Skip to content

Release v0.2.0

Latest
Latest

Choose a tag to compare

View all tags
@Jeffxz Jeffxz released this 14 Oct 00:28
v0.2.0
d60a54d

Wysebee Desktop v0.2.0

Your agents now think for themselves, inside the app.

v0.1.0 gave your own Claude or ChatGPT a set of nodes to call over MCP, and ran no
model of its own. v0.2.0 turns that around: nodes are now agents, and each one has a
Chat tab where a model you choose (OpenAI, Anthropic, Gemini, a local model and
more) decides which tools to call while the app runs them. Agents can also work
unattended as scheduled Tasks, operate your browser through a new extension, query
your databases and work on files on this computer.

It still runs entirely on your machine. There is still no account and no sign-in, and
your agents, tools and data stay in the same SQLite file and workspace folders as before.

Upgrading from v0.1.0? Read Breaking changes first. Inbound
access was redesigned, so MCP URLs from v0.1.0 stop working.

Install

macOS 12 or later, Apple Silicon or Intel

  1. Download Wysebee-0.2.0.dmg from the Assets below.
  2. Open it and drag Wysebee into your Applications folder, replacing v0.1.0 if it is there.
  3. Launch it from Applications.

Signed with a Developer ID certificate and notarized by Apple, with the ticket stapled,
so it opens without a Gatekeeper warning, offline included.

Windows 10 or later, 64-bit

  1. Download Wysebee-0.2.0-setup.exe from the Assets below.
  2. Run it. It installs for your user only, under %LOCALAPPDATA%\Programs\Wysebee, with
    no administrator prompt, and upgrades v0.1.0 in place.
  3. Launch Wysebee from the Start menu.

The installer, the app and the uninstaller are Authenticode-signed with an EV
certificate and timestamped.

Your data carries over. The first launch upgrades the database in place. It first
saves a backup beside the database as wysebee.db.v<N>.bak. Settings → System shows where
the database is.

Docker Desktop is still optional. Only workspaces need it, and agents without a
workspace run without it.

What's new

Chat with your agents

  • A Chat tab on every agent, and the default tab. The model picks tools; the app
    calls them in-process and shows each step.
  • Choose a model per agent. The choice is saved on the agent, so the browser side
    panel and Tasks use it too.
  • Ten AI providers, set up under Settings → AI Providers: OpenAI, Anthropic,
    Google Gemini, Alibaba Cloud (Qwen), DeepSeek, Mistral AI,
    Sakana AI, Sakura AI Engine, Local (Ollama, LM Studio, llama.cpp, vLLM or any
    OpenAI-compatible server) and Jev (TypeSafe AI), which only runs Decision tasks.
    Keys are stored encrypted and never shown again.
  • Instruction tab: standing instructions, in markdown, added to every conversation
    with the agent.
  • Rich answers. Replies render as markdown, and agents can draw charts (d3, in
    a sandbox), show a list of emails and show a calendar week or day.
  • Screenshots. Models that can see images are shown browser screenshots, and you
    see them in the chat whatever the model.
  • Long conversations are trimmed to fit the model's context window, so they keep
    working. Stop cancels the turn on the spot, mid-model-call or mid-tool.

Tasks

A task is a plan plus its own tools, skills, workspace, browser, database and model.
It runs on its own and finishes with a report.

  • Schedules: start by hand, once at a set time, every N minutes or hours, daily, or
    weekly on the days you pick, in your local time. The app must be open for a schedule
    to run.
  • Run history: every run records its trigger, model, steps and final output.
  • Decision tasks: single-shot, with no text written. Input tools run with saved
    arguments, Jev answers your typed questions (choice, yes/no, score) and the tool you
    attached to each answer runs, but only above a confidence threshold you set.
  • Query tasks: no model at all. They run one SQL statement, typed or rendered from
    a SQL Template tool, against a database connection and return JSON or CSV.

Home dashboard

The app now opens on Home: every agent and task as a card, with chat or start
in one click, and a switcher to move between them.

Browser access: Chrome, Edge and Brave

  • A browser extension installed from Settings → Connections. It lets an agent open
    and close tabs, read pages, click, type, scroll, draw, take screenshots and read
    bookmarks.
  • You stay in charge: per-site trust levels, an approval prompt for anything not
    trusted, a switch to pause agent access, request kinds you can block outright, and
    bookmarks that always ask.
  • Web development tools (console, network, evaluate, viewport, audit) are a
    separate opt-in per agent.
  • A side panel in the browser for chatting with your agents, which keeps going on
    long jobs and checks in with you.
  • The desktop app updates the extension files when it updates itself, and the
    extension tells you when to reload. Each browser can be uninstalled and repaired on
    its own.
  • An agent without browser access tells you how to turn it on instead of just saying
    it can't.

Databases

  • Settings → Connections → Databases: SQLite, PostgreSQL and Snowflake, with
    several sign-in methods each. Secrets are stored encrypted, and there's a Test
    connection     button.
  • Database access on an agent or task gives it a query_database tool, read-only
    unless an admin allows writes. Read-only is enforced by the database itself wherever
    it can be.

New tool types

The tool picker is now in three groups: General, Local files and Database.

  • Notification: shows a message in the app, bounces the Dock or flashes the
    taskbar. Useful for a task to say it's done.
  • Local Files: gives an agent folders you choose, to list, read, write, delete and
    run files. In Chat, every action waits for your Approve.
  • SQL Template: typed parameters filled into a query, each value checked against
    its type. It writes SQL but never runs it.
  • SQL from Schema: import a database's tables, describe them, and the agent's own
    model writes the SQL.
  • Tool parameters can be reordered by drag and drop.

Wysebee Assistant

A helper in a panel on the right (the bee button), offered on first launch. It answers
questions about the app and makes changes for you, each one only after you approve
it. Risky changes, such as deleting something or entering keys, become a
step-by-step walkthrough instead. It can also fill in text on the page you have open,
like an agent's instructions or a task's plan, for you to review and save.

MCP server, rebuilt

Want to use an agent from Claude, ChatGPT or Cursor? Turn on its MCP Server tab.

  • One server on 127.0.0.1, off until you enable an agent.
  • Apps sign in with the agent's 8-character pairing key over standard OAuth.
    Connected apps are listed on the tab and can be disconnected.
  • Repeated wrong keys lock pairing for a while.
  • Start ngrok from the tab to reach it from outside, optionally whenever Wysebee
    starts.

Everywhere else

  • Languages: English, French, Japanese, Simplified Chinese and Traditional Chinese
    (Settings → System). The browser extension follows the same setting. The non-English
    text was machine-translated.
  • Unsaved changes bar: agent, task, tool and skill pages no longer save as you type.
    Use Save or Discard.
  • Closing the window keeps Wysebee running in the background, like Slack, so
    scheduled tasks keep running.
  • Update notice at the bottom of the window when a newer release is out, with a
    download link. Nothing is downloaded or installed automatically.
  • A new side navigation, a relaid-out Settings page, and Remote machines renamed to
    Connections.

Breaking changes

  • Nodes are now agents. Your nodes, their tools and skills carry over, renamed.
  • v0.1.0 inbound access is gone. The /node/<name>/mcp and /node/<name>/tools/...
    endpoints, the per-node Access tab, the Settings → Network publishing switches (ngrok /
    your own proxy), the OAuth sign-in with Settings → Users approvals and the per-node
    Authorization tool have all been removed. Their stored settings are dropped during the
    upgrade. To use an agent from Claude or ChatGPT again, turn on its MCP Server tab
    and pair with its key. The URL and sign-in are new.
  • The app no longer listens on a port for its own UI and API. The MCP server above
    is the only listener, and only while you have it on.

Privacy: what now leaves your machine

v0.1.0 sent nothing. v0.2.0 sends three kinds of traffic:

  • Model calls, to the AI providers you set up, when you chat or run a task.
  • Anonymous usage analytics, sent to wysebee.ai: which pages you open and which
    buttons you press, identified by a random install id. It never includes what you
    type, your keys, or the names of your agents, tools, skills or workspaces, and records
    no IP or location. Turn it off in Settings → System.
  • The update check, a request to wysebee.ai for the latest version number. It
    carries no install id. Turn it off in Settings → System.

Uninstalling

Same as before: removing the app leaves your database and workspaces alone. To remove
the browser extension, use Uninstall … connection under Settings → Connections
first. It tells you which extension folder to remove from the browser.

Assets 4
Loading