Installation and usage guide

Before you start

Culon Cowork is a 64-bit Windows desktop application. The application contains its document tools, OCR resources, Chromium and an embedded Node.js runtime, but it does not include a language model.

You need

• A Windows x64 computer. Windows 10 or 11 is recommended.
• About 131 MB for the installer and additional space for the installed application.
• One model connection: Ollama, LM Studio, a company OpenAI-compatible endpoint, OpenRouter or DeepSeek.
• A separate test folder for the first run. Do not begin with your only copy of critical documents.

Download and install

1. Download the installer only from this page. Culon-Cowork-Setup-0.1.0.exe
2. Optionally verify the SHA-256 before running it.
3. Run the installer, choose English or Russian, accept the license and select the installation folder.
4. Keep the Start menu and desktop shortcuts enabled if you need them.
5. Launch Culon Cowork after installation.

Verify the installer in PowerShell

Get-FileHash "$env:USERPROFILE\Downloads\Culon-Cowork-Setup-0.1.0.exe" -Algorithm SHA256

The result must be exactly:

BB320F39BE863F5A127F8CD4BB6D63BED7709997C6308F1363662C122EB8DCB9

Connect a model

Open Settings → Model. Choose a preset, fill in the model id and use the connection test. Cowork attempts to discover the context window and image-input capability from provider metadata.

Ollama

1. Install Ollama and download a tool-capable model.
2. Make sure Ollama is running.
3. In Culon Cowork choose Ollama and enter the exact model id.

ollama pull qwen2.5
ollama list

LM Studio

1. Load a model in LM Studio.
2. Open the Developer / Local Server screen and start the OpenAI-compatible server.
3. Choose LM Studio in Cowork and copy the model id shown by LM Studio.

Company or custom OpenAI-compatible endpoint

Choose “OpenAI-compatible endpoint”. Enter the base URL that exposes /chat/completions, the exact model id and a key if your gateway requires one. Examples include vLLM, TGI and LiteLLM.

OpenRouter / DeepSeek

Select the provider preset, paste your own API key and choose the model id. The key is returned to neither the renderer nor the settings form; on Windows it is stored through Electron safeStorage when OS encryption is available.

Run the first task

1. Create a test folder and put copies of two or three documents into it.
2. Click New task and choose that folder.
3. Keep the permission mode at “In folder”.
4. Describe a concrete output, not a vague conversation topic.
5. Review permission cards on the right and approve only expected actions.
6. Open the ready file from the result panel.

Good first prompts

Read all PDF files in the folder, create a one-page summary with a source table, and save the result as Summary.pdf.

Compare Contract-v1.docx and Contract-v2.docx. List all substantive differences and save a review-ready DOCX with tracked changes.

Use Quick chat only for short disposable questions. Its history is cleared when you leave the section, and it has no workspace or MCP access.

Permission modes

Documents and images

PDF

Plain reading uses the built-in PDF extractor. If a valid PDF has no text layer, Cowork automatically attempts offline Russian and English OCR. Advanced tasks can inspect form fields, fill and flatten forms, analyze text geometry, create print-ready PDFs and perform PDF transformations.

DOCX / XLSX / PPTX

Cowork uses format-specific playbooks. It can read Office files directly, reuse company templates, unpack OOXML for precise edits, validate and repair common package errors, and rebuild the final file. New documents are generated with the embedded Node runtime and bundled libraries.

Images and screenshots

Paste or drag up to five PNG, JPEG or WebP images into a message. A vision-capable model receives the images directly. With a text-only model, Cowork attempts offline OCR and sends the recognized text instead.

Projects and durable memory

Create a project for work that continues across several tasks: a client, contract, tender, rollout or internal process. A project binds its workspace, tasks, history and context.

CULON.md — project rules

Write stable instructions that should apply to every task: tone, templates, approval rules, naming conventions and required sections. Cowork creates a comment-only starter file without overwriting existing content.

# Project rules
- Use the corporate template template.docx.
- All dates use YYYY-MM-DD.
- Never overwrite source files; save results under Results/.

memory.md — durable facts

Cowork can save facts useful in future tasks: requisites, signers, formatting preferences and recurring constraints. You can edit or delete this plain-text file at any time.

.culon/skills/<name>/SKILL.md — reusable workflows

Store a specialized workflow as a project skill. Cowork discovers it in the workspace and loads it only when the task requires it or when the request starts with /skill-name.

Scheduler

A scheduled task is a regular Cowork task with a workspace, permission mode, reasoning and optional web search, launched at a specified time.

1. Open Scheduler.
2. Choose Once, Every day or Every N minutes.
3. Select the workspace and describe the output.
4. Choose “In folder” for file work or “Everywhere” only when trusted MCP/external actions are required.
5. Enable web search only if the report needs current online data.
6. Keep Culon Cowork running in the background or tray and keep the model endpoint available.

Web search and downloads

Turn on the globe toggle for tasks that need current information. Cowork can search with its built-in hidden Chromium, read selected pages without navigation boilerplate and download a file into the selected workspace.

Find the current prices and technical specifications for the three listed products. Use primary sources, include URLs and the access date, then update Report.docx.

Web search is read-only. A separate web download action writes a file and therefore follows file permission rules. The download limit is 50 MB per file.

MCP extensions

MCP connects Culon Cowork to external tools. Local servers use stdio; remote servers use Streamable HTTP. Open Settings → Extensions (MCP), click Add and paste one or more mcpServers entries.

Local stdio example

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "C:\\Work"]
    }
  }
}

Remote Streamable HTTP example

{
  "mcpServers": {
    "company-tools": {
      "url": "https://mcp.example.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      },
      "timeout": 30000
    }
  }
}

1. Paste the JSON and add the server.
2. Enable its card.
3. Start a new task so the available tools are refreshed.
4. Use “Every step” while validating an unfamiliar server.

History, archive and rollback

Normal tasks are saved locally and appear in Task history. The timeline includes requests, answers, reasoning status, tool calls, permission decisions, terminal state and token usage.

Return to a completed point

Use the rollback action on a completed response. Cowork validates its checkpoint first, restores files that existed before the step, deletes result files created by that step and removes later conversation entries.

Archive versus delete

Archive hides a task or project from active lists and allows restoration. Delete removes the saved history permanently. Deleting a project also removes its linked tasks.

Troubleshooting and privacy

The model test fails

• Confirm that Ollama or LM Studio is running and that the Base URL includes the correct port and /v1 when required.
• Copy the exact model id from the provider. Display names and file names may not be valid ids.
• Check the API key for a cloud provider and the firewall for a company endpoint.

The model answers but does not create files

• Select a workspace folder; file creation and commands are blocked without it.
• Use a model with reliable tool calling and enough context.
• Ask for a specific output format and file name.
• Review pending permission cards on the right.

OCR is weak

• Use a higher-resolution scan with a straight page and strong contrast.
• Ask Cowork to mark low-confidence fragments instead of guessing them.
• For diagrams or photos, select a model with verified vision support.

Scheduled tasks do not run

• Keep the application running in the tray.
• Keep the model endpoint and required network connection available.
• Do not use “Every step” for an unattended task.

Where data goes