Zyphra MAIA: a free AI agent for teams in Slack

Felix Laurent 6 Nov 2026 4 min read

AI agents

Zyphra released MAIA in beta on November 6, 2026. The San Francisco company calls it a proactive, multiplayer AI agent for teams. MAIA works in Slack and in a Mac desktop app, remembers what a team decides and can act across its tools. It is free during the beta.

MAIA is the top layer of what Zyphra calls a full stack. Below it sit the company's own open models and its cloud on AMD chips. The launch shows Zyphra moving from research into a product that competes for space in Slack.

What MAIA does for teams

Most AI assistants work for one person in one chat. MAIA is built around a team instead. Everyone talks to the same agent, which keeps shared context and shared memory. Zyphra says it learns people, decisions, preferences and the reasons behind them.

The product page lists four main abilities:

Company memory. Ask about a decision from months ago and MAIA returns the outcome, the reasoning and who was involved, with sources.
Proactive work. It prepares morning briefings, preps meetings, flags conflicts and runs recurring tasks on a schedule.
Actions across tools. "Set up the client kickoff" makes MAIA book the time, draft the invite and share the notes afterward.
Dynamic apps. MAIA builds small apps from a team's real tasks, such as a dashboard, a kanban board or a sprint brief. The team can open, revise and build on them.

In Slack, users mention @MAIA in a channel or thread or send it a direct message. In channels, MAIA first shows its reply privately, and a person approves it before it is posted. Messages it sends for a user carry a "Sent from MAIA" label. The Slack app and the desktop app share one context, so a task started in one place continues in the other.

That reach into team tools raises the obvious question of access.

Connectors, permissions and data

MAIA connects to Slack, Gmail, Google Calendar, Docs, Sheets and Slides, plus GitHub, Jira and Confluence. Zyphra says more connectors are coming. The agent sees only what a user can already see in each tool. In Slack it reads public channels and channels it has been invited to.

Zyphra built several safeguards into the beta:

MAIA confirms before irreversible actions, and users control the approval gates;
private chats with MAIA stay private, and shared objects are shared by choice;
users can view, add and delete anything in MAIA's memory;
disconnecting a tool removes MAIA's access to it at once.

Zyphra says it does not use customer data to train general-purpose models. AI model providers that process data for MAIA are bound by the same rule, according to the FAQ. That wording suggests MAIA uses outside models at least in part. Zyphra has not said which models run the agent.

The beta itself is narrow. Zyphra is looking for small teams that use Macs, Slack and Google Workspace. Users fill out a form and wait for their company domain to be approved. Free beta access comes with daily usage caps. MAIA in Slack's AI assistant panel requires a paid Slack plan.

The company behind MAIA has spent years on models and chips before this launch.

Who is Zyphra

The company was founded by CEO Krithik Puthalath, with Beren Millidge as chief scientist. It is known for small, efficient open models. The Zamba and Zamba2 models came out in 2024, and the Zonos text-to-speech model followed in February 2025. In June 2025, a Series A of more than $100 million valued the company at $1 billion.

Zyphra trains its models on AMD hardware. In October 2025, IBM and AMD announced a large AMD Instinct MI300X cluster for Zyphra on IBM Cloud. The announcement said the cluster would train multimodal models for Maia, then described as a general-purpose superagent. In May 2026 Zyphra released ZAYA1-8B, an open reasoning model with about 760 million active parameters, trained on AMD.

The same month Zyphra launched Zyphra Cloud, an inference service on AMD MI355X GPUs. Forbes reported that Zyphra was raising about $500 million at a valuation of at least $5 billion. The round has not been confirmed as closed.

MAIA is not new as an idea. Zyphra opened a waitlist for "MaiaOS" in December 2024. The beta is its first public release.

MAIA vs Slackbot, Claude Tag and ChatGPT agents

Zyphra enters a crowded market. Several large companies already sell AI agents that live in Slack and team tools:

Agent	Company	Status	Where it works
MAIA	Zyphra	Free beta since November 6, 2026	Slack, Mac app
Slackbot	Salesforce	Generally available since January 2026, Business+ and Enterprise+	Slack
Claude Tag	Anthropic	Research preview since June 2026, Team and Enterprise	Slack
Workspace agents	OpenAI	Research preview since April 2026	ChatGPT, Slack
Custom Agents	Notion	Generally available since May 2026	Notion

Dust uses the same "multiplayer" pitch and raised $40 million in May 2026. Glean sells enterprise agents built on company search.

MAIA's angle is the shared memory and the apps it builds for the whole team. Most rivals also offer memory, connectors and approvals. Zyphra's real difference sits lower in the stack: open models and AMD infrastructure. That only matters if MAIA runs on them, and Zyphra has not said so yet.

Is MAIA worth trying

For a small team on Mac, Slack and Google Workspace, it costs nothing to test during the beta. The private preview of replies in channels and the editable memory are sensible defaults for an agent that acts for a whole team. The dynamic apps could save teams from building trackers by hand.

The open questions are big ones. Zyphra has not named the models behind MAIA, published any benchmarks or announced pricing. MAIA also has to win space next to Slackbot, which Salesforce builds into Slack itself. Teams already paying for Claude or ChatGPT get similar agents in the same channels. MAIA is a reasonable experiment for early adopters, and a test of whether Zyphra can turn its research into a product teams pay for.