Skip to content
DogubakoDogubakoDogubako

Plain words

Plain words lets you type what you want done, such as “convert this PNG to lossless WebP” or “strip the location from these photos”, where you would normally pick a tool. A language model you choose reads the request and runs or opens the right Dogubako tools for you.

It is off by default and sends nothing anywhere until you set it up.

Turn it on

Section titled “Turn it on”

  1. Open Settings → Plain words and switch on Do things in plain words.

  2. Pick a Model provider:

    Provider Base URL Key
    OpenAI https://api.openai.com/v1 Required
    Anthropic https://api.anthropic.com/v1 (its OpenAI-compatible API) Required
    OpenRouter https://openrouter.ai/api/v1 Required
    Ollama http://localhost:11434/v1 Not needed
    LM Studio http://localhost:1234/v1 Not needed
    Custom Any base URL Depends on the server
    This server Set by the administrator (self-hosted only) Kept on the server

  3. Enter your API key if the provider needs one, then choose a Model. Each preset suggests models, and you can type any model id, including your own fine-tunes. Load models fetches the endpoint’s own list.

  4. Click Test connection. It sends a one-word request and shows the reply or a readable error.

Any endpoint that speaks the OpenAI chat-completions API with tool calling works. Pick a model that supports tool calling, or the requests will fail.

Use it

Section titled “Use it”

Type into the search (⌘ K), the box on Home, or the What should happen to it? screen after you drop files. Short queries still search tools as before. When you type a sentence, Do it comes first in the list, and ⌘ Enter (Ctrl Enter) runs it directly.

What happens next depends on the request:

You type Dogubako
A clear request: “convert this PNG to lossless WebP” Runs the tools in place. Each step is listed and can be expanded to show its result, and the outputs appear underneath
A tool name, or a request that leaves a choice to you: “convert this file” Opens that tool with your files and every setting it is sure of filled in, plus a one-line note on what is left to pick
A question a tool can answer: “what DNS records does example.com have” Runs the tool and answers in a sentence

A model switcher sits next to running requests, so you can retry with a different model without going back to Settings.

Undo

Section titled “Undo”

Undo removes what a request made (its output files and History entries) and takes you back. Your original files are never changed by an ordinary run, and every step is also recorded in History.

If a step did change files on disk, such as a rename, a delete or a write to a host folder (always after your OK), the button reads Remove results instead. It removes only the new result files, tells you that the other changes stay, and keeps those steps in History.

Confirmation prompts

Section titled “Confirmation prompts”

Steps that use the internet, write to host folders, read the list of processes or change files always stop and wait for Run or Skip, with every value shown in full. Under Ask before running in Settings you can choose:

  • Risky steps (the default): only those steps wait for you.
  • Every step: you approve every tool run.

Opening a tool on your behalf never pre-fills a tool that would need your OK, because some tool pages run as soon as they open.

What is sent where

Section titled “What is sent where”
  • To the model provider: your request, the names of the files you gave it, and tool results the model needs to answer (such as extracted text).
  • Not sent: file contents. Dogubako processes the files itself, and the model only sees results. With a local model (Ollama, LM Studio), nothing leaves your machine.
  • API keys are saved in the app or browser on your device only. On a self-hosted server they travel with each relayed request and are never stored; signing out forgets them.
  • The model can only pass the files you gave it, or outputs of its own runs, to tools. Any other file id or path is refused, so instructions hidden inside a document can’t make it read other files. Tool results are marked as data, not instructions, in the prompt.
  • Tools run through the normal run API, so the same validation, permissions, quotas and audit log apply as for a manual run.
  • Finished requests are kept for the browser tab only when History is on, with passwords and keys masked.

On a self-hosted server

Section titled “On a self-hosted server”

The browser can’t reach most model endpoints directly (the server’s CSP allows only same-origin requests, and local runtimes send no CORS headers), so the server relays model requests at /api/v1/assistant/*. The desktop app relays them itself.

The relay:

  • follows the network policy. It reaches public addresses only, unless DOGUBAKO_ALLOW_PRIVATE_NETWORK or DOGUBAKO_ASSISTANT_ALLOW_PRIVATE is set. Set the second one to use Ollama or LM Studio on the same host or LAN.
  • is for people only. Agent accounts get 403; they bring their own model through MCP.
  • needs the CSRF token, allows at most 4 requests at once per user, and writes an audit entry with the model id and endpoint host only.

An administrator can offer one shared model to everyone as This server:

Terminal window
DOGUBAKO_ASSISTANT_BASE_URL=https://api.openai.com/v1
DOGUBAKO_ASSISTANT_MODEL=gpt-6-luna
DOGUBAKO_ASSISTANT_API_KEY=sk-…        # stays on the server, never sent to browsers

When a model is set, every request through This server uses it, so people can’t spend the administrator’s key on other models. Set DOGUBAKO_ASSISTANT_ENABLED=false to turn plain words off for the whole server. The Docker setup includes an optional Ollama service for a fully local model.

See Configuration reference for every setting.