v0.4.8: MCP, gateways, storage, and runs you cannot lose
Capabilities
- MCP Servers: Register remote MCP servers once, tenant-wide, and attach them to agents by name. A catalogue of 375 hosted servers prefills the endpoint and OAuth setup, tools are indexed and searchable across every server, and an activity view shows which agents reached which server and what they called.
- LLM Gateways: A gateway in your own cluster that fronts every model provider. Agents get short-lived virtual keys instead of upstream credentials; you get per-agent and per-user spend, token and cache accounting, and budget rules with spend caps and rate limits.
- File systems: PVC-backed disks that outlive a pod. Mount one on several agents, browse it from the dashboard, grow it in place, and move files in and out with exo fs cp.
- Memory: Durable versioned context an agent reads at session start and writes back — now with a full editor, version history and rollback, tenant defaults, promotion into shared knowledge or a tool's notes, and a map of what is attached to what.
Agents
- Schedules: Wake an agent on cron, a plain-English phrase ("weekdays at 07:30"), or once at a fixed time. The operator holds the clock, the control plane authorizes each fire and fails closed, and every decision is audited.
- Runs with a URL: A run has its own page that streams the real terminal, replays what it missed on reconnect, survives a refresh, and can be cancelled for real. Leaving the page minimizes it into a dock tab that flags when the run finishes.
- Terminal dock and beast mode: Shells keep streaming when you navigate away and reattach after a refresh. With several open, beast mode tiles them across the screen or stacks them behind a tab strip.
- Apps and Web Agents: An App pairs a dev server, a coding agent, and an optional Postgres with a live preview. A Web Agent is a browser in your cluster, streamed to the dashboard and exposed to agents as a Playwright MCP endpoint.
- Gemini and Pi runtimes: Two more first-class agent runtimes, alongside per-runtime image pinning in Settings.
- More providers: Google Gemini, OpenRouter, OpenCode Zen, and OpenAI-compatible endpoints join Anthropic and OpenAI as model providers.
- Webhooks: Inbound triggers and outbound notifications are now configured separately, with rate limits, in-place credential rotation, and delivery through the operator for private endpoints.
Governance
- Runtime policies: One YAML policy language enforced at the control plane, at the connector, and inside the agent runtime. Nineteen built-in packs, generation from a description or a pasted guardrail document, monitor mode before you enforce, global and label-scoped policies, and a decisions feed.
- Approvals: A require_approval rule holds the command instead of failing it. Approve or deny from the dashboard, or from the terminal via a QR deep link while the command waits.
- Usage and quotas: Live credits, disk, and node consumption against your entitlements, on its own page and in the sidebar.
- Per-resource access: Agents, apps, web agents, memories, providers, gateways, and MCP servers all share the same model: open by default, restricted to listed users, teams, or roles once you add the first entry.
Getting started
- Personal workspaces: Signing up as an individual provisions a managed workspace and hides the infrastructure: one Integrations page, auto-attached MCP servers, and a home page that is just your agents. Upgrading to a team workspace is one click.