MaxiMinion for VS Code

MaxiMinion.AI is a local-first context optimizer for AI-assisted coding.

Before you paste code into Copilot Chat, ChatGPT, Claude, or any LLM, MaxiMinion:

• Redacts likely secrets: API keys, cloud credentials, tokens, private keys, and connection strings become [REDACTED_…].
• Shows token counts: a status bar counter and before/after estimates warn you when you go over your budget.
• Copies clean, compact context: selections, files, or Explorer multi-selections become labelled Markdown blocks with wasted whitespace removed.
• Warns you while you edit: likely secrets are flagged in the Problems panel with a one-click redact fix.

Everything runs on your machine. Your source files are only changed if you apply a quick fix. Optional summarization uses your own local Ollama model; your code is never sent to MaxiMinion servers, and Ollama cloud models are refused unless you allow them.

Coming in 0.5.0: dependency-graph ranking that picks the most important files in your workspace to fit your token budget.

New here? Run MaxiMinion: Get Started & Help from the Command Palette for a short walkthrough.

Where to find MaxiMinion

Everything MaxiMinion adds is labelled with its name:

Which command should I use?

Copy for AI

Copies the current selection, the active file, or the files selected in the Explorer as labelled Markdown code blocks, after redacting secrets and compacting whitespace (indentation is kept). A notification reports how many values were redacted and the estimated tokens before and after, and warns when the copy exceeds your token budget.

The token counter in the status bar shows the estimate for the active file or selection, turns amber when over budget, and runs Copy for AI when clicked.

Copy for AI and Copy Sanitized File ask before copying:

• Files that usually hold credentials, such as .env, private keys, .npmrc, or Terraform state. If you copy them anyway, likely secrets are still redacted. Templates such as .env.example are not flagged.
• Hidden Unicode: bidirectional controls, invisible tag characters, and zero-width characters can hide text or instructions from you that an AI model still reads ("Trojan Source" and prompt injection). Choose Show to jump to the first one. MaxiMinion never removes them silently.

Activity log

MaxiMinion: Show Activity Log (also in the sidebar title bar) opens the MaxiMinion Activity channel. Each copy to the clipboard, request to your local Ollama model, local preview, and declined copy is listed with file paths, the number of values redacted, whether redaction was on, and estimated tokens. File content is never logged, and nothing leaves your machine.

Security status

The Security section of the sidebar lists each protection with its live state: on, off (turned off in your settings), or needs attention. Hover a row for details. MaxiMinion: Show Security Status opens the full report, including which part of the code enforces each control. Every value is read from that code, not written by hand, so the report shows what this version actually does.

See the Security Model for how MaxiMinion protects your code and its known limits.

Inline secret warnings

Likely secrets in open files are underlined and listed in the Problems panel with the source MaxiMinion and a code such as API_KEY. Keys and credentials are warnings; emails, SSNs, and card-like numbers are information.

• Redact in place: press Ctrl+. on the value and choose MaxiMinion: Redact with [REDACTED_…]. Only the value is replaced, so apiKey = "…" stays valid code.
• Keep the file as is: use Copy for AI, which redacts only the copy.
• Too noisy? Raise maximinion.secretDiagnostics.minConfidence, or turn warnings off with maximinion.secretDiagnostics.enabled. Environment variable references such as process.env.API_KEY are never flagged.

The Refiner

The Refiner is the analysis pipeline behind Run Refiner on Active File (Signal Report). Use it on large files to decide what is worth sending, or to shrink them.

1. Secret scrubbing: likely secrets are replaced with [REDACTED_…].
2. Entropy scoring: Shannon entropy measures information density. HIGH means dense, varied content (usually real logic). LOW means repetitive content such as imports, generated code, or long comment blocks, which is a good candidate to trim.
3. Semantic folding (optional, off by default): a local Ollama model summarizes the code. The summary is lossy, so review it. To enable it, start Ollama (ollama serve), install a model (ollama pull phi3, or use one from ollama list), turn on maximinion.semanticFolding.enabled, and set maximinion.semanticFolding.model if you're not using phi3. Run MaxiMinion: Test Semantic Folding Setup (Ollama) (also in the sidebar and linked from the setting) to check that Ollama is reachable and the model is installed, and to time a sample fold. Turning folding on, or changing its model, runs the same check; if it fails, folding is turned off again and you're told how to fix it. The sidebar shows whether folding is ready. Folding shows a progress notification that you can cancel. Ollama cloud models (for example gpt-oss:120b-cloud) run on another host, so Ollama would send your code there; MaxiMinion refuses them unless you turn on maximinion.semanticFolding.allowRemoteModels in your user settings.

The refined text opens in a preview tab, and your file is not changed. A report with a plain-English explanation of each step is written to View → Output → MaxiMinion.

Commands

All commands are in the Command Palette under MaxiMinion:.

• Copy for AI (Sanitized): see Copy for AI.
• Preview Sanitized File: opens a redacted copy of the active file in a preview tab.
• Copy Sanitized File (Plain Text): copies the redacted text with no labels or fences.
• Run Refiner on Active File (Signal Report): see The Refiner.
• Open Settings: opens the MaxiMinion settings.
• Test Semantic Folding Setup (Ollama): checks that Ollama is reachable and the folding model is installed, then times a sample fold.
• Show Activity Log: see Activity log.
• Show Security Status: see Security status.
• Get Started & Help: opens the walkthrough.

Settings

Notes

• Secret detection is pattern-based and cannot guarantee that every sensitive value is found. Review the output before sharing it.
• Inline warnings skip environment variable references (such as process.env.API_KEY) and lower-confidence patterns by default; lower maximinion.secretDiagnostics.minConfidence to see more.
• Token counts are estimates (~4 characters per token). Actual counts vary by model.
• MaxiMinion works in Restricted Mode (untrusted folders). Semantic folding can only be enabled from user settings until the workspace is trusted.
• Secret protection settings (secretScrubbing.enabled, secretDiagnostics.enabled, secretDiagnostics.minConfidence) can only be changed in your user settings. Values in a workspace's .vscode/settings.json are ignored, so a repository cannot turn redaction off.
• The Refiner preview contains only the refined text (no report wrapper), so it can be copied straight into an LLM prompt with minimal token overhead. Full run details are in the MaxiMinion output channel.

Install

Install MaxiMinion from the Extensions view (search "MaxiMinion"), or run:

code --install-extension maximinion.maximinion

Build from source

From the root of the repository:

npm ci
npm run build
npm run package --workspace packages/vscode-extension
code --install-extension packages/vscode-extension/maximinion-<version>.vsix

Replace <version> with the version in package.json.

Licensing & Compliance

"MaxiMinion.AI Community Edition is an independent, zero-cost tool provided strictly under the MIT License. The provision of this free tier does not constitute an operation in trade or commerce under the Australian Consumer Law (ACL), and the standard statutory consumer guarantees do not apply to this zero-cost release."

This Community Edition is provided completely free of charge for evaluation and individual use. The software is provided 'AS IS' without warranties of any kind. See LICENSE.