AgenticVenus
Agent Works: AI that does the work.

AgenticVenus is an AI agent team. Every chat is a teammate with a name, a role you give it, its own memory and skills, its own connected tools and, when the job needs it, its own cloud computer and phone number.

You hand over work in plain language. The agent plans it, does it, checks it and reports back with proof. You only step in for decisions.

• Founders and operators who want leads, outreach, inbox and reports handled.
• Developers who want features built and tested on a safe computer.
• Agencies and creators who need videos, decks and sites at speed.
• Sales and recruiting teams who need verified lists and follow-up calls.

A chatbot answers. A teammate finishes. It opens real websites, writes and runs code, makes videos, calls people and keeps going after you close the tab.

The same agent answers in the dashboard, on Telegram, WhatsApp, SMS, Slack, by email and on the phone. Memory, role and tools are the same everywhere.

AgenticVenus is young software built by a small team. Expect fast changes. Features marked Beta work but are still being polished.

The agent, tools and interface are open source. The official hosted service keeps plans, limits and payments on its own servers, so they cannot be changed from a browser or a modified copy.

Start with Agent Works to understand the goal, then Quick start. Use the search box (Ctrl or ⌘ + K) to jump anywhere. Every section has a number and a link you can share.

AgenticVenus exists to do work, not to talk about it. The target is that a task a person does at a desk, on a phone or in a browser can be handed to an agent and come back finished.

Work is a task with an owner, an input, a result and a deadline: find 100 leads, answer these 40 emails, build this landing page, call these 20 customers, update this sheet. If it can be described and checked, an agent can be given it.

A chatbot gives text and a person still has to do the task. Agent Works gives outcomes: files, sent emails, finished calls, deployed sites, verified lists.

AgenticVenus can already take over large parts of these jobs, with approvals on risky steps:

• Lead research and list building with verification
• Outbound email drafts and batch sending
• Inbox sorting and replies
• Web research and written reports
• Spreadsheet updates and data cleanup
• Website and app building
• Short motion-graphics videos and decks
• Appointment, follow-up and reminder calls
• Routine reports on a schedule

The roadmap is aimed at removing the remaining human steps:

• Longer unattended missions with less supervision
• Stronger phone conversations with faster interruption handling
• Deeper business-tool coverage through more connectors
• Own agentic model tuned for coding and agent tasks
• Better self-repair of skills so repeat work gets more reliable
• More specialist roles working as one team

1. 1Brief: you describe the goal in plain language.
2. 2Plan: the agent splits it into steps and picks the right tools or specialists.
3. 3Do: it works on its computer, in code, on the phone or through connectors.
4. 4Verify: results are checked by code and by an independent reviewer.
5. 5Report: one message with files, evidence and anything that needs you.

Jobs run as background processes. Close the tab, lock the phone, go to sleep: the report is waiting in chat and on your connected chat apps.

Every computer task can be replayed. Lists show which rows passed verification. Calls come with transcripts and summaries. Nothing is reported done unless it was done.

Goals, taste, relationships, accountability and any decision with legal or financial weight. The agent prepares, you decide. As trust grows you can widen what it may do alone.

Anything that writes, sends, pays or calls can require your yes. You choose how strict this is, and you can approve from your phone with one reply.

1. 1Shadow: give the agent a task you already did and compare results.
2. 2Assist: let it do the work while you approve every write.
3. 3Own: let it run routines and missions alone, approving only exceptions.

Use the Runs page to see what each job did, how long it took and its estimated cost. Compare that with the time a person would have spent.

Sign up with email or Google. No card is needed for the Free plan.

Paste a key from any supported provider. It is stored encrypted, and the browser only ever sees a masked reference.

Press + in the sidebar. Every chat becomes a separate teammate.

Choose Chief, a specialist such as Sales Outbound, or create your own with a name, colour and avatar.

The role is saved in that chat's memory and shapes every answer.

you are my head of sales

Add the cloud-computer key under API keys to unlock Computer, Code, Video and Team modes.

The desktop app shows the real website and can also run a Linux computer on your own PC. See the Desktop section.

Open the site on your phone, or talk to your agent through Telegram, WhatsApp or SMS after connecting a channel.

Open the finished job to see files, the evidence and the replay link of the screen recording.

Name, colour, avatar, role and an optional own email inbox for sign-ups and verification codes.

Mascots in 15 expressions and any colour, plus 150+ buddy shapes, so every teammate looks different.

Say what the agent is to you: head of sales, CEO, assistant. The role is pinned for that chat.

Per-chat memory of facts, preferences, lessons and your owner profile, plus skills with a win and fail record.

Tools from connectors, MCP servers and APIs, plus the computer, Venus Code, Venus Pro, maps and calling.

Approvals for writes, a prompt-injection shield, an encrypted vault, vPassword and an audit log.

Dashboard, Telegram, WhatsApp, SMS, Slack DM, email and phone calls, all with one brain.

At most one proactive message a day: a next step, a reminder or one question to know you better.

Before acting, the agent decides what you actually asked for, and starts heavy work only when you really asked for it.

A work order like "find 100 leads" can never be answered as plain chat, because that is how invented lists appear. It is forced to the computer. And "send these 5 emails" can never open the computer: it uses the mail tool.

After many answers the agent offers short, tappable next steps, so you can continue with one tap.

Choose the model per chat. Computer tasks need a model that can see images; coding and writing work with any provider.

Claude, OpenAI, Gemini, Grok, OpenRouter, Mistral, Cohere, Perplexity, Groq, DeepSeek and Apinex. Bring your own keys.

On a rate limit the next provider is tried. Strong models plan, fast models write reports, when your keys allow it.

Attach documents, images and data. They are available to the agent and stay in the chat for good.

Replies can include cards: maps, lead tables, file cards, calculators and previews of Code, Pro and the computer.

While work runs you see a step list, a progress bar and the current action. Stop or skip at any time.

Each chat gets a private Linux desktop. Files and logins stay between sessions.

The agent works on a real screen you can watch live, not behind a hidden API.

A real browser window: click, type, select, upload, tabs, downloads. Cookie banners are dismissed automatically.

Run commands, scripts and data tools in a real shell.

Chrome, VS Code, LibreOffice, GIMP, Inkscape, ffmpeg, pandoc, OCR and a Python data stack are preinstalled.

Switch the panel from watch-only to your own mouse and keyboard any time, then hand it back.

The screen is recorded and a replay link is posted with the result.

Finished tasks are saved as label-based steps and replayed with no AI model. If a site changes, only the broken step is re-learned.

When work ends the computer pauses to save credits. Files and logins are kept.

Chat, Venus Code and Venus Pro use separate computers, so a video render never blocks a coding job.

Files the agent makes are shown as cards in chat and can be downloaded.

An agent can have its own email address for sign-ups and verification codes, kept apart from yours.

A coding agent that works inside its own computer as a background worker. It keeps going if you close the tab.

Big tasks start with a written plan so the work stays on target.

The agent must read a file before changing it, which prevents blind overwrites.

Errors are checked after each change and fixed before moving on.

On finish the project's tests and build are run. A failing build is not reported as done.

A second pass reviews the change for bugs, missing requirements, placeholders and secrets before it can finish.

Up to 3 writers work on separate parts in isolated copies and are merged back, while read-only explorers search in parallel.

Destructive commands are blocked and secret files are unreadable. Permission modes: plan, strict and auto.

Milestone-based fresh context with saved state, stuck detection and a compute budget.

HTML previews open at once and a live dev server runs for apps.

Use /deploy to publish a site from chat.

Pushes go to a new branch with a pull-request link. Tokens never touch repository config.

/clone owner/repo
...edit with the agent...
push all changes

Add your own checks in the project so every job follows your rules.

A video studio with a timeline editor, a live preview and a render pipeline on its own studio computer.

The agent writes the storyboard: scenes, palette, typography, camera moves, transitions and sound cues.

Scene count is chosen from duration, story complexity and pace, not fixed.

Every scene is custom animated design, checked and repaired when something is wrong.

One narrator voice per scene. Only the script narration is spoken, never the directions written around it.

An ambient music bed and optional sound effects timed to scene beats.

Stock media is downloaded to the studio computer so the render can use it.

Attach a video and ask for edits. The same video is edited, not remade from scratch.

Give a reference video and the agent follows its structure and pacing.

Contact sheets of frames are reviewed and fixes are applied before the final render.

Export at 720p or 1080p.

Split, trim and reorder scenes, change palette and fonts, or ask the in-editor agent to make a scene punchier.

Narration and on-screen text work in many languages, including Hindi in correct script.

Target length is honoured, and long documentary-style videos are built in chapters.

Videos per day depend on your plan. See Plans.

Ask for a deck in chat and open it in Deck Studio.

No fixed template: the agent designs each slide on a 16:9 canvas.

Full-screen presenting from the browser.

Download as HTML, PDF or PowerPoint.

Change slides by hand or ask the agent for edits.

Palette and fonts you chose before are reused, so decks match your brand.

The default chief of staff. It plans, picks specialists, tracks progress and reports in one message.

Say "assemble a team" to choose a roster. Nothing starts until you give a goal.

Strategy, research, sales, marketing, content, engineering, design, video, data, operations, finance and people.

Given a goal, Chief builds the smallest team that can do it, not a crowd.

Work is split into up to 7 steps with dependencies.

Independent steps run in parallel. Steps that share one computer run one after another.

Members map to the computer, Venus Code, Venus Pro, web research, writing or email drafts.

Open any member's thread and talk to them directly.

Watch each member's progress as it happens.

Failed or skipped steps are never reported as done. A verifier gives pass, partial or fail and fails closed.

One agent can read the briefing, memory and files of your other agents, when you allow it.

Memory belongs to one chat. Nothing leaks between chats.

Owner profile facts, preferences and lessons from finished work.

Name, business, tone and goals, so you do not repeat yourself.

Memory is not capped at a small number of facts.

Say /remember or /forget, or tell the agent in plain words.

Reusable procedures with a win and fail record.

Paste a skill link, a GitHub repo, an install command or plain text. Suspicious instruction-like skills are installed paused.

After a multi-step success the agent writes a trial skill.

A trial skill is promoted after 2 wins and paused after repeated failures.

On failure the agent explains the cause, writes a new version and keeps the old one for rollback.

Search earlier messages, missions and call summaries with ranked results.

FOUNT adds a short, proven how-to block for the job at hand, records what happened (opt-in, scrubbed) and promotes only patterns that really worked.

The agent's tool list is rebuilt on every message, and an "Agent updated" note explains what it can now do.

Google (Gmail, Sheets, Drive, Calendar), Microsoft 365, Slack, GitHub, Notion, Linear, HubSpot, Asana, Zoom, Figma, Webflow and Dropbox.

Search a service name or paste a remote MCP address.

Servers that need login use automatic OAuth sign-in.

Give a base URL and a key. Reads are free, writes ask for approval.

Say "connect my bot" and paste the token. The agent sets it up itself.

One tap checks that a connector really works.

Send many emails or write many files in one approved call, with success and failure reported per item.

Anything that changes data asks first, unless you set it to auto.

If a tool is missing, the agent finds a way: it connects it, builds the file itself or does it on the computer.

Connector tokens are never handed to the computer or shown to the model.

Anything with an API can be added, so the agent is never limited to a fixed list.

Create a bot, paste the token, and chat with your agent.

Chat with your agent on WhatsApp.

Plain text messages work too.

Message your agent directly in Slack.

Reports and approvals can arrive by email.

Only your own ID or number gets replies. Strangers are ignored.

Risky actions ask "reply YES to approve". You can approve from your phone.

Payments from chat need a PIN. After repeated wrong PINs payments lock for 30 minutes.

When a job ends, the result is posted on every connected channel.

Your agent can phone people and answer calls with its own voice. It talks like a team member, uses tools during the call and posts a summary in chat afterwards.

No settings page. Tell the agent what you want:

set up calling
buy a number
show my numbers

Calling works with the major phone providers and any compatible provider. You paste the account details once; setup works in every chat.

After connecting, the agent shows your account details, balance and limits in chat.

Use a number you own or ask the agent to buy one. Your balance is spent when a number is bought.

Give each agent its own number, or let all agents share the default.

Save your number so the agent can phone you when work is done, and so it knows when you are the one calling.

Use the free built-in voice or a premium voice with your own key.

The agent answers in the caller's language and follows if the caller switches. Hindi is spoken in correct script.

A call button sits beside attach in the composer. Tap it to talk to the agent by voice right from the chat.

1. 1Ask in chat: "call Rahul and confirm the meeting".
2. 2The agent states the goal and, by default, asks you to confirm.
3. 3The number is checked and given a country code.
4. 4The phone rings. While it rings, the agent's voice is made ready.
5. 5The agent introduces itself as an AI and works toward the goal.
6. 6A summary and follow-ups are posted in chat.

By default the agent asks before every call. You can turn this off.

When someone phones the agent's number, the agent greets them, finds out what they need and handles it.

The first sentence of every call makes clear the agent is an AI assistant.

The agent speaks 1-3 short sentences, lets the other person answer, and does not read lists or links aloud.

A call is treated as yours when the agent dials your saved number, or when it comes from your number and you chose to trust caller ID.

Callers you own can ask for mail, search, calendar or research and hear the result on the line. Anything that changes data is noted as a follow-up and done after the call.

Things the agent promised are collected, then posted in chat. Reply "do the follow-ups" to run them.

While a call runs you can type or say a note, for example "tell them the meeting is booked". The agent hears it at once and says it in its own words.

Hang up from chat any time. The agent can also end the call when the goal is done.

Each call has a transcript, a 3-5 point summary of outcome, key facts and open items, and its length.

In real-time mode the caller can interrupt mid-sentence with almost no delay. If it is not ready, the call falls back to the normal turn-based mode.

A daily outbound limit protects your balance. Default is 50 calls a day. Numbers are validated before dialling.

A mission can find leads, verify them, call each one in turn, update a sheet and report back.

Reception, appointment confirmation, payment reminders, lead qualification and follow-up calls. More complex conversations come with later updates.

• "Connect calling first": say "set up calling".
• "No calling number": say "buy a number".
• Number looks wrong: include the country code.
• Silence at the start: the voice may be waking up; try again in a moment.
• Daily limit reached: wait for the next day or raise the limit.

A mission is a long multi-step job, for example: find leads, verify, call each one, update a sheet, report.

Missions run server-side in short slices and keep notes as memory, so nothing is lost between slices.

You set the number of calls, whether writes need approval and a step cap.

The final report is posted on every connected channel.

A routine repeats a task on a schedule: every hour, day or week.

Cron schedules follow your time zone, including daylight saving changes.

Set hours when routines must not run.

Say "every weekday at 9" and the agent turns it into a schedule.

Telegram, WhatsApp, SMS, Slack, email or a webhook.

Start a routine from another system with a webhook.

Create, list, edit, pause, resume, run now or delete routines by asking the agent.

The scheduler ticks every 5 minutes, so nothing runs more often than that.

Failed runs are posted in chat with the reason.

Free maps and location tools for agents. No map key and no billing.

Find an address or a business by name.

Turn coordinates into an address.

Find categories of places near a point, such as clinics or cafes.

Distance and time for driving, walking and cycling.

Maps with markers and routes appear inside the answer.

Nearby results can be turned into a lead list.

Each lead gets its website, email, phone and a one-line summary of what the business does.

Download any lead list as a CSV.

Every outside call goes through the server. Results are cached, so repeat questions are instant and still work without internet.

Requests are rate-limited and identified so free map services stay available.

Point Vmaps at your own map servers and it runs 100% offline.

A request for N items becomes a contract that the system enforces.

About 1.6x the quota is collected from real pages, each with its source.

Website live and not parked, company found on the page, email format and mail-capable domain, duplicates removed.

The agent cannot finish below quota. Rejected items cannot be edited into passing.

The CSV of verified rows and the CSV of rejects are built by code, so the model cannot inflate them.

The summary starts with how many of N were verified. A short list is reported as short.

A second reviewer checks results for gaps, and fails closed.

Keys and tokens are encrypted and tied to your account. The browser only holds masked references.

Logins, 2FA codes, saved sessions and cards that agents can use but never see.

A password is only typed on the site it belongs to.

Cards need your approval for each payment and obey a per-payment limit.

Computers get a short-lived signed token, never your connector tokens.

Outside content is treated as untrusted data. If it tries to give orders, the run is marked and writes need a human.

Private and internal addresses are blocked, including inside the browser the agent uses.

Tool writes, vault changes, calls and missions are logged.

You decide which actions need a yes: send, write, pay, call.

Chat apps answer only you.

Callers who are not you never get your private data.

Plans and limits are decided on the server, so a modified copy cannot unlock them.

Report security issues privately to the team instead of publishing them.

A desktop app that shows the real website, so plans and limits work exactly as on the web.

Run the agent's Linux computer on your own PC instead of the cloud.

Sign in once in your browser; the desktop app receives access safely.

A small command-line tool starts and stops the local computer.

The same agents, memory and rules apply in cloud and local mode.

A native mobile app is on the roadmap.

Limits refresh every 5 hours, like a session limit rather than a fixed message count.

Every plan can use your own model keys.

Computer-agent time grows with the plan and becomes always-on on higher plans.

Venus Pro videos and Venus Code sites per day grow with the plan.

Higher plans see new features first.

Upgrade from inside the app. Limits change straight away.

The server decides your plan, never the browser.

Every run is listed with its steps and an estimated cost.

Open a run to see each action in order.

A fixed 40-task suite measures success rate over time.

Changes are tested before release.

Tell the agent what was wrong and it saves the lesson.

• Work that must be correct: verified output.
• Work on real websites with a visible computer.
• Long jobs that finish while you sleep.
• Build, create, outreach and calls in one product.
• Keeping secrets safe.

Older open-source agents have bigger communities, more integrations and more maturity. Some hosted bots are easier to set up.

You describe a goal, the agent plans and works, and you get one report with evidence and any items that need you.

• Beta software with rough edges.
• Needs your own AI key and a computer key for full power.
• Community and third-party skills are smaller than older projects.
• Email checks prove a domain can receive mail, not that a mailbox exists.
• Security features have tests but no independent audit yet.

Find leads, write personalised emails, send them in one approved batch and track replies.

A routine that sends yesterday's numbers and today's tasks every morning.

Call customers who have not paid, confirm appointments or collect feedback, then log results in a sheet.

Sort mail, draft answers and flag only what needs you.

Sourced comparison of tools, markets or competitors with links.

Each update removes another human step from the work cycle.

More verification, better recovery from failed steps and more repeatable recipes.

More connectors, a mobile app and a skill marketplace.

Faster, more natural calls and stronger real-time mode.

An agentic model tuned for coding and agent tasks.

After 2-3 updates, agents are intended to run most routine desk work end to end, with humans approving exceptions.

Roadmap items are goals, not promises. Dates may change.

No. They are stored encrypted. The browser only holds a masked reference.

Computer and code jobs continue and the report is posted when they finish.

Only with a card you added, and only after you approve each payment.

No. Memory belongs to one chat.

A vision-capable one for computer tasks. Any provider for writing and code.

It can take over many tasks of a role today, and more after upcoming updates. Humans stay responsible for decisions.

Tell the agent which tool, or open Connectors, search it or paste its address.

Yes. Say "call me when you are done" after setting up calling.

Say what was wrong. It records the lesson and fixes the skill.

Only the broken step is re-learned on the next run.

The voice may still be waking up. Retry in a moment.

Only verified rows are counted. The report says exactly how many passed.

Check it is not paused, quiet hours and its schedule.

Ask the agent first, then contact the team.