Roadmap
What is built, what is in progress and what is planned. Milestones M1 to M4 are built and used; M5 is in progress; M6 onward is planned and may change. This page is a summary: each built area has a reference page, and planned work has no code yet. Last checked against the code at master 436b0c8 (see Architecture for what exists today).
| Milestone | State | What it covers |
|---|---|---|
| M1 Skeleton and chat | Built | Accounts (invite, passkeys, magic links), the model gateway with Anthropic and OpenAI-compatible adapters, routing, streaming chat, the usage ledger, the web app |
| M2 External API and routing | Built | /v1/chat/completions and /v1/messages with API keys, routing policies and budgets with an Admin editor, sandboxed artifacts on a separate origin |
| M3 Agent runtime | Built | Durable runs with steps and checkpoints, resume after restart, tool approvals, conversation compaction, River jobs |
| M4 Coding sandbox | Built | Per-user Docker sandboxes, egress proxy, GitHub App with git tools and open_pr, a files, terminal and preview panel |
| M5 Claude Code and MCP | In progress | Built: secrets from Infisical, an MCP client, the wsj launcher for Claude Code sessions in tmux, and the task router (rules plus a small-model classifier, off by default), and the outbound gateway (a byte-for-byte /v1/messages passthrough for Anthropic endpoints so Claude Code can use ws as its API gateway), the read-only job tab (a page listing the Claude Code sessions wsj reports, with a read-only terminal view, tailnet-gated), and a weekly launch counter that lets the router keep low-value work off the subscription as the week fills. Next: acceptance runs with a real Claude Code client and a real target |
| M6 Design, images and video | In progress | Built: a media gateway with hosted image generation (OpenAI), a generate_image chat tool and an Images page. Next: video, image edit, local engines such as ComfyUI, and design artifacts |
| M7 Long-lived agents | In progress | Built: agents with goals, cron and webhook triggers, a monitor page, and memory with reflection and recall. Next: repo-push triggers and an agent-facing remember and recall |
| M8 Fleet | Planned | vLLM and Ollama/MLX profiles, throughput-aware routing |
| M9 Rented GPUs | Planned | On-demand GPUs from RunPod, Lambda and Vast.ai with spend caps |
| M10 Training flywheel | Planned, optional | Dataset export, fine-tuning, an adapter registry |
Claude Code and subscriptions
Section titled “Claude Code and subscriptions”ws does not forward a Claude subscription credential and does not read Claude Code’s output. Programmatic Claude access goes through an API key and the gateway. A subscription session is launched by the separate wsj CLI, which starts the official claude in tmux and returns a handle. See Trust and security.
How to follow along
Section titled “How to follow along”Each milestone’s status is on this page and in the repository README. Planned items are intentions, not commitments, and they do not carry dates.