Skip to content

Releases

mlx-agent v0.13.0

Choose a tag to compare

@github-actions github-actions released this 16 Nov 17:17
v0.13.0
96995d8
  • Add serve front-door: one authenticated loopback port that lists discovered routes sharing its API key by path-free model name and routes chat, Messages, Responses and embeddings to them. The mlx-agent OpenClaw provider plugin discovers its chat models with the gateway key.
  • Add Splash as a serving runtime for local GGUF, MLX and Strata profile weights and cached repositories, with authenticated --jit serving behind the front door, strata and splash bench labels and the Splash family of each quantized MLX output in convert scan.
  • Serve classifiers, text embeddings, speech, music, image, video and 3D models through typed local endpoints that keep weights loaded between jobs or release them after each request, including OpenAI POST /v1/embeddings with float or base64 output. convert speak, music, generate, video and describe run on a live endpoint that already serves the same model, and task capabilities --model-path reports model-bound media settings.
  • Add the Kev classification port (jaredpalmer/kev-4b over Qwen3-4B-Base, 8-bit or unquantized), and serve TypeSafe System One decisions at POST /v1/systemone on classification task routes and the front door.
  • Protect MLX-LM, vision, Strata, accelerated chat and typed endpoints with API keys, separate control credentials and private Unix or HTTP workers. Add serve private-worker and a foreground authenticated MLX service for existing service managers. A live model load refuses when its files or settings no longer match the reviewed serving plan.
  • Share memory admission across resident endpoints and tracked agent processes under one budget that includes Metal physical memory; --memory-budget-gb caps task jobs. A guard stop releases only the model worker and the endpoint keeps admitting requests, exited owners are not charged, idle task allocator cache is reclaimed before a refused admission is retried, and serving status keeps numeric reasons for memory refusals.
  • Add named serving workflows with per-role models and resource profiles, read-only previews, owned activation and memory forecasts that include resident models and pending allocations.
  • Managed chat gateways apply serving defaults without overriding explicit client settings and report token usage and response throughput through protected status.
  • Convert to FP16 or FP32 and 2, 3, 4, 5, 6 or 8-bit where the converter supports them. Model update preparation fetches pinned sources with companion and support files, requires observed completion and fresh output evidence, and reuses completed downloads instead of starting overlapping ones.
  • convert scan reports the disk each model occupies with every file counted once, bytes under each root that belong to no model, families of copies of one model across MLX, GGUF and Ollama, vision projector GGUFs with their declared base model, and standalone .safetensors and .ckpt files; --compare-assets compares same-size files by full SHA-256.
  • Add the mlx-task command and skill to every provider for durable local inference tasks, widen mlx-advisor to the whole model lifecycle, and document agent integration and the MCP server.
  • TRELLIS.2 convert mesh includes the texture stage and writes a closed, outward-facing solid with base color as COLOR_0 and occlusion, roughness and metallic as COLOR_1, takes repeatable --view images and a --guide GLB, and loads only the components each resolution needs.
  • Add the optional ace-step backend: convert music runs ACE-Step 1.5 checkpoints with their planner, and takes --bpm, --time-signature and --reference for them.
  • Add opt-in adaptive denoiser caching for Wan2.1 T2V and TripoSG with feature-memory budgets, refresh limits and reuse statistics, and speed up TripoSG and MiniMax Music3 by reusing conditioning projections, depth prefixes and prompt encodings.
  • Fix MLX-VLM tool calls from templates that write <tool_call><function= without a newline, such as MiMo-V2.6 and Qwen3.5, size paged Strata MTP expert caches from packed draft-head storage, reap exited serving and worker processes, and update opentelemetry-api to 1.45.2 and the ACE-Step runtime locks.

mlx-agent-docs-v0.13.0.tar.gz

sha256 bebf498dac02b1033ccf12d397076f1e66fbd63dbfa3cc43394d4d92e6ebf683

mlx-agent v0.12.0

Choose a tag to compare

@github-actions github-actions released this 07 Nov 23:00
v0.12.0
fc585f3
  • Add Strata total-memory budgets, live budget changes and saved usage profiles. Choose plain, verified prompt-lookup or MTP decoding; unused draft heads stay unloaded.
  • Add image inputs to the Messages and Responses APIs and opt-in image-feature reuse for Qwen2.5-VL and Qwen3-VL dense/MoE models.
  • Improve M5 code-edit decoding and prompt processing, reduce idle worker CPU use, and protect active requests during memory-pressure handling and worker reloads.
  • Add a coding-edit benchmark and loopback endpoint measurements for repeated prefixes, prefill, concurrency and cold loads. Benchmark exports distinguish end-to-end throughput from native stage rates.
  • Add native MLX TripoSG, TRELLIS.2 and Pixal3D mesh generation, including octree surface extraction and camera-aligned glTF output.
  • Accept 5-bit and 6-bit conversion where supported, and report converter-specific bit widths and model sizes.
  • Rank discovery by downloads, trending, likes or recency, and search repository names.
  • Discover MLX Agent services across receipt locations and scope doctor, watch and stop commands to a selected service.
  • Fix JIT tool-call identifiers, stream keepalives, end-token trimming, KV-cache admission and readiness deadlines.
  • Add mlx-agent --version and refresh installation, serving and compatibility documentation.

mlx-agent-docs-v0.12.0.tar.gz

sha256 f764773cac549eb05e3ca0138cf6d833c2e2f5458820fd6cd69bcdb2c661a624

mlx-agent v0.11.1

Choose a tag to compare

@github-actions github-actions released this 02 Nov 23:52
v0.11.1
fb9b13b
  • Fix: convert start --repo converts a model downloaded with intake fetch --revision from its cached snapshot instead of downloading it again.

mlx-agent-docs-v0.11.1.tar.gz

sha256 b31c628bcc541d5110e129ef2c221193b76122444f77160e382b51a2089585a8

mlx-agent v0.11.0

Choose a tag to compare

@github-actions github-actions released this 30 Oct 14:49
v0.11.0
cb0aeb4
  • Add serve start --m5-prefill: an opt-in MLX-LM prompt kernel for affine 4-bit Qwen2 MLP projections on M5 GPUs with macOS 26; other hardware, dependencies, models and inputs keep native MLX, and the plan and receipt record prefill_kernel.
  • Add serve start --prompt-lookup: opt-in MLX-LM decoding for code edits that drafts up to eight tokens from repeated prompt text and verifies them with the model; plain greedy requests run one at a time, other requests keep native generation, and the plan and receipt record decode_profile.
  • JIT serve gateways answer the Anthropic Messages API (POST /v1/messages) and the OpenAI Responses API (POST /v1/responses), streamed and unstreamed, with text, system prompts and tool calls translated to the model worker's chat completions; unsupported content and parameters return an invalid_request_error.
  • The core requires Python 3.11 or later; the mlx-agent and mlx-agent-mcp launchers exit with a message on older interpreters, including the python3 bundled with macOS.
  • Releases also notify cavi-ai/plugins, which points every host in the plugin catalog at the release.
  • convert start --gguf converts Qwen3.5-family GGUFs (qwen35) through a bundled port to MLX-VLM's qwen3_5_text, or qwen3_5 with the matching mmproj projector from the same folder; GGUFs holding only a multi-token-prediction block are classified as speculative drafters and refused.
  • Update filelock to 4.2.4 and setuptools to 84.2.1 in backend locks.

mlx-agent-docs-v0.11.0.tar.gz

sha256 56bc78c3191ec71f68a90bddcf4a6330abed9852105c55c9b7b43916dbfdd943

mlx-agent v0.10.0

Choose a tag to compare

@github-actions github-actions released this 22 Oct 18:50
v0.10.0
6e6effd
  • Fix: credential redaction covers convert transcribe, convert generate and convert decide failure messages, the token after an Authorization scheme such as Bearer, prefixed Hugging Face, GitHub, OpenAI-style and Slack tokens, and text inside JSON values.
  • Fix: the coding role probe checks a model's answer in a separate isolated interpreter with a 5-second limit instead of inside mlx-agent.
  • Fix: capture volumes work while other disk images attach or detach. The mount check reads the hdiutil listing again instead of failing and leaving the image attached, and a removal preview no longer changes when the capture remounts on another device.
  • The mlx-agent and mlx-agent-mcp launchers cache compiled bytecode under $XDG_CACHE_HOME/mlx-agent/pycache (default ~/.cache/mlx-agent/pycache) instead of recompiling the package on every run, and each command loads only its own modules, so mlx-agent starts in less than half the time.
  • Inference timeouts report <verb>_timeout for every convert command and task kind; describe, speak, video and music reported <verb>_failed.
  • Music tasks reject truncated WAV output, and a changed music output reports output_artifact_changed like other media tasks.
  • Discovery inspects up to six listed models on Hugging Face at a time instead of one after another.
  • CI and release workflows pin every action to a commit and only the release and PyPI jobs receive write permissions; a Coverage / Python 3.12 job runs the suite with the experiment runtime installed, so the lab tests run in CI, and fails when line coverage of the package drops below 84%.
  • Configuration transactions record each file in an append-only step journal instead of rewriting the whole receipt, so provider installs no longer slow down as files accumulate; rollback and recovery replay the journal of an unfinished transaction. Provider artifacts are listed with one directory scan, receipt loads and target locks check each folder once instead of once per file, previews are built only up to the preview limit, and provider CLI versions are checked in parallel.
  • Fix: a provider CLI newer than its last tested version is detected as native-visible, with a provider_untested warning in doctor, instead of unsupported; versions below the minimum stay unsupported.
  • Update opentelemetry-api to 1.45.1 in backend locks.

mlx-agent-docs-v0.10.0.tar.gz

sha256 6c2fe1d521bcbfd29fb754fa38081e32f6046524a24f661991d40dcdd8e3177c

mlx-agent v0.9.0

Choose a tag to compare

@github-actions github-actions released this 21 Oct 00:16
v0.9.0
222ea7f
  • Add openclaw and hermes providers: the installer places the seven skills, each with its launcher and runtime, in $OPENCLAW_STATE_DIR/skills or <workspace>/skills, and $HERMES_HOME/skills or <project>/.hermes/skills.
  • Publish the mlx-agent package to PyPI with the mlx-agent and mlx-agent-mcp commands; provider installation from the package reports plugin_files_unavailable and points to a checkout.
  • Add read-only mlx_agent_capabilities discovery for MCP clients and mark invalid CLI arguments as non-retryable.
  • Enforce OpenCode extension ownership and clean up partial capture volumes.
  • Add policy checks for coding-capture commands.
  • Refresh the certifi pins and hashes in the backend dependency locks.
  • Update Pillow to 13.0.0 in backend locks and the experiment engine.
  • Run the full 27-case MCP battery with MCP Eval 0.8.0 in CI.

mlx-agent-docs-v0.9.0.tar.gz

sha256 da2f6a0312d711c11f1c36b906b6ec9937631c0c2a690e1b724166bed1e9d6bf

mlx-agent v0.8.0

Choose a tag to compare

@github-actions github-actions released this 19 Oct 21:31
v0.8.0
b243e09
  • Add durable inference tasks (task start|status|list|result|cancel|verify-model|capabilities, also through MCP) for classification, image and video generation and understanding, music generation, transcription, and speech synthesis: stable IDs, recent history, cancellation, and verified result artifacts; detached workers keep running after the caller exits and share the compute lease.
  • Add optional Strata serving for prepared Qwen3.8-Flash-Next GSQ-RCO GGUF profiles in an isolated runtime, with expert-memory and context-capacity controls checked against memory headroom; speculation is off by default.
  • Serving previews describe the weight format and declared quantization.
  • Add serve start --training-receipt and --adapter-path: serve a LoRA adapter on its pinned base; previews, startup, and each load-on-request reload recheck the receipt and the adapter's config and tensor hashes, and adapter tensors count toward the memory estimate.
  • Add lora inspect: read-only checks of the training executable, the cached base architecture, and both dataset splits.
  • Add adapter evaluation prompt suites: up to 16 prompts compared between base and adapter, with per-prompt preferences recorded apart from the generation report.
  • Add lab training-data validate|freeze|inspect for content-addressed media datasets with caller-selected readiness limits.
  • Add study recipes captured from cached Wan 2.1 Diffusers revisions, with weight, asset, runtime, and platform identities rechecked before and after generation.
  • Add serve start --thinking off: mlx_lm serves with the chat template's reasoning switched off; other runtimes record unsupported in the reviewed plan.
  • convert scan outputs report the backend that runs them for their task and whether it is installed; cached mlx-community snapshots count as MLX outputs even unquantized (Kokoro bf16), and snapshots whose weight files are missing do not.
  • intake resolve reports cached when the Hugging Face cache already holds every file a fetch would download.
  • Fix: training refuses splits smaller than one batch and base models whose architecture MLX-LM does not support.
  • Fix: convert scan --mlx-root inventories MLX models when no GGUF root is configured.
  • Fix: convert scan groups duplicate MLX model folders and reads Ollama stores under its roots: identical copies as exact with reclaimable bytes, other bit widths of one model as variant, Ollama models listed under ollama.
  • Fix: training and model diagnostics honor HF_HUB_CACHE, HF_HOME, and XDG_CACHE_HOME.
  • Fix: the MCP server negotiates the protocol version, declares which tools change state, and returns structured results with an output schema beside the text response; a CLI exit is returned as the result's exit_code instead of ending the server.

mlx-agent-docs-v0.8.0.tar.gz

sha256 c8ff5905a4e06aa0ce4f7b39bdd16532aecc6cd539926f6f479b78018200fd55

mlx-agent v0.7.0

Choose a tag to compare

@github-actions github-actions released this 12 Oct 18:20
v0.7.0
097326c
  • Add local music generation and requantization with checkpoint tokenizer validation.
  • Add shared experiment, blind review, workspace adoption, and text LoRA evaluation workflows.
  • Add native MLX Wan LoRA training and loading of converted Wan training inputs.
  • Add mlx_conversions to intake resolve: MLX models already quantized from the repository on the Hub (repo, downloads, library), for repositories that are not already MLX.
  • Select isolated serving backends from local model metadata.
  • Organize installation, model lifecycle, examples, troubleshooting, and contribution documentation into focused guides.
  • Remove undeclared Click imports from laboratory regression tests.
  • Record GitHub release deployments on the hosted publishing job.

mlx-agent-docs-v0.7.0.tar.gz

sha256 fd628c02791813cd36c5fa27f12e4a8ce02a049687615f207fa14dd28f555587

mlx-agent v0.6.0

Choose a tag to compare

@github-actions github-actions released this 08 Oct 14:10
v0.6.0
b95e972
  • Add serve start --jit: the loopback endpoint stays reachable while its local model is unloaded; chat and completion requests load one owned worker from the confirmed local files, concurrent requests share the load, and active streams block unload.

  • Add serve unload (receipt-fenced, --expected-pid) to release JIT weights while the endpoint stays reachable, and separate residency in serve status.

  • Add JIT memory policy: --idle-timeout, --keep-loaded, and --min-headroom-gb on serve start and serve policy; serve policy updates a running endpoint without a restart; cold loads check fresh OS headroom against estimated weights, runtime allowance, and reserve, and unknown measurements block loading.

  • Fix: convert describe decodes the answer with the tokenizer's special tokens skipped; moondream3 answers no longer start with <|md_reserved_4|>.

  • Fix: convert transcribe with Whisper vocabularies before large-v3 ends at the end of speech. mlx-audio resolved their missing <|nospeech|> to <|endoftext|> and suppressed it, so every transcript ran to the token limit with invented text; the runner uses <|nocaptions|> instead.

  • Add convert speak: synthesize one text with a converted text-to-speech model to a new WAV (voice, speed, language), returning audio length, synthesis time, real-time factor, and peak memory (read-only).

  • Add convert describe: answer one question about an image or a video with a converted vision-language model through its backend, returning token counts, throughput, and timing (read-only).

  • Add convert video: render one prompt with a converted text-to-video model to a new MP4 (size, frames, fps, steps, seed), returning frame count, duration, generation and load time, seconds per frame, peak memory, and pixel spread (read-only).

  • Add the mlx-video backend (hash-locked mlx-video-with-audio 0.1.37) with a t2v port: intake resolve and convert start take the original Wan2.1 T2V checkpoint (Wan-AI/Wan2.1-T2V-1.3B) to an MLX directory at 4 or 8 bits; new task type and backend category video_generation (pipeline tag text-to-video).

  • Add a DeepSeek-V4 DSpark drafter port (deepseek_v4_dspark): convert start --gguf on llama.cpp's dflash drafter GGUF writes an MLX sidecar, tensors named as DeepSeek's checkpoint (mtp.N as stages.N), FP8-sourced projections affine-quantized, MXFP4 routed experts repacked to MLX mxfp4 exactly, config.json with the DSpark fields and text_config. dspark-DeepSeek-V4-Flash-0731-Q8_0 at 4 bits: 4,730 tensors, 10.67 GB; F32, BF16, and FP4 tensors equal DeepSeek-V4-Flash-0731's own safetensors.

  • convert scan labels speculative-decoding drafters (dflash, eagle3 GGUFs; configs with dspark_target_layer_ids) speculative_draft (use case speculative_decoding), reports draft (port, target, block_size), and keys a drafter apart from its target so it is never grouped as the target's duplicate. Drafters no port converts are refused at plan time (unsupported_draft).

  • Committed provider packages hold one runtime per installed package instead of one per skill: tracked files under providers/ drop from 1871 to about 200. providers/claude/ is removed (Claude Code and the installer's claude provider use the repository-root plugin files); providers/codex and providers/agy carry one runtime at the package root, which their skills resolve as <plugin-root>/scripts/mlx-agent; providers/opencode and providers/agentskills carry none.

  • install agentskills copies the root launcher and runtime into each installed skill, and install opencode copies root src/mlx_agent; the cp -R providers/agentskills/... route is removed from the docs.

  • update and uninstall remove receipt-owned files that the current manifest no longer declares (the per-skill runtime copies of earlier releases), refusing any whose content changed. Transaction changes accept remove, receipted with a backup and restored by rollback.

  • Scout command and skill text point at the reference packs under the package runtime root (${CLAUDE_PLUGIN_ROOT}/src/..., <plugin-root>/src/..., <skill-dir>/src/..., <config-dir>/src/...) instead of a bare src/....

  • Fix: the OpenCode mlx_agent_command tool accepts all seven capabilities (it accepted only scout, adopt, and wire while bench, doctor, watch, and fleet commands called it).

  • Fix: the Codex plugin author is generated from the manifest publisher (organization_url) instead of a hardcoded GitHub profile.

  • Add the mflux backend (0.20.0, hash-locked) with an image_generation category and a Qwen-Image 2.1 port (qwen_image_21): its converter merges an optional PEFT LoRA into the transformer (W + scale·B@A, rounded to the weight dtype), quantizes with mflux, saves mflux's component layout, and writes a root config.json (components, quantization). Diffusers repos resolve by model_index.json class (port_pipelines); curated port_recipes name a repository's pinned base plus its LoRA (abenzerps/Qwen-Image-2.1-Uncensored-GGUF: Qwen/Qwen-Image-2.1@d26bb61 + its LoRA at 1.0, equal bit for bit to its BF16 GGUF on the checked tensors). intake fetch downloads a recipe's base snapshot and LoRA; convert start reads both from the cache. intake resolve reports recipe.

  • Add convert generate: render one prompt with a converted image-generation model to a new PNG (size, steps, seed).

  • convert scan lists outputs saved as per-component folders when config.json names the components.

  • Add the mlx-embeddings backend (0.1.0, hash-locked) with a classification category, and a Laya port (laya, convaiinnovations/laya root checkpoint): ModernBERT encoder plus the option-marker decision head, with the shipped temperatures. Repos without a root config.json resolve through a port's port_signatures; convert start --model-type runs a port's own converter (port_convert); port_quantize takes a group_size (Laya: encoder and head linears at group 32).

  • Subfolder checkpoints: org/name/folder and tree/<rev>/<folder> links resolve against that folder's files (config, weights headers, port signature, download size); intake fetch downloads only that folder; convert start --subfolder converts from the cached folder with any backend's converter. Laya's multilingual (mmBERT-base) and typed-decisions checkpoints convert this way.

  • The convert receipt schema allows backend (written by optional-backend conversions) and source.subfolder; tests validate real receipts against it.

  • Ports can declare the bit widths their conversions keep accurate (port_bits); intake resolve reports them as q_bits and convert start refuses others. Laya converts at 8 bits only: at 4 bits the multilingual checkpoint changes 3 of 25 reference answers.

  • intake resolve reports download_bytes, what a snapshot fetch takes; intake fetch --model-type downloads only a ported type's files (Laya: 846 MB of the 2.4 GB repo).

  • Add convert decide: answer typed questions (choice, score, noul) about a state with a converted classification model through its backend (read-only).

  • Task labels: text-classification and zero-shot-classification repos, and models a classification backend implements, are classification (use cases: moderation, routing, classification).

  • Add convert transcribe: transcribe one audio file with a converted speech-to-text model through its backend (read-only).

  • watch diff no longer reports a gated change when either side is unknown or missing.

  • intake resolve reports estimated_output_bytes (4- and 8-bit) for convertible repos from the safetensors headers; ports declare what they quantize (port_quantize). Edge0/Audio8-ASR-Infinite: 3.75 GB at 4 bits, matching the real conversion.

  • Add backend model ports: mlx-agent ships an MLX implementation of Edge0/Audio8-ASR-Infinite (audio8_asr_infinite: Voxtral Realtime audio tower, Qwen2 decoder with delay conditioning, projector, frame-length embedding, semantic VAD heads, streaming greedy decode with the 30 s rolling window) for the mlx-audio backend. intake resolve now reports Audio8 as convertible through mlx-audio; convert start --backend mlx-audio installs the port into the backend before converting. Conversion quantizes the decoder only.

  • Task labels: a model is vision-language only with a vision config or a vision module (from the mlx-vlm registry's vision files); GGUF main weights are never vision, and GGUF multimodal projectors (clip/mmproj) are labelled other.

  • Add local-path serving: serve start --path DIR serves a local model
    directory (e.g. a converted output outside the Hugging Face cache) through
    the same preview → confirm → receipt flow as --repo. The two are mutually
    exclusive; path plans are gated on the directory existing instead of cache
    membership, receipts carry path alongside repo, and status/stop argv
    matching accepts either.

  • Add intake resolve: classify a pasted Hugging Face link and report which declared MLX backend (mlx-lm, mlx-vlm, mlx-audio) converts it, or which components have no MLX implementation, without downloading weights.

  • convert scan items carry a task label (type, use cases, source, confidence).

  • Add backend list|install|remove: optional mlx-vlm and mlx-audio converters in isolated, hash-locked environments (confirmation-gated; removal moves to the Trash).

  • Add convert start --backend for repo conversions through an installed optional backend.

  • Add intake fetch and intake status: confirmation-gated, receipt-tracked downloads.

  • intake fetch skips weight formats MLX converters never read (TensorFlow, Flax, ONNX, Core ML, GGUF) in snapshot downloads.

  • Add intake port-analysis: deterministic component, weight-prefix, and custom-code analysis for architectures no backend implements.

  • Add intake port-plan: draft a porting plan from that analysis with a locally served model (loopback endpoints only).
    -...

Read more

mlx-agent v0.5.2

Choose a tag to compare

@github-actions github-actions released this 26 Aug 03:24
v0.5.2
3498f6e
  • Replace the discontinued Gemini CLI extension with a first-class Agy plugin package, including isolated native validation, user/project lifecycle coverage, and generated seven-skill bundles.
  • Centralize canonical CAVI publisher metadata and project it into generated Claude and Codex metadata without unsupported host fields.
  • Rename the former Gemini-only argument parser to shared command parsing, remove its retired file transport, and fix installer secret detection so canonical publisher URLs do not produce false positives.
  • Build release documentation artifacts in staging from the exact tagged commit, avoiding impossible self-referential committed provenance.

mlx-agent-docs-v0.5.2.tar.gz

sha256 d6900bb8d7fc4fda6ded9c76116014664dec6f8cb1973b336b1167b33d546d2a