Repository navigation
Releases
v0.1.8
am2Strata v0.1.8: a Workshop for making models and training data, a sandbox VM for computer use, HTTPS, API keys
with usage, one live WebSocket channel, installers for six platforms, and a regression suite.
Install or update
Linux (x86_64, arm64) and macOS (Apple silicon, Intel):
curl -fsSL https://github.com/abhaymin/am2strata/releases/latest/download/install.sh | shWindows (x64, Arm64), in PowerShell:
irm https://github.com/abhaymin/am2strata/releases/latest/download/install.ps1 | iexRunning it again over an earlier version keeps your models and settings. am2strata-run doctor then lists each
optional feature and what this machine still needs for it.
On Linux arm64, Intel Macs and Windows Arm64 the Rust engine runs on the CPU only. The Computer tab's desktop and the
sandbox VM are Linux-only; the VM also needs an x86_64 PC with KVM.
New
Workshop (a new tab, and am2strata-run workshop)
- Convert a Hugging Face Llama-family model to GGUF, byte for byte as llama.cpp's converter writes it; quantize
it (Q8_0, Q6_K, Q5_K, Q4_K, Q4_0, asllama-quantize --pure); merge a LoRA adapter from Unsloth or PEFT into its
base; install the result so it appears in the model list. All in Rust, no Python. - Training data written by the model you have loaded: questions and answers or multi-turn chats from your
documents (text, Markdown, code, HTML, PDF), or more examples like your own. Near-duplicates are dropped, an optional
judge pass scores each answer against its source, and every dataset gets a card. Review, edit and delete records in
the tab. - LoRA training from the command line (CPU):
am2strata-workshop trainfine-tunes Llama, Mistral and Qwen
models on a chat dataset, with checkpoints that resume exactly and a PEFT adapter the Workshop's merge takes.
For small machines: a half-precision base, training only the top layers with their inputs cached on disk, and
profile/autopilotto estimate memory and time and pick settings that fit;find-lr;soup.yamlconfigs
are read, and what is not supported is named. Its outputs match the reference implementation to 1e-5. - Jobs run one at a time in a separate process, and only inside the Workshop's folders.
Sandbox VM for computer use (Linux)
- A Debian 13 virtual machine am2Strata creates and runs itself (QEMU with KVM; no root, no Docker). The agent is root
inside it and can install or remove anything; this PC is untouched. Turn it on with
Settings › Computer and browser use › Mode: Sandbox VM, or use it yourself from the Computer tab. - Light: it only gets CPU time that nothing else is using, has a memory ceiling, no GPU, pauses when unused, and is then
saved to disk; a start restores it with its open apps in about a second. - Apps: Chromium, Firefox ESR, Bitwarden, Tailscale, NetBird, Mullvad, Surfshark, WireGuard and OpenVPN, VS Code,
LibreOffice and more with one click; anything else by Debian package, Flathub id,.debor AppImage link. - Ads: a DNS filter with Pi-hole's default list, the AdGuard DNS filter or both, a let-through list for jobs that
must see a site as it is, and an "Ads and trackers" app on the VM's desktop. am2strata-run vm status|start|stop|install|exec|ads|settings.
The Computer tab
- See the desktop computer use works on, live; Take control to use it with your own mouse and keyboard while
the agent waits; copy and paste between your clipboard and its clipboard.
HTTPS and secure WebSockets
- Settings › Secure connection (and
am2strata-run tls): a self-signed certificate made for you, your own
certificate, or a Tailscale certificate browsers trust. HTTPS andwss://are on port 8001 by default. - Plain
http://andws://keep working beside it. You can instead send other devices, or everyone, to HTTPS; HSTS
is sent only when everyone is.
API keys and usage
- Make keys while the server runs (Settings › Clients,
am2strata-run keys): a name, tool sets, an expiry, daily
request and token limits, allowed models. A key is shown once and stored only as a hash. - A generated key opens the model APIs and never the server's settings.
- Usage per key and per model, per day: requests, tokens, media jobs, errors.
One live channel
GET /v1/liveoverws://orwss://: the server pushes status, the queue, jobs and Monitor figures when they
change, and any API request can be sent over the same socket with streamed answers and cancel.- The web app's chat runs over it: answers stream over the open socket, Stop cancels there, and tool approvals,
media job cards and the canvas go the same way. Without the socket (or for a message over 16 MB) the chat uses HTTP
as before; a connection that drops mid-answer keeps what arrived and offers Retry, never resending by itself. The
Monitor tab uses it too.am2strata-run liveprints the pushes for scripts.
Installers and platforms
- Downloads for Linux arm64, Intel Macs and Windows Arm64 join the existing three. The installers detect the OS and
CPU, check the download's SHA-256, install without sudo, and can uninstall.
Regression suite
- A contract table of every route (who may call it, and what each kind of caller gets), recorded response shapes for
the compatible APIs, and one test per bug class fixed before.am2strata-run regressionchecks a running server
after an update and prints a pass/fail table.
Fixed
- A chat request whose JSON body was not an object caused a server error; it is now a 400.
- Seven owner routes that change state did not refuse requests from another site's page.
- A request with a broken body is now refused for its missing key before its body is looked at.
- Training-data chunks could exceed their size limit when they held blank lines or indentation.
- Merging an adapter into a base stored in F16 or BF16 read the weights as F32; it now reads each dtype, takes a
singlemodel.safetensorsas well as shards, and copies the config and tokenizer into the output. - The zip library is updated past five published advisories (one critical).
- Writes from the web app's own page over HTTPS (HTTP/2) were refused as another site's; they work now.
- Tests no longer race on shared fixtures or on the process environment.
Notes
- Sign-ins inside the sandbox VM (a password manager, a VPN, a browser account) are yours: open the app in the
Computer tab, take control and sign in. - Training is command-line and CPU-only in this release: no Train card in the tab yet, no GPU training, no QLoRA
(4-bit base) and no MoE expert adapters. - The installer downloads a headless Chrome for the browser tool without a checksum;
--no-browserskips it.
What's Changed
- Tests: the agents setup/reset test gets its own home folder by @abhaymin in #11
- HTTPS, HSTS and wss on port 8001: self-signed, your own certificate, or Tailscale by @abhaymin in #12
- Tests: no environment changes in test threads (the flaky agents test); a longer worker wait by @abhaymin in #14
- Model Workshop: merge, convert, quantize and install models in Rust — tab, CLI and server jobs by @abhaymin in #13
- Computer tab: see and use the sandbox desktop live — take control, clipboard both ways, UI and CLI by @abhaymin in #15
- Workshop: training data from documents (UI, API, CLI) by @abhaymin in #16
- HTTPS: keep http:// and ws:// working beside https:// and wss:// by @abhaymin in #17
- API keys made at run time, with usage per key and per model by @abhaymin in #19
- Sandbox VM for computer use (Debian, self-managed, light on the host) by @abhaymin in #18
- The live channel: one WebSocket for pushed updates and requests (GET /v1/live) by @abhaymin in #21
- Installers for Linux, macOS (Apple silicon, Intel) and Windows (x64, Arm64), with a doctor report by @abhaymin in #20
- Regression suite: a contract for every route, golden response shapes, bug regressions and a live runner by @abhaymin in #22
- README, banner and social preview for v0.1.8 by @abhaymin in #25
- README: an at-a-glance infographic of v0.1.8 by @abhaymin in #26
- Tests pass on macOS and Windows; regression targets run there in CI by @abhaymin in #28
- Security: patch zip to 9.2.2 in the main workspace (5 Dependabot alerts, 1 critical) by @abhaymin in #29
- Workshop trainer, milestone 1: LoRA training in Rust after Soup (CPU) by @abhaymin in #31
- Docs and help: present am2Strata as its own product, one Credits section by @abhaymin in #30
- Web app: the chat over the live channel (HTTP as the fallback) by @abhaymin in #33
- Remote models (off by default) by @abhaymin in #32
- Release v0.1.8 by @abhaymin in #23
- Trainer CLI test: free RAM is unknown outside Linux by @abhaymin in #34
Full Changelog: v0.1.7...v0.1.8
v0.1.7
am2Strata v0.1.7: computer use (a sandboxed desktop the chat can drive, with rules you choose), a fuller
browser, narration dubbed into other languages, and a quality check for media.
Install or update
Linux (x86_64) and macOS (Apple Silicon):
curl -fsSL https://github.com/abhaymin/am2strata/releases/latest/download/install.sh | shWindows (x64), in PowerShell:
irm https://github.com/abhaymin/am2strata/releases/latest/download/install.ps1 | iexRunning it again over an earlier version keeps your models and settings.
New
Computer use (off by default)
- Settings › Computer and browser use: switches for Browser use and Computer use, both off until you turn them on.
- Rules you choose: presets Strict (the default), Creator and Custom. Strict never types into password
or payment fields and asks before anything that sends, deletes, installs, buys or publishes. Every step is checked
against the rules before it runs; a refused step is never run and says which rule stopped it. - Modes: sandbox (a private virtual screen, never your real desktop), watch-only (looks, never clicks or types),
assist and trusted-job. One approval covers a job; a job stops at 200 steps (computer.max_steps). - The sandbox's own screen: started on first use and stopped after 10 idle minutes, refusing network
connections. Needsxvfbandxdotoolon Linux; without them the tools are simply not offered. macOS and Windows
desktops come later. - "Using the computer / browser" shows while a chat is at work, with a Stop button.
The browser
- More steps:
browser_tabs(list, switch, close; the last tab stays),browser_selectfor dropdowns, and
browser_waitfor text to appear (up to 30 s). - Site allow and block lists (
computer.site_allow/computer.site_block); a blocked site is refused before
it loads, and the block list wins. - Downloads are refused unless you set a download folder. With one, files land only there, with safe names, are
never opened, and Strict asks before keeping each.
Media
dub_narration: narration translated and spoken in another language, timed to the original beats, then mixed
under the music as a second audio track.check_media: checks a video or song the way a reviewer would: length, frame rate, still or frozen stretches,
loudness peaks and silence, and the words actually heard against the script.- The video maker: thumbnails render without a font installed, and the media-job hourly limit message is clear.
Made by am2Strata itself
Everything in this release was written, tested and pushed by am2Strata's own coding agent (opencode on the local
model), from its own plan for computer use. It was reviewed by Claude and by GitHub Copilot and Kilo Code, and the
agent fixed each review round itself, including a safety fix that keeps sandbox input off the real desktop.
Thanks
Everything listed in NOTICE, as before: Strata and PolyStrata, on which am2Strata is built; ACE-Step 1.5; mold.
What's Changed
- Media: compose_video review warnings from PR #2 by @abhaymin in #5
- Media: check_media and dub narration by @abhaymin in #6
- Computer use: rules, settings, desktop backend and run-loop wiring (off by default) by @abhaymin in #8
- Computer use builds on macOS and Windows; CI checks them on every PR by @abhaymin in #10
- Release v0.1.7 by @abhaymin in #9
Full Changelog: v0.1.6...v0.1.7
v0.1.6
am2Strata v0.1.6: clusters of PCs, a model optimizer, a video maker for the chat, long videos as chains of clips,
songs that continue, a new voice, and a browser that can upload a file and asks before it publishes.
v0.1.6 is the first release since v0.1.4 (a v0.1.5 tag was made, but its builds failed on macOS and Windows and it
was never released).
Install or update
Linux (x86_64) and macOS (Apple Silicon):
curl -fsSL https://github.com/abhaymin/am2strata/releases/latest/download/install.sh | shWindows (x64), in PowerShell:
irm https://github.com/abhaymin/am2strata/releases/latest/download/install.ps1 | iexRunning it again over an earlier version keeps your models and settings.
New
Clusters
- A master and its workers: a master routes chats to the workers that have the model (Settings › Cluster shows
each worker, whether it answers, its models and its queue). - Joining is simple and private: a worker joins with a pairing code and is found on the network by mDNS; master
and workers talk over mutual TLS 1.3 with the paired certificates pinned.
The model optimizer
- Tuned for this PC: it reads the machine (GPUs, the GPU link at its real width, the disk), searches the engines'
own speed flags within bounds, and keeps a tuned profile that is used at load. Settings › Models: tune a model,
follow the run, and switch its tuned settings on or off. - A quality floor: a tuned setting is kept only if the model still agrees with its default run.
- Quantization advice: each quant marked by how it fits this PC, with the recommended and the fastest.
Long videos and media
- A video maker the chat can use:
compose_video(andPOST /v1/media/videos/compose) joins pictures, clips,
audio, captions and a title and end card into one MP4, landscape or vertical for Shorts;make_thumbnail(and
POST /v1/media/thumbnails) puts big words on a picture for a 1280x720 thumbnail. - Long video as a chain of clips: Wan's one-frame carry between clips, the colour held to the first clip,
crossfaded joins, the carried frame redrawn by the image model (on by default, measured), and each step shown in
Monitor. - Big Qwen Image pictures decode on the CPU, and a video can start from an image.
- Media jobs survive a restart, ids never repeat, and "Make it again" re-runs one; media jobs wait for memory.
- Productions ready for YouTube: one request, many parts, one audio file, with chapters in the description's
form and captions from the finished audio.
Songs and voices
- Songs continue: a song continued from the one before (repainted past its end), starting from the last sound and
singing its lyrics; in the engine, the music API and productions. ACE-Step's audio encoder now runs in Rust
(110 dB SNR against the reference). - The planner writes a missing tempo, key and time signature first.
- Gautham, a new named voice: a warm Indian-English male. Indic Parler speaks each language with its own speaker;
Parler's pauses stay under 0.2 s and a held last sound is trimmed.
The browser (computer use)
browser_upload: attaches a file to a page's upload field, only from an upload folder the operator names
("browser": {"upload_dir": ...}); nothing outside it, no..or links pointing out.- Publishing asks first: clicks on Publish, Save, Upload, Delete, Share, "Make public" and the like wait for your
yes, as passwords and form submits already did. With"browser": {"approve": "job"}one yes covers that job on
that site for 30 minutes; any other risky step still asks each time. - Steadier: a hung page no longer takes the tool down; typing replaces a field's text; the same step three times
on an unchanged page is stopped.
Release fixes
- macOS and Windows build again: the disk speed's page-cache drop is Linux only.
- The video maker's captions and cards appear and
make_thumbnailworks (found by GitHub Copilot's review).
Clients and APIs
- Client profiles: their tools, memory and limits (image size, video length, web searches per answer), in
Settings › Clients;run_python, the browser and/v1/memoryfor profiles; media jobs per hour. - Responses API hosted tools:
image_generationandmcp. Gemini'sgoogle_searchis answered by the server's
web search. - Downloads show their progress: percent, done and left, speed and time left.
Made by am2Strata itself
The browser upload and approval features and the video maker were written, tested and published by am2Strata's own coding agent
(opencode on the local model), reviewed by Claude and by GitHub Copilot and Kilo Code, and fixed by the agent through
four review rounds before merging.
Thanks
Everything listed in NOTICE, as before: Strata and PolyStrata, on which am2Strata is built; ACE-Step 1.5; mold.
What's Changed
- Machine: the disk speed's page-cache drop is Linux only (fixes the v0.1.5 macOS and Windows builds) by @abhaymin in #4
- Release v0.1.6 by @abhaymin in #7
Full Changelog: v0.1.5...v0.1.6
v0.1.4
am2Strata v0.1.4: songs that sing their lyrics, MCP servers from the app, the server's tools for every client,
and a server that keeps going: requests routed by priority, models switched quickly, overload told plainly.
Watch what's new (4 min): https://youtu.be/HP4ztKaqOyo
Install or update
Linux (x86_64) and macOS (Apple Silicon):
curl -fsSL https://github.com/abhaymin/am2strata/releases/latest/download/install.sh | shWindows (x64), in PowerShell:
irm https://github.com/abhaymin/am2strata/releases/latest/download/install.ps1 | iexRunning it again over an earlier version keeps your models and settings.
New
Songs
- The song planner, in Rust.
- ACE-Step 1.5's 5 Hz planner (a 1.7B model) now writes a song's structure before the music is made, so the
vocals follow the lyrics. - Measured with Whisper on one RTX 5060 Ti:
- 60 s English songs: 47% of the lyrics heard before, 91% now (4 seeds);
- 3-minute Hindi songs: 48% before, 60% now (5 seeds).
These figures were measured with a tempo, key and time signature given. Takes vary a lot by seed: on the
live server, two 60 s songs scored 81% and 17%. A request without a tempo and key leaves the planner less to
go on.
- Give
bpm,keyandtime_signature, and ask for"takes": 2or more: the server keeps the take whose words
come through best. - Next: the planner writes a missing tempo and key itself, as ACE-Step does.
- Checked against ACE-Step's own Python, step by step.
- A minute of song takes about 27 s with the planner (7.5 s without).
- Turn it off with
"planner": falsein/v1/audio/music. - The planner's files (3.7 GB) join the song model's download.
- ACE-Step 1.5's 5 Hz planner (a 1.7B model) now writes a song's structure before the music is made, so the
Tools
- MCP servers from the app (Settings › Tools):
- add a server by its address or as a program on this PC, or paste a Claude Desktop / Cursor / VS Code
mcpServersblock; - test a server before saving it, edit it, turn it off, remove it, or turn single tools off;
- changes apply at once, with no restart.
- Header and environment values are always shown masked.
- A program server can be added only from this PC, unless the run config sets
"mcp_remote_commands": true.
- add a server by its address or as a program on this PC, or paste a Claude Desktop / Cursor / VS Code
- Client profiles:
"clients"in the run config gives each client its own API key and tools (memory, web,
image, video, speech, music, MCP servers).- Agents that send only a URL and a key (
unsloth start, Codex, Claude Code) get those tools without flags;
x-strata-toolsnarrows them for one request. - Each profile keeps its own memory.
- A client key opens the model APIs, not the server's settings.
- Agents that send only a URL and a key (
- The agents' own web search works against am2Strata for a key whose profile has
web:- Claude Code's WebSearch (Anthropic's
web_searchserver tool); - Codex's Responses
web_search; - Chat Completions'
web_search_options.
- Claude Code's WebSearch (Anthropic's
- Memory for API calls, on by default: apps calling with the main key get remember, recall and forget, and
matching facts in the prompt, in their own "api" space ("strata_memory": falsegoes without; Settings › Memory
has the switch). am2strata agentrun <agent> --tools web,imageasks for tools from the command line.
A server that keeps going
-
A crashed or killed engine is replaced: the next request loads the model again instead of failing until a
manual reload. The interrupted request is told what happened, and the alerts say so. -
No silent failures: an internal error is answered as an error, never a dropped connection, and a cut stream
ends with an error event. -
The memory guard:
- am2Strata learns how much memory each model needs;
- a model that would not fit is refused plainly, before the working model is stopped, which keeps answering.
-
Overload told plainly:
- answers carry
x-strata-warningwhile the PC is low on memory or requests are waiting; - every "busy" refusal says when to retry (
Retry-After).
- answers carry
-
Request routing: fixed rules, applied in order (docs/PRODUCTION-PLAN.md):
- a person at the web app goes before API calls, and those before background media jobs (
x-strata-priority
asks for a class); - a job that waits rises a class every minute;
- jobs for the loaded model go first, within a bound, to save a switch;
- different clients take turns.
/api/queueand a waiting stream give the expected wait. - a person at the web app goes before API calls, and those before background media jobs (
-
Quick switching:
- while a job waits for another model, that model's files are read into memory ahead of the switch (when there
is room); - each model's load time is learnt and used in the estimates.
- while a job waits for another model, that model's files are read into memory ahead of the switch (when there
Plans
New plans in the ledger (docs/LEDGER.md):
- production readiness (docs/PRODUCTION-PLAN.md);
- every web app feature through the API (docs/SERVER-TOOLS-PLAN.md);
- clusters of machines, phones included (docs/CLUSTER-PLAN.md).
Thanks
ACE-Step 1.5 (MIT, ACE Studio and StepFun), its planner included; and everything listed in NOTICE, as before:
Strata and PolyStrata, on which am2Strata is built.
Full Changelog: v0.1.3...v0.1.4
v0.1.3
am2Strata v0.1.3: songs, speech in your language with native voices, deeper chats, and a Monitor that pages.
Install or update
Linux (x86_64) and macOS (Apple Silicon):
curl -fsSL https://github.com/abhaymin/am2strata/releases/latest/download/install.sh | shWindows (x64), in PowerShell:
irm https://github.com/abhaymin/am2strata/releases/latest/download/install.ps1 | iexRunning it again over an earlier version keeps your models and settings.
New
- Chat can dig deeper: an answer may take 100 tool rounds (searches, pages, tools), up from 8. The run config's
mcp.max_roundssets the default, and the chat settings' Tool rounds per answer (orstrata_max_rounds) up to
500. On long runs, older tool results are shortened before the conversation outgrows the model's context, and the
answer says so. - Speech in the language of the text:
/v1/audio/speechfinds the language itself (language: "auto", the
default) and hands it to a model that speaks it. The web app's Text to speech gets a Language picker, and Read aloud
speaks each answer in its own language. Speech may now be up to 20,000 characters (about 20 minutes). - Native female and male voices (
voice: "female"or"male"; Laura and Lea are female, Jon and Gary male) in
Spanish, French, Portuguese, Polish, German, Italian, Dutch and Hindi, made by Parler-TTS mini multilingual and Indic
Parler-TTS, loaded in Rust. Indic Parler-TTS also speaks Malayalam, Tamil, Telugu, Bengali and other Indian
languages. - Each sentence of speech is checked: Whisper hears it, and a take that adds or drops words is made again.
- Songs (ACE-Step 1.5, ported to Rust): a Song tab under Media Generation and
POST /v1/audio/music. Write
lyrics, pick one of 32 genres (Bollywood, bhajan, ghazal, qawwali, Carnatic fusion, Malayalam melody, bhangra, pop,
rock, jazz, lo-fi, cinematic and more) or describe the sound, choose the vocal language and a length up to 10
minutes. On an RTX 5060 Ti a minute of song takes about 12 seconds. Several takes can be made, keeping the one whose
lyrics come through best. - Chat makes songs and music: with media on, the chat model can call
make_song(sung) andcompose_music
(instrumental), as it callsgenerate_image. - API calls are kept across restarts (the newest 2,000 on disk, under the data folder's
api-calls), so old calls
are there to trace a problem; Clear all empties them. Calls from earlier days show their date. - Monitor: the request history comes in pages (25 to 200 a page), and the table scrolls inside its card.
- Settings › Media models gets a Music filter, and speech models can be downloaded from the web app again.
Fixed
- The media engine no longer gives an answer to the wrong request after a client goes away mid-job (it showed as a
400 or a speech job that never ended). - Chatterbox speech that ran on with babble at the start or end of a sentence.
- A media route (
/api/media/genres) answered without the API key; every route needs it.
Good to know
- Songs, video, images and the new voices need the build from source with an NVIDIA card (the media engine is not
in the downloads; see the README). - Songs are an early version: the song planner is still to be ported, so sung words do not always follow the lyrics
(best on songs of about a minute; try another take). Hindi and Spanish sing well; Malayalam less clearly.
Each model has its own licence, which is yours to read (ACE-Step 1.5: MIT; Parler-TTS: Apache-2.0). am2Strata itself
is MIT.
Problems and ideas: issues.
Full Changelog: v0.1.2...v0.1.3
v0.1.2
am2Strata v0.1.2: fixes to the web app's Monitor and media pages.
Install or update
Linux (x86_64) and macOS (Apple Silicon):
curl -fsSL https://github.com/abhaymin/am2strata/releases/latest/download/install.sh | shWindows (x64), in PowerShell:
irm https://github.com/abhaymin/am2strata/releases/latest/download/install.ps1 | iexRunning it again over an earlier version keeps your models and settings.
Fixed
- Monitor: with a long history, or media players in its rows, the requests table spread under the cards below it.
It now scrolls inside its card, and the row is taller. - Media models: while an image or a video was being made, Settings › Media models waited behind the job and said
"the media engine is not built". It now shows the last list at once, says that a job is running, and says
"Loading" until the server answers.
New
- The web app shows the server's version beside its name (
/metricscarries it as"version"). - Monitor: a System RAM card shows the RAM used of the total, the share used, what is free, and its history.
Each model has its own licence, which is yours to read. am2Strata itself is MIT.
Problems and ideas: issues.
Full Changelog: v0.1.1...v0.1.2
v0.1.1
am2Strata v0.1.1: coding agents in one command, the Gemini API, a Monitor that keeps its history, and Chatterbox
speech in the media engine.
Install
Linux (x86_64) and macOS (Apple Silicon):
curl -fsSL https://github.com/abhaymin/am2strata/releases/latest/download/install.sh | shWindows (x64), in PowerShell:
irm https://github.com/abhaymin/am2strata/releases/latest/download/install.ps1 | iexAlready on v0.1.0? Run the same command again: it replaces the program and keeps your models and settings.
New
- Coding agents in one command:
am2strata agentrun <agent>starts an agent on your models, on this PC or on
another one (--url http://<host>:8000 --api-key …). Everything after the agent's name goes to the agent.- Started for you: Claude Code, Codex, opencode, Pi, Qwen Code, Aider, Goose and the Gemini CLI.
am2strata agent setup <agent>writes an agent's own config once (it keeps a backup, andagent resetputs it
back): Claude Code, Codex, opencode, Pi and Factory Droid.- Oh My Pi, Crush, Hermes, OpenClaw, Continue, Cline, Kilo Code, Zed and VS Code get the exact settings to enter,
in Settings › Connect › Coding agents and witham2strata agent show <agent>. unsloth startreaches am2Strata too (UNSLOTH_STUDIO_URLandUNSLOTH_API_KEY).- Not supported: the Antigravity CLI and the Cursor CLI, which take no custom model endpoint.
- See docs/AGENTS.md.
- The Gemini API:
/v1beta/models/<model>:generateContent,:streamGenerateContentand:countTokens, with
the key asx-goog-api-keyor?key=. Gemini clients and the Gemini CLI work by changing the base URL. - Monitor:
- the request history is kept on disk until you clear it;
- each request shows the model it ran on;
- media jobs keep their output: images, video and audio play in the table;
- a running job is named by its kind ("Speaking · part 2 of 3", "Making an image").
New in the build from source (the media engine)
- Chatterbox text to speech (Resemble AI, MIT), through ONNX Runtime with the
onnxfeature:- Chatterbox Turbo (English, with tags such as
[laugh]) and Chatterbox Multilingual (20 languages here, Hindi,
Spanish and French among them); - the voice of a 5 to 10 second clip (
voice_audioin/v1/audio/speech), or a built-in voice; - it runs on the CPU, beside the text model, so chats do not wait. Measured on a Ryzen 7 7840HS (8 threads):
9.2 s of speech in 10.7 s, 5 GB of RAM at its peak; - it needs ONNX Runtime 1.23 (
sudo apt install libonnxruntime1.23, orORT_DYLIB_PATH); - clone only a voice whose speaker has agreed to it.
- Chatterbox Turbo (English, with tags such as
Fixed
- A server whose default model's run config has no
api_keynow uses the key of the newest other run config, and
says which. Before, changing the default model to one without a key left a server on--host 0.0.0.0open to the
network. - A media job whose client went away no longer shows as running for ever in the Monitor.
- The README's "Not yet" claims only what was measured.
What to expect
As in v0.1.0:
- the download's engine runs on the CPU, and suits small models;
- the fast NVIDIA path, and image, video and speech generation, need the build from source
(docs/DEPLOYMENT.md); - by default the server answers only this PC: set an API key before
--host 0.0.0.0.
Each model has its own licence, which is yours to read. am2Strata itself is MIT.
Problems and ideas: issues.
Full Changelog: v0.1.0...v0.1.1
v0.1.0
The first public release of am2Strata: run AI models on your own PC, behind a web app and OpenAI-, Anthropic- and
Responses-compatible APIs. One download for Linux, Windows or macOS. No Python, no Docker, no CUDA toolkit.
Install
Linux (x86_64) and macOS (Apple Silicon):
curl -fsSL https://github.com/abhaymin/am2strata/releases/latest/download/install.sh | shWindows (x64), in PowerShell:
irm https://github.com/abhaymin/am2strata/releases/latest/download/install.ps1 | iexThen:
am2strata pull mistralai/Ministral-3-3B-Instruct-2512-GGUF Q4_K_M
am2strata # open http://127.0.0.1:8000Each installer checks the download's SHA-256 and needs no admin rights. To get the files by hand instead, take the
archive for your system below; QUICKSTART.md inside takes a few minutes.
In this download
- A local server with OpenAI
/v1/chat/completionsand/v1/responses, Anthropic/v1/messages, streaming and
tool calls. Existing clients and SDKs work by changing the base URL. - A web app: several chats, a canvas (HTML previews, Mermaid flows, a whiteboard), a monitor and settings.
- A Rust inference engine for GGUF models: Llama and Mistral, Qwen3 and Qwen3-MoE, Qwen3.5-MoE, DeepSeek-V2-style
MoE and Nemotron-H. It runs on the CPU, and on any GPU through--devices(wgpu; new and less tested). - A Hugging Face downloader:
am2strata pull <repo>lists the quantizations, andpull <repo> <quant>
downloads one, checks it and writes its run config. - Tools the model can use, when you switch them on:
- web search and page reading;
- a browser: a headless one, or your own Chrome with your logins;
- memory across chats;
- MCP servers;
- PDFs as text and page images. Windows carries Poppler; on Linux and macOS,
install.sh --pdfsets it up.
What to expect
- Speed: the CPU engine suits small models. A 3B model answers at a readable pace; a 35B MoE runs at a few tokens a
second on a desktop CPU. - Fast on NVIDIA: building from source on Linux with PolyStrata's CUDA engines runs large MoE models at 60-160
tokens a second on one RTX 5060 Ti 16 GB. Image generation and editing, video and voice need that build too. See
docs/DEPLOYMENT.md. - Safety: by default the server answers only this PC. Set an API key before using
--host 0.0.0.0. - Unsigned binaries: Windows SmartScreen or macOS may warn about a file downloaded in a browser. The installers
avoid the macOS warning.
Each model has its own licence, which is yours to read. am2Strata itself is MIT.
Problems and ideas: issues.
Full Changelog: https://github.com/abhaymin/am2strata/commits/v0.1.0