Verzeta^™ · local-first AI team workspace

Your AI team, running on your machine, in one chat.

Verzeta Studio is the host app. Create a project room, add named members, give each member its own model and tools, then watch the team hand off work, produce artifacts, vote on decisions, and leave an audit trail. Verzeta for Android and Verzeta for VS Code pair in as wire clients.

See the workflow Read the docs

No telemetryNo cloud syncNo accountsOpen source

Verzeta Studio

Launch Strategy Teamproject group chat · VERZETA LAUNCH

anthropic/claude-opus

You

@Alice, build the release narrative. Ask the right named members for research, writing, engineering review, and a launch decision.

Product Manager · @Alice

I will own the launch goal. @Maya, collect the provider, Android, skills, project, and activity evidence before @Clark drafts.

anthropic/claude-opus · 144 tokens

Researcher · @Maya

I will gather the evidence Alice requested and hand the notes back into the launch brief.

anthropic/claude-opus · 127 tokens

Writer · @Clark

I will turn Maya's notes into a Project Brief, then ask @Robin for implementation risk review.

anthropic/claude-opus · 110 tokens

Group chat· @Alice, @Maya, @Clark, @Robin · use @everyone to broadcastVerify agent work: Project Brief, decision poll, activity timeline, tool log

Ask the launch team to continue

Verzeta Studio · Project Rooms

Start a working AI team from a room template.

Project Rooms turn setup into a guided workflow. Choose a room, start from a curated template, edit the goal and teammates, then launch a project with roster, tasks, artifacts, and tools ready.

Choose a room or templateBrowse existing rooms, filter templates by workflow, or start a blank project.

Project RoomsHome · New project

Search projects and templates

Your rooms 2

idle

Launch Strategy Team

Campaign, release notes, release evidence

AMCR

idle

Sales Pitch Workshop

Opening hook, objections, customer stories

PRSA

Start from a templateAllSalesEngineeringMarketingExecDiscovery

Research

Research Brief

A researcher gathers material from multiple sources, an analyst structures the findings, and a reviewer checks the claim trail.

PRAC

Discovery

Customer Interview Synthesis

A researcher reads through every transcript and pulls out direct quotes. An analyst clusters the quotes into themes.

PRAW

Eng. review

Code Review War Room

An engineer reads the diff deeply and identifies bugs. A reviewer applies senior-engineer judgement on risk.

EECE

Eng. planning

Technical Spec

An engineer drafts the proposed design. A designer makes sure the user-facing surfaces and trade-offs are clear.

EEDC

Marketing

Marketing Copy Studio

A writer drafts every variant with a clear voice. A critic flags weak claims, vague benefits, and tone drift.

PWCA

Product launch

Product Launch Plan

A writer owns the launch narrative across surfaces. A designer makes sure landing-page and email visuals align.

PWDC

Local first

One install is the whole product.

Many agent workspaces need a separate server before the app can do anything. In Verzeta the app is the server: install it, add a model provider, and your team is ready. Want it always on? Run the same app headless on a home server or spare machine and pair your devices to it. Nothing leaves your machines unless you use a cloud model or web search.

Requirement	Verzeta Studio	Server-based agent workspaces
What you install	One app, nothing else	An app, plus a server or a hosted account
Separate server required	No, the app is the server	Yes: your own, or the vendor's hosted one
Where it runs	Your own machine	Your server or the vendor's cloud
Database and storage	A local file, built in	Run by whoever hosts the server
Account or identity	None	Needed to join the workspace
Where conversations live	On your own machine	On the server
Working with local models	Anywhere, even offline	While the server is reachable
Always-on server (optional)	The same app, run headless	A separate server, yours or the vendor's

A team that stays on track

Turn limits, rounds, polls and task plans coordinate your agents, so a conversation always moves forward and always ends.

Careful by default

Agent commands go through an allow-list and safety checks, with optional write protection on Linux.

Native on your desktop

Built with Qt 6 and KDE libraries, not a browser in a window, and follows your desktop theme.

Provider routing

Each member can use a different model in the same conversation.

Verzeta supports cloud APIs, remote servers, local GGUF models, and any OpenAI-compatible server. Provider choice happens per member, not only per app.

@AliceProduct Manager

Ollamaqwen3:14b

@ClarkWriter

Anthropicclaude-opus

@RobinEngineer

OpenRouterdeepseek-chat

@MayaResearcher

Geminigemini-flash

Cloud providersOpenAI, Anthropic, Gemini, OpenRouter, DeepSeek

Remote serversOllama and llama.cpp over HTTP on your machine or LAN

Local modelsGGUF models run by the bundled llama.cpp engine on your machine

Any OpenAI-compatible servervLLM, LM Studio, TabbyAPI, and more, each as a first-class provider

Tool whitelists

Permissions are part of the team design.

The app can expose tools globally while each member sees only the tools allowed for that member. This lets one agent review, one write, one search, and one coordinate.

Tool surface@Alice@Clark@Robin@Maya

Read files

Write files

Shell commands

Web search

Canvas

Polls

Skills

Cascade chat

The team hands off work by @alias, where you can see it.

One message can trigger a structured chain: mention routing, intent checks, per-member model dispatch, and a bounded cascade that stays inspectable.

Group chat · @Alice, @Clark, @Robin, @MayaProvider: per member

You@Alice, prepare a launch plan. Ask the team for writing, feasibility, and sources.

@Alice · Product Manager@Clark draft the brief. @Robin check implementation risk. @Maya gather supporting notes.

@Maya · ResearcherI collected the provider, skills, project, Android, and activity timeline notes.

@Clark · WriterI opened launch-brief.md in the canvas and wrote the first pass.

@Robin · EngineerFeasibility check complete. I recommend a poll before final positioning.

Artifacts, polls, timeline

The conversation leaves evidence behind.

Verzeta is strongest when a team produces something concrete and you can inspect what happened afterwards.

Canvas · Project BriefRevision 4

# Launch plan

- Position Verzeta as a local-first AI team workspace.
- Lead with privacy: no telemetry, no cloud sync, no accounts.
- Show Android and VS Code as wire clients of the desktop host.

Decision pollOpen

Build demo first

Publish docs first

Ship client story first

Activity timelineAppend-only

agent_turn@Alice routed the launch requestOllama / qwen3:14b

tool_invoked@Maya searched and saved notessearch_web success

canvas_edited@Clark updated launch-brief.mdrevision 4

poll_vote@Robin voted for Build demo firstsingle-choice poll

file_written@Clark wrote final-summary.mdartifact folder

Skills

Turn repeatable work into approved team playbooks.

Skills extend what the team can do, under your review. Install a domain workflow, read what it instructs, approve it, then expose it only where that project or conversation needs it.

Skills

Skills are reusable instruction bundles that agents can follow. They never grant tools and never run scripts on their own. Every install stays Blocked until you review and approve it.

Search the skill catalog. Every install lands as Blocked until you review and approve it. Agents can only use a skill once you approve it.

release-notes-from-repoBlocked · waiting for review

---
id: release-notes-from-repo
description: Extract user-facing release notes from repository evidence.
tags: [release, changelog, writing]
---

When the user asks for release notes:
1. Inspect the changed files and docs.
2. Separate user-visible changes from internal work.
3. Call out risks, setup steps, and upgrade notes.

01Browse or import

Find a skill in the skill catalog or drop a local skill folder into the app data skills directory.

02Review SKILL.md

Read the instructions, tags, declared workflow, and supporting files before the model can use it.

03Approve to unblock

New installs stay Blocked until you review and approve them. Agents cannot use a skill until it is Approved.

04Expose by scope

Make the approved skill global, project-only, or conversation-only so the right agents see it.

05Agent discovers it

The prompt layer marks skills READY or BLOCKED, then agents read and apply them when relevant.

Host and wire clients

One host runs the work. The clients are remote surfaces.

Verzeta Studio runs the tools, agents, and conversation state on your machine and sends every model request. Verzeta for Android and Verzeta for VS Code pair in over a WebSocket on your network, with optional TLS, and show the same team. No Verzeta cloud service sits in between, and the clients never talk to a model provider directly.

Verzeta StudioDesktop host · the control plane

Connects every model and provider
Runs tools under an allow-list you control
Routes agent handoffs and cascades
Holds conversations, projects, and the activity log

WebSocket · port 9180 · TLS for remote 6-digit pair code · self-signed certs pinned by fingerprint

Verzeta for VS CodeChat, editor actions, and a workspace mount so agents work on your real code.

VS Code Marketplace Open VSX

Verzeta for AndroidThe same conversations and rooms on phone, tablet, desktop, and TV. The host always does the work.

Verzeta for VS Code

Bring the whole team into the editor you already use.

Verzeta for VS Code pairs with the host app and shows your conversations and project rooms in the sidebar. Register a workspace as a virtual mount on any conversation, group or 1:1, and the team can read, edit, and create files as if the project were local. Editor actions route a selection straight to a member, and the extension can generate commit messages and send terminal output into the chat. No model runs inside VS Code; the host does the work.

VS Code Marketplace Open VSX

payments-service - Visual Studio Code

VERZETA

Verzeta

Home Chat Settings

ConnectedWorkstationws://localhost:9180/

Project RoomsMulti-agent teams on shared goals. Pick a room or start one.

RECENT · WORKSTATION

Launch Strategy TeamGroup · 2m

Research Brief · market scanGroup · 18m

Code Review · payments APIGroup · 1h

Release notes v0.2Chat · 3h

Competitor analysisChat · Tue

Onboarding email copyChat · Mon

main.cpp build.sh

12345678

int main(int argc, char** argv) {
  Host host = Host::open(argv[1]);
  while (host.running()) {
    auto turn = host.next();
    route(turn);            // hand off by @alias
  }
  return 0;
}

PROBLEMSOUTPUTDEBUG CONSOLETERMINAL

[10:42:18] workspace-mount registered: payments-service (44 files, tier: ask)[10:42:19] wire-sync: @Robin opened main.cpp through the mount

main0 errors, 0 warnings

Ln 5, Col 18Spaces: 2UTF-8LFC++ Verzeta: Workstation

Editor actions

Select code and route it to an agent from the lightbulb menu.

Explain Fix bugs Refactor Add tests Add docs Fix this problem

Workspace mount

Register a VS Code workspace as a virtual mount on a chat, group or 1:1. The team works on the real project files under your control. Writes follow a permission tier, with a built-in secret blocklist and an optional path allowlist.

AskSmartBypass

Smart auto-applies clearly safe edits and escalates anything large or unusual to Ask.

Command safety

Shell commands follow a per-conversation policy and run sandboxed where the platform supports it.

OffAskAllow

Off refuses commands, Ask confirms each one, Allow runs without a prompt. Commands run under bwrap on Linux when it is available; otherwise you are told the command ran unsandboxed.

Verzeta for Android

Carry the same AI team across your devices.

Verzeta for Android pairs with Verzeta Studio over a WebSocket, with TLS for remote access. It runs on phone, tablet, desktop, and TV, streaming conversations and project rooms so you can send messages, manage members, vote in polls, and review activity. The device never runs models or tools. The host does that.

6-digit pair code, self-signed certs pinned Live conversation streaming Rooms, polls, members, activity

VVerzeta

CONNECTEDWorkstationws://192.168.1.20:9180/ws

RECENT · WORKSTATION

Launch Strategy TeamGroup

Research Brief · market scanPlan Group

18m

Code Review · payments APIPlan Group

Competitor analysisAgent

HomeChatSettings

Research Briefollama · qwen3:14b

Loaded 8 messages

VResearcherqwen3:14b@Maya

I pulled twelve sources for the market scan and tagged each claim with a citation. Handing the notes to @Clark for the brief.

VWriterqwen3:14b@Clark

Draft one is in the canvas, every claim linked to a source. @Robin, can you sanity-check the data points before we ship?

Ask anything

The desktop host runs models and toolsAndroid is the remote surface, on phone, tablet, desktop, and TV.

System design

Built for coordinated AI work.

Aliases, projects, providers, tools, skills, polls, artifacts, wire clients, and timeline evidence work together as one local-first system.

The team is explicit

Verzeta models the team directly: named aliases, reusable templates, coordinator behavior, member routing, and visible handoffs.

Provider choice is part of the team

One project can mix local Ollama, cloud APIs, remote servers, local GGUF models, and any OpenAI-compatible server instead of forcing one model across the workspace.

Permissions are designed per member

Tool whitelists let one member search, another write, another review, and another coordinate without giving every agent every capability.

Work leaves inspectable evidence

Canvas artifacts, plans, polls, files, and activity timeline entries turn a chat into a traceable project run.

Local-first by default

No telemetry, no cloud sync, no accounts. The desktop stays the control plane and the clients are paired surfaces.

Skills need your approval

Skills are reviewed before use and only exposed where they belong, so repeatable workflows never run unseen.

Capability map

Named members, providers, tools, and a record of the work.

Named members

Aliases like @Alice and @Robin are the routing handles, not fixed job titles.

Reusable templates

Product Manager, Researcher, Writer, Reviewer, or your own, used more than once.

Provider routing

Each member can use a different provider and model in one chat.

Tool whitelists

Allow tools per member instead of opening every capability to everyone.

Project rooms

Goals, shared documents, and rosters travel with the room.

Skills

Approve an instruction bundle, then scope it to a project or conversation.

Polls

Members and you vote when the team needs a decision on the record.

Activity timeline

Replay turns, tool calls, votes, file writes, and member changes.

Verzeta for VS Code

Chat, editor actions, and a permissioned workspace mount inside the editor.

Verzeta for Android

The same team on phone, tablet, desktop, and TV while the host does the work.

Sub-agents

Members delegate self-contained side tasks to private background workers that report back.

Heartbeats

Schedule background agent runs so the team keeps working on a cadence you set.

Conversation memory

Long group chats stay coherent through automatic compaction. Nothing you see is deleted.

Headless server

Run the host without the desktop UI on an always-on machine and pair clients to it.

Documentation

User docs for Verzeta Studio, Android, and VS Code.

The guide is synced from each product repository. Pick a surface, then read setup, providers, teams, projects, editor integration, pairing, and troubleshooting in one place. Search spans every surface at once.

Introduction to Verzeta Studio

What it is

Verzeta Studio is a local-first AI chat workspace where several AI agents work together as a team in one conversation. Each agent has its own model, role, system prompt and tool list.

The app is open source under a triple-license arrangement (GPL-3.0-or-later for the differentiated components, LGPL-3.0-or-later for the supporting infrastructure, and a commercial license for organisations that want neither copyleft variant). It runs on Linux and Windows, and pairs with two wire clients, Verzeta for Android and Verzeta for VS Code, over a connection you can encrypt with TLS.

Key benefits

Multi-agent group chat. Run multiple AI agents in one conversation. Each agent can use a different provider (e.g. Alice on Ollama locally, Bob on Claude, Carol on Gemini) and a different tool whitelist. They route work to each other with an @mention.
Local-first. Your conversations, attachments, and skills are stored on your own machine in a SQLite database. API keys stay in your local settings in obfuscated form (see how keys are stored). Local models via Ollama and llama.cpp run on your hardware without sending data to anyone.
Eight providers supported out of the box. Cloud: OpenAI, Anthropic, Gemini, OpenRouter, DeepSeek. Self-hosted servers: Ollama, llama.cpp (HTTP). Built-in: a local llama.cpp engine in the Local AI edition. Any OpenAI-compatible server can be added as well.
Real agent capabilities. Agents can read and write files, run allowed shell commands, search the web, edit code in a side canvas, run multi-step tasks, use the skills you approve, and vote in polls. Every action is recorded in the activity log.
Skills. Installable instruction packs (a SKILL.md file with frontmatter, plus optional supporting files). Agents discover them and apply them when relevant. Each skill is approved by you before any agent can use it.
Polls. When the team needs a decision, they open a poll instead of looping in discussion. Votes are recorded, and the chat keeps moving.
Activity timeline. Every agent turn, tool call, file write, member change, and poll vote is recorded. Open the timeline for a project or a single conversation to see what happened.
Mobile companion. Pair an Android device with your desktop using a one-time code. Continue chats on your phone over a WebSocket connection, encrypted with TLS when you turn it on.

Use cases

Drafting a document. A writer agent drafts, a reviewer agent critiques, and a researcher agent fact-checks. Each step lands as a real file in the project's artifact folder.
Code review. A coder agent edits a file in the canvas, and a reviewer agent runs it and writes a feedback note. The plan tracks unfinished steps.
Research with audit trail. A search agent uses the web and a summariser agent writes notes. Every search query, every file read, and every conclusion is logged.
Team decisions. When two agents disagree, the coordinator opens a poll. The team votes. The decision is recorded on the chat timeline.

Product context

Attribute	Value
License	GPL-3.0-or-later, LGPL-3.0-or-later, or commercial, declared per file. See LICENSING.md
Source code	Open source. Browse and modify under GPL or LGPL, per the SPDX header on each file
Platforms	Linux and Windows
Wire clients	Verzeta for Android and Verzeta for VS Code
Telemetry	None. The app never sends usage data anywhere.
Data location	`~/.local/share/Verzeta/verzeta-studio/` on Linux, `%APPDATA%\Verzeta\verzeta-studio\` on Windows
Connectivity	Required only for cloud providers, web search, downloads and remote pairing

Adapting to your screen

Every dialog has the same layout: a title with an icon and a close button, an optional search box or subheading, and action buttons along the bottom.

On large screens dialogs open as centred cards. On small or touch screens they expand to fill the window, and you can drag a dialog down from its title strip to dismiss it.

Where to read next

New user? Start with 01-quick-start.md to get a working chat.
Want to run a team of agents? 04-multi-agent-teams.md covers the multi-agent workflow.
Need to set up multiple providers? 02-providers.md covers all 8.
Pairing a phone? 08-remote-android.md covers pairing and what Android can and cannot do.

Get Verzeta

Get the desktop host, then pair the clients.

Download a release or build from source. The fastest first run is Ollama locally or one configured cloud provider, then a group chat with named members. Pair Verzeta for Android or Verzeta for VS Code once the host is running.

Releases and source Quick start

Fastest first run1Open releases or source2Configure Ollama or one cloud provider3Create a Project Room from a template4Pair Verzeta for Android or VS Code

Verzeta Studio · Linux

The primary host build. Use GitHub release artifacts or build from source.

View releases

Verzeta Studio · Windows

The Windows host build, from the same source and release channel.

View releases

Verzeta for Android

Pair the wire client with the host on a phone, tablet, desktop, or TV.

See Verzeta for Android

Verzeta for VS Code

Install the extension and pair it with the running host to work inside the editor.

See Verzeta for VS Code