tag (hooks.server.ts), so this script<br>leaves them untouched. Every other route is prerendered as<br>`light`; here we upgrade to the visitor's stored preference<br>before first paint to avoid a flash. Kept inline + blocking<br>and dependency-free on purpose. Storage key must match<br>THEME_STORAGE_KEY in src/lib/config/theme.ts.<br>-->

content entirely when scripting is on —<br>so the scripted path is never affected. They hide the JS-only<br>[M]/[X] triggers, expose the + checkbox fallback, and key<br>the overlay's visibility off the checkbox's :checked state. The `~`<br>selectors rely on #nojs-nav-toggle, .header and the .overlay-*<br>surfaces all being siblings inside .header-wrapper.<br>-->

Centaur 2.0: Paradigm's Open Source AI Agent, Rebuilt

[M]<br>[M] [X] Menu

Centaur 2.0: Permissions, Context, and MCP<br>08.10.2026<br>By Matthew Slipper, Georgios Konstantopoulos<br>[L] Listen [S] Share Link, Markdown, X, LinkedIn

AI models change quickly, but a company’s tools, data, permissions, and shared history shouldn’t have to change with them. To address that, today we're announcing Centaur 2.0, giving every agent the same permissioned access to a company's tools and context, wherever people choose to work.<br>Centaur 2.0 introduces:<br>Company-wide context archival, so that your Centaur can quickly search across all configured data sources including Gmail, Granola, Slack and more.

Granular permissions, so that you can configure who has access to what resources, allowing you to give Centaur access to sensitive data without worrying about unauthorized access from its users.

MCP, so that you can query Centaur from your local agent, whether it is Amp, Claude Code, Codex, or your custom harness.

Two months ago, we open-sourced Centaur, our self-hosted runtime for secure, multiplayer agents. We believed agents would become shared infrastructure: present where decisions happen, equipped with a team's tools, and able to build context across an organization.<br>That thesis is playing out. More than 80% of Centaur sessions now happen in shared channels rather than DMs, and questions that once meant digging through years of messages, notes, emails, and databases get answered in minutes.<br>Putting an agent in a shared channel was the easy part. But making it useful inside a real organization requires access to sensitive systems, an understanding of company context, and the ability to carry those capabilities across models and interfaces. 2.0 ships the permissions and MCP support today, on a smaller, more reliable Rust core.<br>Permissions that follow the conversation<br>The previous version of Centaur assumed one organization-wide set of credentials. Companies don't work that way. Teams have shared systems, and access changes depending on where a conversation happens.<br>Centaur 2.0 derives access from the context in which it's invoked. A DM can use a person's connected accounts; a shared channel gets only the credentials and data sources granted to that team. Each sandbox receives its own identity, and grants are enforced at the network boundary without ever exposing credentials to the agent.<br>You can now connect systems that would be unsafe to place behind a single organization-wide credential. We use this ourselves: in a DM, Centaur can query your personal Granola notes and email; tagged into a channel, the same agent switches to the documents shared with that team. An engineer never sees a partner's meeting notes. The same rules apply to workflows, which receive only the access appropriate to their context.<br>Use Centaur’s tools anywhere<br>Switching models or clients shouldn't mean leaving your tools, permissions, and context behind.<br>MCP support ships today. You can use Centaur's tools and context from Claude, ChatGPT, or any other MCP client, under the same identities and grants used inside Centaur. We can ask Claude about a portfolio company, and it will pull the same meeting notes and documents Centaur would surface in Slack.<br>MCP support also lays the groundwork for cross-client handoff, which should ship in the coming weeks. Start a task in Claude, send it to Centaur, then close your laptop while the work continues. Long-running tasks will no longer depend on one device or client, and you can resume them without losing context or progress.<br>The same principle applies to Centaur itself. Claude Code, Amp, Codex, and nanocodex all run behind one harness interface. Slack remains Centaur's primary multiplayer interface, but teams have also contributed integrations for Linear, GitHub, Discord, and Microsoft Teams. We're also building a richer web console for persistent workspaces, files, traces, approvals, and long-running executions that don't fit inside a chat thread.<br>These are different ways to inspect and steer the same underlying work, not separate products with separate context and permission models.<br>Organizational context that compounds<br>The next step is helping agents understand not only what is stored in company systems, but how it all fits together.<br>We've shipped...