot Chat 0.32, the JetBrains Copilot plugin 1.5.57, and OpenClaw newer than 2026.1.29 |
| Build tools | Only on old Linux such as CentOS 7 and 8 and RHEL 8, where the SQLite addon builds from source |
| An account or key | None. The README says no account is required |
/plugin marketplace add mksglu/context-mode /plugin install context-mode@context-mode
Restart Claude Code, then run /context-mode:ctx-doctor. The README says every check should show [x]. For most other agents the first step is this line, and the rest is a config entry and a hooks file.
npm install -g context-mode
The MCP entry has the same shape in Cursor, Kiro, Antigravity, Pi and OMP, each in its own file. VS Code Copilot uses a servers key and Zed uses context_servers.
{
"mcpServers": {
"context-mode": {
"command": "context-mode"
}
}
}Work as before. The hooks steer the agent toward the sandbox tools when a command or a file would flood the context, and you can ask for them by name. Two commands show what happened.
ctx stats ctx doctor
Typing ctx stats shows context savings and call counts. ctx doctor checks runtimes, hooks, FTS5 and versions. In Claude Code the same commands are slash commands, such as /context-mode:ctx-stats. ctx index and ctx search add your own files to the index and query them.
You want the Hacker News front page grouped by domain. The README says that without hooks one unrouted curl or Playwright snapshot can dump 56 KB into context. With the plugin, you send this prompt, which comes from the README’s Try It list:
Fetch the Hacker News front page, extract all posts with titles, scores, and domains. Group by domain. Then run /context-mode:ctx-stats.
The README lists this prompt as one tool call, and says web pages are processed in the sandbox and never enter the context window. You get the posts grouped by domain, and ctx-stats then reports the savings.
If you are new. A coding agent reads whatever its tools return, and all of it takes room in a window with a limit. When the window fills, the agent compacts the conversation and can lose the thread. Context Mode keeps the bulky part, such as a log, a web page or a test run, on your machine and hands the agent a short answer. It also keeps a local record of your edits and tasks. Start with the Claude Code plugin and run ctx doctor.
For experts. Search runs two strategies on SQLite FTS5, a Porter-stemmed match and a trigram substring match, and merges them with Reciprocal Rank Fusion. Results are reranked by how close the query terms sit, and a typo is corrected by Levenshtein distance before a second search. Fetched pages are cached for 24 hours by default, and a call can override that with ttl or force.
The SQLite backend is picked at runtime: bun:sqlite on Bun, node:sqlite on Linux with Node 22.5 or newer, and better-sqlite3 elsewhere. Hooks feed the session database. Before compaction a snapshot is built in priority tiers, and SessionStart restores it. Searches throttle: from the fourth call results shrink, and from the ninth they are blocked and sent to ctx_batch_execute.
Most of it runs on your machine, and a few paths reach out.
This kit holds setup notes only. Mindket wrote them. Context Mode itself, its plugin files and its npm package come from its maker's repository: https://github.com/mksglu/context-mode
These notes are not made or endorsed by its maker. The GitHub description reads "Free, local plugin that keeps tool output out of your coding agent's context window." The README calls Context Mode "an MCP server". The maker's site is https://contextmode.ai/
License: Elastic License 2.0. The LICENSE file (https://github.com/mksglu/context-mode/blob/main/LICENSE) names "Copyright 2026 Mert Koseoglu" and the npm entry lists "Elastic-2.0". The README says this is source-available, not MIT: "You can use it, fork it, modify it, and distribute it. Two things you can't do: offer it as a hosted/managed service, or remove the licensing notices." Read the LICENSE for the full terms.
context-mode npm package, an MCP server entry in the agent's config file, and, where the agent has hooks, a hooks file. Source: README.md (Install) and https://github.com/mksglu/context-mode/blob/main/docs/platform-support.mdAGENTS.md, GEMINI.md, KIRO.md, QWEN.md, a Cursor rules file, copilot-instructions.md or SYSTEM.md) that you copy by hand where the README says to. It tells the model where data should go.~/.context-mode/content/ and that session data lives in your home directory. Its storage table lists adapter defaults such as ~/.claude/context-mode and ~/.codex/context-mode, and platform-support.md lists a context-mode/sessions/ folder for each agent. CONTEXT_MODE_DIR moves the root: see setup/codex.md.>=22.5.0. Linux with Node below 22.5 is unsupported (README, Build Prerequisites).npm install -g context-mode. platform-support.md (Prerequisites) says every platform except the Claude Code plugin needs it, but the README's steps for the Codex plugin, OpenCode, KiloCode, the OMP plugin and OpenClaw do not include it, and platform-support.md says the Cursor plugin's hooks run through npx -y with no global install. Follow the page for your agent.The README's install steps ask for no account or key, and its Privacy section says none is required.
agy command-line toolctx commands and the README's try-it promptsThe README groups platforms by how much of the setup is automatic: where the agent has hooks, routing is enforced for you, and where it has none, you copy one routing file once. It says without hooks "~60% saved" against "~98% saved" with them, so turn hooks on wherever the agent allows it.
The README's utility commands include an upgrade. Inside an agent session, type this and the model calls the tool:
ctx upgrade
From a terminal, the README gives:
context-mode upgrade
The Copilot CLI note in the README gives npm install -g context-mode@latest for the global package. Claude Code also has /context-mode:ctx-upgrade. Source: README.md (Utility Commands, Install)
ctx purge permanently deletes all indexed content from the knowledge base. Source: README.md (Tools)
The README gives no general uninstall steps. For the Antigravity CLI it gives agy plugin uninstall context-mode. For the Copilot CLI MCP entry, platform-support.md names copilot mcp remove. The maker's site, https://contextmode.ai/, lists a remove command for each agent on its own install route (npm tag @next), such as context-mode setup --remove --agents claude-code; the README does not document it. The notes below are from Mindket, not the maker: to remove Context Mode by hand, undo what the steps for your agent wrote: the global package, the MCP entry, the hooks file, a copied routing file and the data folders above.
The README says: "Nothing leaves your machine. No telemetry, no cloud sync, no usage tracking, no account required." Source: README.md, Privacy & Architecture. The GitHub description says "Zero telemetry".
One command goes to a hosted page: ctx insight (/context-mode:ctx-insight in Claude Code) opens the hosted Insight dashboard at https://context-mode.com/insight in your browser, which the README describes as "org analytics for AI-assisted engineering teams". The README says nothing more about what that page reads or stores. The maker's site describes Context Mode Insight as "How a team works with coding agents, from Engine events", where Engine is its name for this free plugin, and does not say how those events reach it.
ctx_fetch_and_index fetches the web addresses you or the agent give it. ctx_upgrade pulls the latest version from GitHub. See setup/permissions-and-safety.md.
README.md 7 KB setup/antigravity.md 3 KB setup/build-prerequisites.md 3 KB setup/claude-code.md 3 KB setup/codex.md 5 KB setup/commands-and-tools.md 4 KB setup/cursor.md 3 KB setup/gemini-cli.md 2 KB setup/github-copilot.md 6 KB setup/kimi-code.md 3 KB setup/kiro.md 2 KB setup/openclaw.md 3 KB setup/opencode-and-kilocode.md 3 KB setup/permissions-and-safety.md 5 KB setup/pi-and-omp.md 3 KB setup/qwen-code.md 2 KB setup/zed.md 2 KB
Updated by Mindket to the current version of this kit.
Listed GitHub projects in the same category that share the most topics with this one.
| Product | Stars | Activity | Language | License | Last change |
|---|---|---|---|---|---|
| Context Mode This page | 33,650 | — | TypeScript | — | — |
| UI UX Pro MaxShares kiro · copilot · antigravity | 146,944 | Active | Python | MIT | |
| Cheap AgentShares claude-code-hooks · claude-code-skill · claude | 310 | Active | TypeScript | MIT | |
| Agentic Resource RegistryShares kiro · antigravity · codex | 4,320 | Active | Python | Apache-2.0 | |
| GraphifyShares antigravity · codex · claude-code | 144,528 | Active | Python | Apache-2.0 | |
| Agent SkillsShares antigravity · codex · claude-code | 116,705 | Active | JavaScript | MIT |
Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search and on Trendshift. Mindket doesn’t write or check them. Open one to read it where it was posted.
well9472 · Kit
Makes illustrations from text prompts with a small 670M-parameter diffusion model, run through ComfyUI. Built as a low-budget research model for people who wan…
Waqf-AI · Kit
Reads handwritten Arabic text from images and returns it as text. A fine-tuned PaddleOCR-VL model for people digitizing Arabic handwriting.
Wan-AI · Kit
Makes video clips from a text prompt or an image plus text, at up to 720P and 24 frames per second, in the Diffusers format for a single consumer GPU.
nextlevelbuilder · Skill
Generates tailored design systems, including color palettes, typography pairings, and layout patterns, based on specific project requirements and industry stan…
AgenticResourceRegistry · Kit
Keeps a team's coding agent skills, hooks, prompts and agent files in one self-hosted registry, with review of new entries and usage insight. Suits engineering…
muaaz9329 · Kit
Moves routine coding in Claude Code or Codex to a low-cost model that must prove its work with a test, so your main model's usage lasts longer.