Slotstream adds local Codex and Claude Code support to its SSD-streamed Mac runtime

Slotstream 0.2.20 added the OpenAI Responses API that current Codex expects, including function tools, Codex namespace bundles, freeform custom tools such as apply_patch, streamed reasoning and image inputs. The project documents Codex through a custom local provider pointed at Slotstream's server, with a generated model catalogue supplying the real context window and the apply_patch declaration that Codex needs for file edits.

Version 0.2.21 broadened that interface to Anthropic Messages and added `slotstream launch` for Codex, Claude Code, Pi, OpenCode and Hermes. The command can start a local server, connect the selected coding agent for that run and keep shared prompt state on disk. Version 0.2.22, published later on 18 September and now the latest tagged release, focuses on context-selection and cached-conversation correctness rather than adding another agent protocol.

Slotstream is an MIT-licensed Swift runtime built around Qwen3.8-Flash-Next on Apple Silicon. Instead of requiring the full model to remain in unified memory, it keeps shared weights in memory and streams the mixture-of-experts records needed by each layer from SSD into a cache. The project targets Macs with 16 to 64 GB of memory, requires roughly 110 GB of free storage for the model, and uses MLX and Metal without a Python runtime on the inference path.

The agent launch path is designed to keep model traffic local. For Claude Code, the release says it disables cloud-provider switches, clears API-key paths and denies the hosted WebSearch tool; OpenCode is limited to the Slotstream provider for that launch, Hermes side tasks are pinned to the local model, and `codex cloud` is refused. Those controls are visible in the source and release notes, but Vakker Wire has not independently packet-captured the full set of clients, so the privacy boundary remains a maintainer-implemented mechanism rather than an external audit result.

The most consequential current fix is about reproducibility across turns. Slotstream says the arithmetic for a prompt could differ depending on whether tokens were read cold, split across prefill passes or resumed from state left by generation. In one recorded 1,430-token agent turn, the resulting logit difference was enough to invert two candidate characters in tool-call syntax and produce a malformed `file.edit` call.

Version 0.2.22 now resumes only from state aligned to one of the request's own prefill-pass boundaries and re-reads the remaining tokens. The project reports that a continued conversation then produces the same tokens and bit-identical prompt logits as a cold read, and adds a `prefix-exact-check` gate plus a weights-free aligned-resume check to CI. The 3.7% to 5.9% logit-spread figures and the measured follow-up timing come from Slotstream's own tests; they have not been independently reproduced here.

The current arm64 release archive is built by CI from a named source commit, published with a SHA-256 digest and accompanied by GitHub attestation instructions. Slotstream's own code is MIT-licensed, while the model weights and other external components retain their own terms. The repository also publishes the wire-protocol documentation, agent guides and test gates used for the release, making the integration surface inspectable rather than a closed local gateway.

There are practical limits. The project supports Apple Silicon rather than Windows, Linux or Intel Macs, and the Codex integration uses a custom provider because `codex --oss` performs Ollama-specific version checks that Slotstream does not satisfy. The maintainer verified the 0.2.20 Codex path with Codex 0.148.0, but latency, tool reliability, compaction behaviour and the later Claude Code path still need independent reproduction on representative 16, 32, 48 and 64 GB Macs before the project's performance or reliability claims can be generalised.