AgentOS

If your sales lead works in Claude, your engineers live in Cursor and marketing uses ChatGPT, none of those tools know what the others did or what a client was promised on Tuesday's call. Everyone ends up re-explaining the same background to a different assistant.

AgentOS, from a small AI consulting company called Devcore, tries to fix that. You connect the places work already happens, such as Gmail, Slack, Google Drive and Google Calendar, plus GitHub and meeting recorders, and it builds a searchable knowledge base from them. You can ask questions inside AgentOS, or connect Claude, ChatGPT, Cursor and custom agents over MCP so they read the same context.

The useful detail is that answers point back to their sources. The site's example asks what a client was promised on pricing and shows the email and call behind the reply. One thing to know when you go looking: several unrelated products use the AgentOS name, and the company says so itself. This one lives at tryagentos.net.

A built-in manager agent called Otto keeps records current, chases follow-ups and messages people when something needs a decision. Skills are written procedures, such as a client prep brief, that every connected agent then follows the same way. Bots are agents that wake up on their own for recurring work.

You choose what each agent can reach, so Slack might be read and draft while a finance folder stays off limits. The company says consequential writes to business systems wait for human approval, and its privacy policy says email goes out only after you approve it. Slack is looser, since AgentOS can post replies in channels you grant. The terms also warn that AI output can be wrong and that you're responsible for how it's used.

You can start without paying. The free plan has 200 credits a month, 300 extra at signup and 100 workflow runs a month, with no card required. Starter is $20 a month for 1,000 credits, Pro is $99 for 6,000, and Team is $299 for three people sharing 20,000 credits, with each extra person at $79. Yearly billing gives two months free. Monthly credits don't roll over, and $10 buys 500 credits that never expire. Credits cover model usage and vary by model and task, so it's hard to know from the pricing page how far a plan will go.

The privacy policy, effective October 29, 2026, names its providers one by one, and it's worth reading because this product takes in a lot. Workspace content and connected-app data stay for as long as your workspace exists. To answer requests, AgentOS sends content mostly through OpenRouter to model providers such as OpenAI, Anthropic and Google, along with Moonshot AI, DeepSeek and Z.ai. Devcore says it doesn't train models on your content, though some providers keep request data briefly to monitor abuse.

Composio holds the access tokens for Gmail and Drive, and Recall.ai captures audio, video and transcripts when you send the recorder to a meeting. In the app, PostHog records session replays with typing and workspace text masked. Sign-up and payment events go to OpenAI Ads with a hashed email address. Deleting a workspace means writing to Devcore by email.

A team with client confidentiality duties should go through that list before connecting anything. If the list sits fine with you and your team's context is scattered across three assistants, the free plan is a low-stakes way to see what a shared memory feels like.