A
licenseB
qualityD
maintenanceAn MCP server that lets AI assistants read and visually analyze local documents — PDFs, Excel spreadsheets, CSV files, Word documents, PowerPoint presentations, and images.
4
21 npm
49 PyPI
MIT
Malyn
Officialby synops-one
Run bounded local AI jobs through Malyn using Markdown recipes, then check or manage their results.
Discover and read available recipes in Malyn’s Kitchen.
Import reviewed local Markdown recipes and export shareable recipe files; importing never runs a recipe.
Submit jobs with a recipe, root path, files, images, and task instructions.
Check job status and get report links, or request a summary return mode when needed.
Cancel jobs and inspect available models or components.
Configure local model settings such as model key, context length, output tokens, and cache quantization.
Use built-in defaults for image description and Windows system log checks.
Access legacy loop actions and loop metadata where still supported.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Malyncheck my system logs."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Meet Malyn, your local AI helper. She helps you make sense of files, images and logs using the GPU in your own PC.
Your AI agent can hand her the time-consuming reading while it plans the next step. Ask her to investigate recurring errors, pull useful details from a file or explain a screenshot. You can also ask her for help directly from Windows.
Malyn gives you a way to put your own hardware to work and keep bulk file processing on your computer, without pasting everything into a chat.
Download Malyn for Windows · Watch the 67-second demo
What can she help with?
Try asking | What she can do |
“Malyn, check my system logs.” | Look through a snapshot of recent Windows events and highlight problems worth investigating. |
“Help me understand this file.” | Summarise a text or code file, pull out key facts and suggest next steps. |
“What's in this image?” | Describe a photo or explain what is visible in a screenshot. |
“Transcribe this recording.” | Turn a short WAV recording into a draft transcript, with the optional Whisper component installed. |
She can also help your coding agent review files and draft changes. Your agent handles commands and applies changes after review; Malyn reads the files and prepares her suggestions.
Related MCP server: CascadeGateway
Get started on Windows
This preview is built around Windows 11 and the Intel Arc B580 with 12 GB of memory, using Gemma 4 12B by default. Connect your GPU and install its driver first.
Download and extract Malyn-0.0.68-Betaz-Windows-x64.zip.
Run
Install-Malyn.cmd. Setup downloads the local runtime and model if needed, checks that they work, and connects Malyn to Codex.Restart Codex, then try asking, “Malyn, check my system logs.”
Open Malyn from her tray icon to choose your settings and enable the Windows features you want.
Allow about 7.2 GB for the default model, plus space for the runtime and your results. Setup reuses existing downloads. Transcription and embeddings are optional downloads.
For another model, a different GPU, optional components or installation from source, see the setup guide.
Ask her from Windows
Right-click a file. Enable Ask Malyn in Preferences, then right-click a text/code file or a PNG, JPEG or WebP image. On Windows 11, choose Show more options → Ask Malyn. She'll work through the selected file and open a page where you can follow her progress and read the result.
This preview accepts text/code files up to 2 MB and images up to 10 MiB. PDF, Office documents and other binary formats need conversion first.
Give her a screenshot. Switch on Screenshot mode, then copy a screenshot with Snipping Tool. She'll analyse it locally and request a Windows notification when she's finished. Windows notification settings may hide that notification.
Screenshot mode is off by default. While enabled and her tray is running, she sees newly copied images from any app; she doesn't read clipboard text or take screenshots by herself. Saved images and results stay on your computer until you delete them. If she's busy or paused, copy the image again when she's ready.
Keep her close. Use the tray to pause new work, open settings or check her status. A job already running can finish after you pause her. You can choose to start her when you sign in to Windows. See the desktop guide for more.
Your files and privacy
Malyn processes files locally and keeps detailed results on your computer. You can open and read them yourself. If your cloud AI agent reads a result, its contents can enter that provider's context. Filenames and paths can also be shared with your agent, and anything you upload to a cloud chat has already left your computer.
Give her only the files or small folder needed for the task. Her file controls aren't a complete secret scanner or an operating-system sandbox. Read the data and security guide for the boundaries.
Give her a new recipe
Recipes are reusable instructions for a job. Each is a small Markdown file that describes the work, the model she needs and how to check the result. You can cook up your own and share them.
She comes with recipes for file summaries, Windows logs, image descriptions, audio transcription and text/image embeddings. The embedding recipe prepares data for developers creating search tools; it doesn't include a search app.
See how recipes work, example jobs and optional audio and embedding components.
About this preview
Release: 0.0.68Betaz. Package and plugin metadata use 0.0.68-Betaz. This is an early Windows release tested on a B580. Fresh-account/fresh-machine installation and GPU disconnect/reconnect still need verification. Other GPUs and model combinations are experimental.
Codex has been tested. Claude Code and other compatible MCP clients can be configured, but they haven't been validated for this release. See connecting your agent.
One job at a time. Malyn needs her configured GPU. She won't silently switch to your CPU or a cloud model, and she doesn't queue new jobs while busy.
The demo shows the 0.0.41 development build. It uses real product footage and synthetic files. Processing waits are cut out and labelled; it isn't a speed benchmark.
Cook with her
Her Kitchen is where your recipes live. Import a Markdown recipe, read what it does and export a copy to share or adapt. Kitchen and recipe import/export are included in 0.0.68Betaz.
For model settings and advanced options, see configuration and the setup and command-line guide. To compare speed or token use, see measurement; local token counts aren't a measure of cloud tokens saved.
Contributions are welcome. Start with contributing and the release guide.
Malyn's code is MIT licensed. Bundled dependencies keep their license notices. LM Studio, optional runtimes and model downloads have their own terms.
Available Tools
1 toolmalynB
Malyn helps with bounded local AI work using shareable Markdown recipes in her Kitchen. Use recipes to discover them, recipe to read one, import_recipe with recipe_file to add a reviewed local Markdown file, and export_recipe to get its shareable filename and Markdown. Import never runs the recipe. An image plus @malyn defaults to describe-my-image; Windows log checks use windows-system-logs. Use models/configure for local model choices and submit with recipe/root/files/images to run. Malyn checks and loads a compatible installed model. check returns report links by default; open or quote reports only when the user asks for details. Use returnMode summary only when requested. Collection steps belong to the supervising agent. Legacy loop actions remain supported.
| Name | Required | Description | Default |
|---|---|---|---|
| loop | No | ||
| root | No | ||
| task | No | ||
| files | No | ||
| action | Yes | ||
| images | No | ||
| job_id | No | ||
| recipe | No | ||
| modelKey | No | ||
| settings | No | ||
| loop_file | No | ||
| returnMode | No | ||
| recipe_file | No |
TDQS
B3.1/5.0
Behavior3/5
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description must carry the full behavioral burden, and it does disclose important traits: import never runs the recipe, Malyn loads a compatible installed model, check defaults to report links, and collection steps belong to the supervising agent. It still omits destructive effects of import/configure/cancel, permission needs, rate limits, and what status or cancel actually do.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Conciseness4/5
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is front-loaded with purpose and then packs the main action guidance into roughly 140 words, which is efficient for a 14-action tool. It is dense and jargon-heavy, but most sentences add routing or behavioral value rather than repeating structured fields.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Completeness2/5
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Without annotations or an output schema, the description must cover a complex 14-action, 13-parameter tool on its own. It explains several core workflows but leaves many actions, most parameters, and nested settings undocumented, so an agent would have to guess at large parts of the contract.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Parameters2/5
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 0%, so the description must compensate for 13 parameters including a nested settings object. It mentions recipe_file, recipe, root, files, images, and returnMode, but leaves action, task, job_id, loop, loop_file, modelKey, and all settings fields unexplained. This is not enough semantic coverage for a 0%-documented schema.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Purpose4/5
Does the description clearly state what the tool does and how it differs from similar tools?
The description states a specific overall purpose: bounded local AI work using shareable Markdown recipes, plus action-specific routing for recipes, import/export, models/configure, submit, and check. The opaque tool name 'malyn' is clarified, though several enum actions such as status, cancel, loops, and components remain unstated.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Usage Guidelines3/5
Does the description explain when to use this tool, when not to, or what alternatives exist?
It gives explicit routing for key workflows: recipes to discover, recipe to read, import_recipe with recipe_file, export_recipe, models/configure for model choices, and submit to run. It also says check returns links by default and returnMode summary only when requested. However, guidance is absent for status, cancel, loops, loop, install_loop, and components, leaving action selection only partially covered.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
1 tool update
v0.0.42+kitchen.1- First observed
malyn
TDQS
C2.8/5.0
Scored across 1 tool
Disambiguation2/5
There is only one tool, so nothing can be misselected by name, but the tool internally multiplexes many distinct operations (recipes, recipe, import_recipe, export_recipe, models/configure, submit, check, open, quote, plus legacy loop actions) that an agent must disambiguate purely by reading prose. The boundaries between when to use submit vs. recipe vs. check vs. open are only vaguely stated, making action selection error-prone.
Naming Consistency2/5
The single tool name 'malyn' is a proper noun, not a verb_noun action name, and the sub-actions referenced inside it mix conventions (snake_case import_recipe/export_recipe, slash-style models/configure and recipe/root/files/images, and free-form verbs like submit, check, open, quote). There is no predictable naming pattern for an agent to generalize from.
Tool Count2/5
The described domain spans recipe discovery, import, export, execution, model configuration, and report retrieval, yet the entire surface is exposed as one mega-tool. That is far too few tools for the apparent scope and forces all routing into a single parameterized entry point.
Completeness2/5
The underlying lifecycle (discover, read, import, export, configure, run, check/report) arguably exists, but as a single opaque tool none of these operations is independently discoverable or typed, so agents must infer valid action names and parameters from a dense paragraph. This creates real dead-end risk for anything not explicitly enumerated, such as editing or deleting an imported recipe.
Maintenance
ActivityMaintained
ResponsivenessNo issues
Related MCP Connectors
OCR, transcription, file extraction, and image generation for AI agents via MCP.
Connect AI assistants to authorized Windows and macOS devices for files, commands, and management.
Private-by-default, local-first memory/context/task orchestrator for MCP apps and agents.
Remote MCP server exposing 330 production AI-agent services for web/data processing, validation, AI utilities, blockchain/crypto utilities, and x402 pay-per-use access.
Related MCP Servers
- AlicenseAqualityAmaintenanceProvides an MCP endpoint that lets clients like Google Antigravity and Claude Desktop dispatch LLM tasks to local GPU models with automatic hardware-adaptive routing and cloud fallback, enabling near-zero token cost execution.34MIT
- AlicenseNot gradedqualityCmaintenanceIt enables MCP-compatible AI clients to run local device, filesystem, shell, PowerShell, Git, browser, HTTP, clipboard and application actions on a Windows PC and a paired Android phone, with session-based device routing. Sensitive operations are gated behind configurable one-use approvals and recorded in a local SQLite audit.MIT
- AlicenseAqualityFmaintenanceLets MCP clients delegate private or bulk text tasks to a local OpenAI-compatible LLM while keeping data on the machine.11MIT