What changed in Player One

Platform and admin updates from 19 September to 3 October 2026, newest first. See the latest updates.

Release v0.3.286

Backups you can move, Slack replies as the agent, and one task at several times

Admin has full backups: download one, or move your whole installation to a new one with a code. A scheduled task can run at several times, each rule with its own days and times. Slack replies come from the agent under its own name and face, and Player One does less work when it starts.

Featured update
Back up everything, or move to a new installation
Admin > Backups prepares a backup of the whole installation, or of one project, some agents or chosen chats. Nothing that signs in on your behalf travels unless you tick Include secrets, which needs a passphrase. To move instead, get a move code on the old installation and paste it into the new one.
New
Download a backup in parts, or as one file
The size is measured in the background and each part appears as it is written. Save the parts to a folder or as one file. Show what and why lists anything left out, and the files that look like they hold a password or key.
New
Move to another Player One by code
The code works for one hour and is claimed by the first installation that uses it. The parts travel directly between the two installations, so you can close the page.
New
A restore that waits for Go live
A full restore brings everything in held. No hosted app starts and no secret provider runs a program until you select Go live.
New
One task, several times
Mondays at two and four, Thursdays at one and eight is now one scheduled task with two rules. A Times tab in the editor sets each rule's days and times, and agents write one job with rules rather than one per time.
Improvement
Slack replies come from the agent
A reply is posted under the agent's own name and face, as one message edited from its preview. Buttons arrive with their text above them, and a message to a person goes to their direct message.
Improvement
Approvals where you are
When you message an agent from one of your own channels and it needs your say-so for a command, it asks you there, in that chat, with Once, Always and No.
New
Voice notes transcribe from the start
Player One Provider, included with Player One, transcribes voice messages with nothing to connect. Voice notes also transcribe on xAI and ElevenLabs.
New
GPT Transcribe, Claude Sonnet 5.5 and Eleven v4
Voice notes and live calls on OpenAI use GPT Transcribe. Claude Sonnet 5.5 succeeds Sonnet 5 on Anthropic's own API and is available on six providers, and ElevenLabs Eleven v4 and v4 Turbo join the speech models.
Improvement
Models say how you pay
Every model and account choice says whether it is an API key, a subscription or included.
Improvement
Less work at start-up
An upgrade copies only the skills that changed, and the Slack and MCP clients and the backup code load only when they are needed.
Improvement
Staying signed in while you work
A browser tab you are using keeps its sign-in alive beyond a day.
Improvement
A Skills tab that names missing skills
An agent's Skills tab shows the skills it is set up with that exist nowhere.
New
Gmail filters and vacation replies
Google Workspace has opt-in Gmail settings for filters and the vacation reply.
Improvement
Buttons answer the right question
In Slack and Discord, the buttons on a question or an approval answer the one they name, in your direct message.
Release v0.3.257

Share agents between installations, connect your own tools, and a catalogue that grades every model

Agents can now be shared with other Player One installations, by a pairing code or by an open link. Coding tools such as Claude Code and Codex can connect to your installation and work with your agents. Every model is graded for power, cost and speed, and many more providers join the catalogue.

Featured update
Share an agent with another installation
You publish an agent and send the other person a link. They ask for access, their page shows a code, and you enter it on your Agent Sharing page. The agent keeps running on your installation, with your model, tools, files and memory, and they see only what it sends them.
New
Open a sandboxed agent to anyone
A sandboxed agent can be opened to anyone with its address, within limits you set. A visitor's turn gets the read and search tools and nothing that schedules, notifies, writes or speaks. Nobody asks you first, so open one only for an agent you are happy to have anyone talk to.
New
Team rooms across installations
Invite another installation into a chat you host, and join one hosted elsewhere. A shared agent is marked wherever you meet one.
New
Everyone who talks to your agents is a contact
Contacts now include everyone who speaks to an agent here, with their picture, and you can merge two into one. Your own card holds your name, picture, time zone and the notes agents read about you.
New
Connect Claude Code, Codex or another tool
On the MCP page you make a key and paste one prompt into your coding tool. It can then ask what decisions are waiting for you, start a chat with one of your agents, read what they are doing, and take up work they leave for it.
New
A catalogue that grades every model
Every model is graded for power, cost and speed from 0 to 10, and a rule can ask for a fast or cheap model by need rather than by name. The catalogue is signed and refreshes in the background, and news of new models appears on your home page.
New
Many more providers
Venice, Together, Chutes, DeepSeek, Mistral, Xiaomi MiMo, Alibaba Model Studio, NVIDIA, OpenRouter, Anthropic on Vertex, Vercel AI Gateway, OpenCode Zen and Go, Hugging Face, GitHub Copilot, MiniMax, Baidu Qianfan, Volcengine Ark and Synthetic join, each saying how it bills and what happens past its plan.
New
Grok 4.7, GPT-6 Sol and Luna, and Claude Opus 5.5
Grok 4.7 joins on xAI, GitHub Copilot and OpenRouter. GPT-6 Sol and GPT-6 Luna join on OpenAI, OpenAI Codex and others, and Claude Opus 5.5 on Anthropic and others. A model that a better one at the same price or cheaper replaces moves to its successor.
New
Permissions you set for each agent
An agent's Advanced > Permissions tab sets what it may run, with one control, one sentence and one list. An agent that needs more asks you on a card that asks one small question, and no agent gives itself more.
New
Use your own browser
With Sidekick's browser extension, an agent can work in the browser you are sitting in front of. The Sidekick panel says which chat it is on and follows the tab.
New
Fast voice calls on OpenAI Realtime
When an agent's switch is on, a fast voice call runs on OpenAI Realtime and both sides are saved in the chat. Live calls on OpenAI use GPT Realtime 2.1.
New
Choose how messages between agents show
A chat can draw messages between agents in full, folded into one line per run, or as a small marker. Your card holds the choice, and any agent can change it when you ask.
Improvement
Clearer model pages in Admin
Voice Models, Judgement Models and Embedding Models have their own pages, and Image Models has two tabs, Generation and Vision.
Improvement
A broken scheduled task tells its owner
When a scheduled task breaks, the person who owns it is told.
Improvement
Searches rank what they find
When an agent searches your memory, files and chats, the results are ranked, say where each one came from and say how to read them.
Improvement
A limit on always-on instructions
An agent's always-on instructions are limited to 25,000 characters. A save over the limit is refused, with what to cut.
Improvement
Smaller interface changes
The Usage page opens on Tokens, the notifications panel groups its permissions under their own header, and your card's time zone picker searches by city, region or offset.
Improvement
See and resume paused Memory Spaces
You can see each paused Memory Space and resume it.
Release v0.3.255

Telephone calls, Player One on your own computer, and cheaper long conversations

Player One can now answer and place telephone calls, and work directly on your own computer: your screen, your files, and a side panel with its own chat. On Anthropic models the tools sent with every request sit in one stable cached block, so a long conversation costs less and starts faster. The sidebar groups the chats each agent started, with a map of everyone working for you.

Complete fix log
Every customer-facing fix, fully listed
282 fixes in this release are published in one permanent record.
Read all 282 fixes
Improvement
Long conversations on Anthropic models cost less
The tools sent with every request now sit in one stable cached block, so a long chat is cheaper and quicker to start, and picking up a new tool part-way through no longer rebuilds it.
New
Player One works on your own computer
Sidekick lets Player One work on the machine in front of you: your files, your screen, and a shell that stays where you left it.
New
Player One answers and places telephone calls
A call runs as an ordinary conversation, kept in the chat, with a live banner while it is in progress.
New
A side panel on your desktop
The panel carries its own chat, your chats from every project in folders, and the full desktop composer.
New
A map of the agents working for you
Open a map from the row of faces to see every helper under the chat you are in, with a turning ring on any that is still working.
New
See your screen, and act on it
Player One can capture your displays and windows, then click, type and drag where you ask it to.
New
An agent speaks on a call instead of typing
Only the part meant to be spoken is read aloud, and the rest stays in the chat.
New
Your agents' chats, grouped together
The sidebar gathers the chats an agent started beneath it, as a tree you can open, and filters by what a chat is.
Improvement
A clearer view of what Player One is doing
A live progress line replaces the old activity boxes, and opens into a log of what is happening.
New
Chats are included in backups
Chats have their own tab and a chat history box, so a conversation can be exported and brought back.
New
Microsoft 365 mailboxes
Connect a Microsoft account with proper sign-in, and it stays connected.
Improvement
Approvals you can answer once
A command approval now carries an Always allow answer and gives you five minutes to respond, and Always allow never hands over an interpreter.