Roadmap
What is built, what is being built, and what is deliberately not planned.
Status as of 6 October 2026. This page is maintained alongside the design and is the single answer to "is X available yet".
Phase 1 — Foundation
Goal: one tier, end to end, with correct billing.
| Item | Status |
|---|---|
| Control plane: gateway, identity, admission, session-mgr | In progress |
Data plane: wbnode, session gateway | In progress |
| Tier 1 — gVisor sandboxes | In progress |
| Sessions, runs, exec, streaming | In progress |
| Filesystem API | In progress |
| Metering pipeline with reconciliation | In progress |
| Usage API | In progress |
One region: bom1 | In progress |
| Customer console: overview, sessions, usage | In progress |
| OpenAPI spec and generated reference | Done |
Exit criteria: a tier 1 session starts in under 250 ms p95, runs code, is metered to the second, and appears on an invoice that reconciles exactly against the node write-ahead log.
Phase 2 — Tiers and isolation
Goal: the full tier ladder with real isolation boundaries.
| Item | Status |
|---|---|
| Tier 0 — tool executor, built-in handlers | Planned |
| Tier 2 — microVMs with snapshot restore | Planned |
Warm pools and pool-mgr | Planned |
placer with filter-then-score placement | Planned |
| Egress proxy, default-deny, allowlists | Planned |
| Policy engine: egress, CLI allowlist | Planned |
| Volumes and snapshots | Planned |
Exposed ports and edge-router | Planned |
| Secrets vault and injection | Planned |
Templates and image-svc | Planned |
Exit criteria: tier 2 starts in under 600 ms p95 warm, pool hit ratio above 97%, and a session with
egress: none has demonstrably no network path out.
Phase 3 — Browser, desktop, billing depth
| Item | Status |
|---|---|
| Tier 3 — headless browser, action API, profiles | Planned |
| Tier 4 — desktop VMs, warm pool, input and capture | Planned |
| Human handoff for tier 4 | Planned |
| MCP server registration | Planned |
| Tool allowlist and confirmations | Planned |
| End-user attribution, quotas, caps | Planned |
| Invoices, credits, spend caps | Planned |
| Webhooks | Planned |
| Infra console | Planned |
Second region: iad1 | Planned |
Exit criteria: a reseller can meter, cap and rebill per end user from the API alone, and a tier 4 session starts in under 6 s p95 warm.
Phase 4 — Enterprise and scale
| Item | Status |
|---|---|
fra1, sfo1, sin1 | Planned |
| Residency enforcement across all four layers | Planned |
| SSO: SAML, SCIM | Planned |
| Private link to customer VPCs | Planned |
| Customer-managed encryption keys | Planned |
| OpenTelemetry export | Planned |
| SOC 2 Type II, ISO 27001 | Planned |
| Committed-use agreements and reserved pool depth | Planned |
| GPU-attached sessions at tiers 2 and 4 | Planned |
| SDKs: TypeScript, Python, Go | Planned |
Under consideration
Not committed. Listed so the shape of the argument is visible.
| Item | The case for | The case against |
|---|---|---|
| Tier 2.5 — shell without a microVM | Many tier 2 workloads never need a kernel; gVisor plus a shell would be cheaper and faster | Another boundary to defend, another tier to explain. The ladder's value is that it is short |
| Session migration between hosts | Survive maintenance without failing sessions | Moving live memory is slow and a residency hazard. Snapshot-restore already covers the use case |
| Multi-attach volumes | Several sessions sharing a workspace | Needs a cluster filesystem. A different product |
| Agent framework integrations | Lower integration cost | The API is deliberately below the framework line. Integrations belong in the frameworks |
| Model inference | One vendor for the whole loop | Not our competence, and it would make the platform a competitor to the customers' model providers |
| Windows at tier 2 | Some CLI tooling is Windows-only | Licensing and image size. Tier 4 already offers Windows desktops |
| Per-second billing granularity below one second | Marginally fairer for very short runs | Nothing meaningful changes; the complexity is real |
Deliberately not planned
Stated so nobody builds a proposal around them.
| Why | |
|---|---|
| Long-lived VMs as a product | Not what the platform is. There are good VM vendors |
| Implicit tier escalation | Would move a workload across a security and price boundary without anyone deciding to |
| A managed agent runtime | The orchestration layer belongs to the customer |
| Shared sessions between organizations | No. The isolation model is the product |
| Billing end users directly | What a customer charges its customers is its pricing decision |
| Reading customer session content for product improvement | Content is the customer's. Break-glass with two-person approval is the only path |
How this page is maintained
Each phase has exit criteria that are measurable, not aspirational. A phase is not complete because the items shipped; it is complete when the criteria are met in production.
Changes to scope, and the reasons, go in the decision log.