Contributing
Roadmap
What is being worked on, how items are prioritized, and what shipped recently.
The roadmap lives in roadmap/README.md, grouped by area (Kernels, Inference, Harness, Loading, Networking, Room, UI, Infra, Tests). Design notes sit next to it as numbered files, and the pinned issue #83 has the same list as a checklist.
RoadmapEvery item with its priority, design note and issue.
Pinned issue #83The same list as a checklist on GitHub.
Priorities
Section titled “Priorities”| Priority | Meaning |
|---|---|
| P0 | Now. Broken, wrong, or the biggest thing holding rooms back. |
| P1 | Next. Planned and ready to build. |
| P2 | Later, or needs a design note first. |
Issues get the area: label of their section.
Top of the list
Section titled “Top of the list”- SSD expert offload (P1, design note). Stream a MoE’s experts from NVMe when VRAM and RAM can’t hold them, the next model size up on a Mac and a PC.
- Batched multi-session decode (P1, #72). Agents with subagents send requests at the same time; today they queue behind each other.
- Host-authoritative room page (P1, #17). Only the host may load a model or start the room’s model; the room node already checks.
- Reproducible benchmarks (P1, #21). A protocol, one result format and a device matrix; a contributor is taking the first slice.
- Pinned model revisions (P1, #19), so an upstream re-upload can’t mix old and new ranges.
Shipped recently
Section titled “Shipped recently”As of October 11, 2026, every roadmap line was checked against the code:
@pooled/cli0.4:pooled servewith OpenAI and Anthropic endpoints and tool calling,pooled host/join/chaton a native peer, and the OpenClaw plugin (#4, #11).- Bigger models: the Qwen3.5 122B, and MoE experts kept in system RAM (
--ram), so a Mac Studio and an RTX 5070 PC run the 122B at about 17 tok/s. - Security: channel-bound join auth, where keys and passes are proved, never sent.
- Networks: the TURN relay, in the browser and now the CLI (#1).
- Speed and sessions: the prefill kernels (#2); checkpoints on disk and prefix reuse across agent turns (#69, #73); GPU sampling (#26); the GEMV 8-row fix (#47).
Pick something up
Section titled “Pick something up”- Find an item in the roadmap and read its design note, if it has one.
- Comment on its issue before you start, so two people don’t build the same thing.
- P2 items and anything that changes the protocol, the model output or what peers can see need a design note first. See Contributing.
Benchmark reports from hardware we don’t have always help. See Benchmarking.