Skip to content

Changelog

Built from the git history and the three tags in the repository. Pull request numbers link to GitHub. Full release notes are on the releases page.

Unreleased (on main since v0.3.0)

Server fixes (2026-10-09)

  • Think blocks are filtered out of the live stream, and the tool-call retry runs on a fresh KV cache so it no longer lands on top of the first attempt (#71).
  • stop_sequences and tool_result is_error are honored; tool parsing is skipped for requests without tools (#70).
  • Errors use Anthropic's error shape; a bad request body gets a 400 instead of a dropped connection (#69).
  • Arguments sent as a JSON string are decoded; Gemma 4 arguments are parsed as JSON (#68).
  • Images in tool_result are no longer pasted into the prompt as base64 (#67).
  • A JSON array inside <tool_call> tags is handled (#66).
  • Gemma 4 number and boolean arguments are kept (#65).
  • Parallel calls to the same tool with different arguments are kept (#64).

Code mode and parser (2026-09-14 to 09-19)

  • Code mode keeps the Skill tool so slash commands work, and fenced JSON arrays of tool calls (Hermes 4 14B) are parsed, with a standalone parser test, by @zwolf25 (#56).

Features and models

  • keep going: resume the current folder's Claude Code conversation on a free OpenRouter model or a local one, plus an optional status line reminder (2026-09-17).
  • Model lineup refresh: Qwen 3.8 27B (8-bit) for 96 GB Macs and Gemma 4 for 32 and 64 GB, older models retired (2026-09-19).
  • bin/ helpers added (fia, keepgoing) (2026-09-07).
  • CONTRIBUTING.md and an "other ways to do this" section in the README (2026-09-19).

Docs

  • README rewritten around a one-line summary and accurate setup and mode caveats (#61), plus fixes to links and the lsof privacy check (#60, #62).

v0.3.0 (2026-08-22)

  • Native Engine (agent/agent.py): an in-process MLX terminal agent with a fixed short prompt and a KV-cache fix for Gemma's rotating cache; 0.36 s turn starts on long sessions as measured in the repo (#51), with bench/agent_bench.py to reproduce.
  • 16 GB Macs get Hermes 4 14B instead of a 4B model; one-line install.sh added (#49).
  • Real terminal demo GIF in the README (#50).
  • Qwen 3.8 27B and Muse-Glimmer 30B documented; Agent-12 leaderboard linked from the README.
  • MLX_KV_BITS=4 tool-argument degradation documented (#42).

v0.2.0 (2026-08-05)

  • Prompt cache correctness (#46, #47): the cache was trimmed using .step (a fixed allocation size) instead of .offset, so the trim never ran and later sessions were answered from the previous session's cache. The release notes record 1/8 correct before and 8/8 after on one long-lived server. An empty reply after a cache hit now drops the cache.
  • Launchers find the claude binary instead of hardcoding ~/.local/bin/claude (#40, #48).
  • Gemma 4's collapsed <tool_call:Name{...} opener is parsed (#45).
  • Prefill chunk lowered from 8192 to 512 tokens to cut peak memory (#44).
  • Bare top-level JSON tool calls are parsed (Qwen 2.5 Coder), by @KaoCSC (#43).
  • Generation stops on the tokenizer's real end-of-sequence token, and empty numeric environment values fall back to defaults, by @KaoCSC (#41).
  • Live token streaming for requests without tools (#39).
  • Gemma 4 thinking suppression and slimmer tool descriptions in code mode, by @kevbarns (#33).
  • MIT LICENSE added.

v0.1.0 (2026-05-08)

First tagged release, mostly from @tadrianonet's #32:

  • Offline launchers: the four CLAUDE_CODE_DISABLE_* / DISABLE_AUTOUPDATER variables stop Claude Code's own startup traffic to Anthropic.
  • 16 GB Mac support with Claude Chat.command and Claude Agentico.command (Qwen 2.5 Coder 14B, 4-bit KV cache, login workaround).
  • Server-sent events for tool responses, which Claude Code 2.1 requires.
  • ChatML and Llama 3 stop markers, the Qwen 2.5 <tools> format and fenced JSON tool calls.
  • Tool inputs filtered against each tool's schema; the empty text block that made Claude Code drop tool calls removed.
  • uninstall.sh by @tripathiprateek (#23).

Earlier, untagged work (March to May 2026):

  • 2026-03-25: initial release, built on llama.cpp ("TurboQuant") behind a proxy.
  • 2026-03-26: the MLX server that speaks the Anthropic API directly replaces the proxy.
  • 2026-03-29: tool-call reliability work: garbled JSON recovery, retries, lower temperature.
  • 2026-04-11: launchers share a library with model-aware restarts and local model folders; the server is installed as a symlink to the repo; Gemma KV quantization crash fixed by @asdmoment (#7); client disconnects handled by @0xshugo (#4).
  • 2026-04-13: ArraysCache crash fixed by @kulveersingh (#10).
  • 2026-04-23: scripts/doctor.sh added.