One of Jamaica’s first AI agent apps
An Android app built in Jamaica for running a whole company of AI agents (hiring them, giving them work, and approving what they do) from a phone. A Jamaican AI agent platform that fits in your pocket.
Your AI company, in your pocket. Hire AI agents, chat with them like teammates, hand them tasks, let them build apps, draw pictures and search the web, and approve anything that matters. On free AI models with no key, no account and no card, or on your own Claude, GPT, Gemini, DeepSeek and Grok keys.
Version 1.9.15 · Android 8.0 or newer · Signed APK, 25.3 MB · Made in Jamaica by Nathan, CorX Labs



Automa is an Android app from CorX Labs, made in Jamaica, that runs a whole company of AI agents for you. You create an organization with a goal, hire agents into an org chart, give them work in plain words, and watch them plan it, do it and check in when they need you. Every task traces back to the organization’s goal, hires, budgets and plans wait for your approval, and every token is counted against a budget with a hard stop.
The agents work on an Automa server you control: your computer, a cloud machine, or the phone itself inside the free Termux app. Nothing sits in the middle: no Automa cloud account, no ads, no tracking. And it runs on free AI models with no key, no account and no card, including Moan, Automa’s own Jamaican model that talks Patois.
Automa 1.9.15 is here to download. Get the signed APK from this page: 25.3 MB, for Android 8.0 or newer. It has also been submitted to APKPure, and will be released there once APKPure finishes its review. Download and install.
com.corxlabs.automaWhat Automa is, in the words people search for, each one something the app can back up.
An Android app built in Jamaica for running a whole company of AI agents (hiring them, giving them work, and approving what they do) from a phone. A Jamaican AI agent platform that fits in your pocket.
Designed, built and shipped by Nathan Lyttle, the 14-year-old founder of CorX Labs, working alone in Jamaica.
CorX Labs also built CorX3.8-27B, Jamaica’s first large open-weight language model, and the NatureV1 hurricane model.
Moan answers in Patois first, and reads Jamaican news, law and government sources into a live index so every reply cites source, date and link. Meet Moan
The server the agents work on can run on the phone itself, in Termux: no computer and no cloud bill. How
Twenty-plus free chat models and two free image models, ready the moment the app opens. The list
Build, run and share apps in 69 languages with an AI pair or an AI team, from the phone. App Studio
From 1.0.0 on 27 September 2026 to 1.9.15 on 22 November 2026. Every release
Version 1.9.6 was a full audit of the app across 16 areas, with every confirmed finding fixed. Security
Automa is not one chatbot. It is a team of agents that each have a job, and between them they cover most of the AI work a small business, a student or a developer does in a week.
Agents search the web across many results, read and crawl pages, and answer with tables and linked sources. Chats
Launch posts, Instagram and WhatsApp captions, emails and plans: written as tasks you can read and approve. Tasks
Real Excel files with formulas, PowerPoint decks with charts and pictures, and Word documents, and they read the ones you attach. Free models
A paired image model draws from a description, or redraws an attached photo with the change you ask for. Pictures
App Studio builds, runs and shares apps in 69 languages, with an AI that sees the page and fixes what breaks. App Studio
Claude Code, Codex, Gemini CLI and OpenCode in their full terminal UI, on the phone. Terminals
Projects, goals, routines on a schedule, approvals and budgets, the plumbing of a company, run by agents. Projects
Moan answers in Patois and knows Jamaica’s news, government, courts and statutes, with source, date and link. Moan
A shop, a bakery, a salon: agents check prices against the market, write the posts, draw the logo and work out how to register the business. An example
Research with sources, spreadsheets, decks and documents, and an app studio to learn to code in, on free models, with no card.
Captions, launch posts, pictures and logos, and routines that run the weekly work on a schedule.
Claude Code, Codex and Gemini CLI in a terminal on the phone, App Studio, App Inspector, and any OpenAI-compatible endpoint. Terminals
Most AI tools assume a US-dollar credit card, a desktop computer and an English that is not the one spoken at home. Automa was built in Jamaica, and it shows in three places.
The free models need no key, no account and no card. A first agent can start on them with one tap, Skip for now: start on the free models, and you add paid keys later only if you want them.
The Automa server can run on the phone itself, inside Termux. No computer, no cloud bill: the app, the agents’ server and your data all live on one Android phone.
Moan speaks Patois first and switches to Standard Jamaican English for formal documents and code. It knows Jamaica’s history, government, courts, statutes, economy and culture.
Moan runs on Nemotron 3 Ultra with a Jamaica primer and a living knowledge index. While Moan is in use, Automa keeps reading Jamaican sources (the Gleaner, the Observer, JIS, the Office of the Prime Minister, Parliament, the Bank of Jamaica and the Laws of Jamaica) into a local, searchable index. Every reply gets the passages that match the question plus the newest news, with source, date and link.
Elsewhere in the Caribbean, the same agents search and read the web with sources, so a question about Trinidad and Tobago, Barbados, Guyana or the Bahamas gets a researched, linked answer too. Moan’s built-in index is the Jamaica-specific part.
The screenshots on this page come from a sample organization: Yaad Bites, a family bakery in Kingston selling patties and hardo bread, with the goal Grow Yaad Bites online. It runs three agents. Moan is the Jamaican assistant, and answered a Patois question about registering the business. Nova, the marketing lead, checked patty and bread prices against the market and wrote the launch post for the order page. Pixel, the brand designer, drew the logo.





More AI built in Jamaica: CorX3.8-27B, a 27B Jamaican Patois assistant and the island’s first large open-weight language model, and NatureV1, a research model for Caribbean hurricanes.
http://192.168.1.20:3100,
or tap Run Automa on this phone and the app sets up a server inside Termux,
with a progress bar for each step. A computer can also be linked over a USB cable or Wireless
debugging.



A bottom tab bar (Home, Chats, +, Tasks, Files) stays in thumb reach, a sidebar holds every section and the organization switcher, the bell opens the Inbox, and your avatar opens the account sheet. Long-press the app icon for shortcuts: New chat, New task, App Studio, Change server, Files. Every screen is edge-to-edge, keyboard-aware, and works in portrait and landscape on phones, tablets, foldables and Chromebooks. Destructive actions come with Undo, not “are you sure?” dialogs.
Home is your day at a glance: what needs you, who is working right now, what the free models saved, and live run activity.


Talk to an agent like a teammate. Replies stream token by token; the model’s thinking streams into its own panel and folds away when the answer starts. Tables wider than the screen scroll sideways. Source links show the site’s favicon. When a reply fails, the chat says why in plain words, and you can switch model and send again.




Pictures. Ask any free chat model for a picture and its paired image model draws it. Attach a photo and say what to change, and the model redraws it. Tap a picture to see it full size or Download it to the phone; Draw again asks for a fresh take.
The model switcher in the chat header lists every provider with its key status and the maker’s logo, a Free filter, thinking depth (Auto, Quick, Balanced, Deep, Deeper, Max), and switches for what a free agent may do: web search & reading, files & commands, teammates & tasks, connected apps. A paired image model gives a chat model the power to draw; a small “Bonded” mark shows the two makers’ logos together.



Messages. Copy, retry, delete and thumbs up or down on the newest reply; press and hold any other message for the same menu. Attach photos, screenshots, PDFs or any file with +. Unsent drafts survive leaving the chat. ⋯ on a chat pins, moves or deletes it.
Slash commands: /plan (plan first; nothing runs until you say
so), /ask (answers only, no changes), /auto, /model,
/think deep, /context, /stop, /clear,
/skills, /help, /new.
Notifications. When an agent finishes a reply, or an App Studio job ends, while you are outside the app, a notification says so. Tap it to jump straight to that chat, task or project.
Pick a free model in the switcher and the agent runs on it; your other models stay where they are. Every model stays listed and says how it is doing: ready, busy, at today’s free limit and when it is back, or closed by its provider. Nothing is hidden. On 17 November ten free versions were offline at their providers, DeepSeek V4.1 Flash’s among them. From 1.9.15, Check on Terminals and in the chat’s model picker asks every free model for one word and marks the ones closed at their provider.
| Kind | Models |
|---|---|
| Chat | Moan (Jamaican, on Nemotron 3 Ultra) · Nemotron 3 Ultra 550B · Nemotron 3 Super 120B · Nemotron 3 Nano Omni 30B Reasoning · Nemotron 3.5 Lightning · Nemotron 3.5 Super VL · Nemotron 3.5 Content Safety · DeepSeek V4.1 Flash · Gemma 4 31B · Gemma 4 26B A4B · Qwen 3.8 27B · Qwen 3.8 27B Abliterated Cyber · Inkling · Inkling Small · Flint v1.0 · North Mini Code · Atria Dawn Preview · Apodex 1.1 Mini · LFM 2.5 2.6B · Xing 4.0 29B A4B · Ternary Bonsai 27B · Ling 3.2 Flash · ERNIE 5.5 Flash Preview · Particle 1 · LFM3 1.7B · Free Router, which sends each question to an available free model |
| Image | SDXL Lightning (fast, the default pairing) · SDXL Base 1.0 |
Search the web across many results, read pages, crawl a site and download files; read, write, edit and search files and run commands in its workspace; make real Excel spreadsheets with formulas, PowerPoint decks with charts and pictures and Word documents, and read the spreadsheets, documents, decks, PDFs and CSVs you attach; draw with its paired image model; message teammates, hand off work, and create and update tasks; and request the connected apps it needs. It knows its role, its teammates and your name, and it keeps going through busy providers and dropped connections without ever switching models behind your back.
The Que engine sits under every free-model call: up to ten calls at once, at most two at a time on any one model, a pool of app IDs that grows on rate limits, retries on busy answers, a five-minute answer cache, conversations and batches. Many questions at once all get answered instead of some failing.
Everything the agents are doing, each one a thread you can read, steer and approve. A task has one assignee, a parent and sub-tasks, blockers, comments with @mentions, a plan document, attachments and work products, labels, a priority and a status : backlog, to do, in progress, in review, blocked, done. The task page shows the live run with the agent’s thinking and tools, its cost and its full history. An agent can ask you a question with options right in the task; pick one and it carries on. When it finishes, it asks you to approve the result: approve to mark the task done, or send it back with what to change.




Inbox and approvals. Decisions waiting on you: hires, budgets, plans, join requests, failed runs and anything an agent got stuck on. Approve, decline or retry from the phone. The bell shows the count and clears when tapped. The Decisions page keeps the record of every automated and governed decision.


Each agent has a role and title, an engine and model with the maker’s logo, a system prompt and instructions, a budget, skills, tools and a manager. The agent page shows its live runs, logs, progress, instructions and tools. Pause, change or offboard an agent any time. New agent is a short wizard; the teams catalog installs a whole team, a CEO with executives and engineers, in one go.
Engines: Claude Code · Anthropic API · Codex · Cursor (local and cloud) · Gemini CLI · Grok · Hermes · Kimi · OpenCode · OpenClaw gateway · Pi · Free models · metered API engines · a REST or webhook engine for an agent you host elsewhere. On a phone, a Claude agent answers through the Anthropic API with Claude Code’s own tools running in Termux.



Projects group tasks under one goal and keep their own workspace. Goals are the top of the task tree, the “why” every task hangs from. Routines run work on a schedule, a daily summary, a weekly review, and each run is a task you can read. Pipelines chain stages of work, cases collect related work into one file, and workspaces show where an agent’s run happens, with its live service ports.




Artifacts are every file the agents produced, grouped by task, with a viewer
for text, images and code. Downloads go to Downloads/Automa.
Files shares one folder on the phone with your agents. With auto-send on, a file an agent asks for is sent without asking; otherwise the request waits for you to pick a file or decline. Agents can send files back. Use Share to Automa from any app’s share sheet and the file waits on the Files page until you confirm.
Auto-send removes you from the loop, that is the feature and the risk. With it off, every file leaves the phone because you tapped Send. With it on, anything inside the shared folder can leave without a decision at the moment it happens. Share a folder that exists for this, not your whole Documents tree.


Skills are what agents know how to do. Install them from the catalog or build one in Skill Studio; a new skill is available from the agent’s next run. Connectors are the apps agents can reach: GitHub, Slack, Gmail, Notion and more. You choose which agents get access, which accounts they use and which actions they may take : three choices, not one switch. Connecting an app stays inside Automa.


Build, run and share apps with an AI pair, or a whole AI team, that writes the code and fixes what breaks. No key needed: the free models work in the AI panel and for the editor’s code suggestions.




The AI assistant reads and searches the files, writes and edits code, runs commands and tests, starts and stops the app, reads the console, searches the web, and keeps going until the job is done. It sees the page like you do: it opens the running page at phone or desktop size, reads what is on screen, the console errors and the layout trouble, fixes what it found and looks again.
AI teams. Pick teammates in the model sheet and the lead model hands them work: several at once, each with a role and the files it owns, talking to each other while they work. The Stats tab shows a running clock, steps, files changed and every model at work with live tokens, then the totals and what the free work would have cost.
Undo and safety. Undo any change the assistant made, step by step. Checkpoints are saved before every AI message, Stop is instant, and nothing is ever written outside the project folder. Build together: invite people with a join code or link, for an hour up to no limit. Guests see your code as you type, chat, send files and suggest changes you review as a diff; they can’t save files or reach the rest of Automa.




Terminals give you a real coding terminal on the device: Claude Code, Codex, Gemini CLI, OpenCode and others in their full terminal UI, on a Debian layer the page sets up for you. Wire a terminal to an agent so it works with that agent’s key, pick a model per tool, or open the free models chat with nothing to install.
App Inspector takes an app on your phone, or any APK, apart: it decodes the manifest, resources and code, shows the code as Java, lets you search, edit and ask Claude to explain any file, and rebuilds a signed APK to your Downloads. It is for apps you made or are allowed to analyze.


Every token is counted: agent runs, chats, App Studio replies and code suggestions land on Costs by agent, by task and by month. The organization and each agent have a budget, and a hard stop pauses spending at the limit. The usage limits card shows how much of the free models’ monthly allowance is used and when it renews, each free model’s daily limit with a countdown, and Claude’s session and weekly limits when Claude Code is signed in. The free models card on Home and Costs shows what the free work would have cost at paid prices.


Search finds tasks, comments, plan documents, artifacts, agents and projects in one place. Audit holds the log of everything that happened, who did it and when, with the runs, costs and budgets behind it. The work timeline shows every agent’s runs side by side.



Organization settings: general (name, goal, logo, budget), members and access (invites, roles, join requests), API keys (one place for the keys agents, chats and App Studio run on, with a monthly spending limit) secrets, adapters, environments, plugins and Experimental switches. The account sheet shows your profile, your API balance, the tour, the dark theme and sign-out, and About Automa keeps the full version history.




| Term | What it means |
|---|---|
| AI agent | An AI model given a job, tools and the freedom to take steps on its own. In Automa every agent is an employee: a role, a title, a manager, a model, a system prompt, a budget and skills. |
| Organization | One company: a goal, agents, tasks, a budget and settings. One server can host several, and every page is scoped to the current one. |
| Adapter | How an agent runs: Claude Code, Codex, Gemini CLI, Grok, Hermes, Kimi, OpenCode, Pi, the free models, a metered API, or a REST or webhook adapter for an agent hosted elsewhere. |
| Task | A unit of work with one assignee, a parent, sub-tasks, blockers, comments, a plan document, attachments and work products, and a status that moves through review and approval. |
| Chat | A conversation with one agent: technically a task in conversation mode, so it keeps the same history, attachments and cost tracking. |
| Approval | A governed action waiting on you: a hire, a budget, a plan, or anything an agent got stuck on. Agents never spend or hire without a yes. |
| Routine | Work that runs on a schedule, such as a daily summary or a weekly review. Each run is a task you can read. |
| Project and goal | Projects group tasks under one goal; goals are the “why” the whole task tree hangs from. |
| Connector | An app agents can reach, such as GitHub, Slack, Gmail or Notion, with per-agent access and per-action permissions. |
| Skill | Something agents know how to do, installed from the catalog or written in Skill Studio. |
| Workspace | The checkout or sandbox where an agent’s run happens, which can expose live services such as a dev server. |
| Artifact | A file an agent produced, attached to a task and downloadable from Artifacts. |
Three parts, and it matters which one is where. The phone holds the app and one shared folder, the whole board UI ships inside the APK, so the layout is the same whatever server it talks to. The server holds the agents, their tasks and their routines, and does the work. The connections are how the server reaches services that are not yours to host. Solid lines carry work outward; dashed lines are the three things that come back to you.
No computer needed. Install Termux from F-Droid, tap Run Automa on this
phone in the app, and paste the command Termux is given. The installer sets up Node.js
and PostgreSQL, the server and an automa command, then starts Automa on
127.0.0.1:3100, locked to the app. The first install downloads a few hundred
packages and takes 10 to 45 minutes; updates are much faster.
| Command | What it does |
|---|---|
automa run | Run in the foreground, restarting the server if it stops |
automa start · stop · status | Run in the background, stop it, check it |
automa update | Update the server; it backs up the database first |
automa backup · restore | Save a database backup now, or put one back |
automa link | Hand the app key to the Automa app again |
automa adb link | Let agents drive the phone over adb: tap, type, screenshot |
The app starts Automa for you when it isn’t running and shows Automa is starting over your page while the server comes up. Backups are made every hour while Automa runs and before every update, and a restore never loses the database. Set Termux’s battery use to Unrestricted; with Termux:Boot, Automa starts when the phone boots.
You need Android 8.0 or newer, about 100 MB of storage, and internet for the AI features.
Signed APK · 25.3 MB ·
Android 8.0 or newer · package com.corxlabs.automa · version code 37
SHA-256
e6acca41e926446a7331c56787b5438bc9276b4511f21bb895c1ced84c2faa7d
APKPure listing pending review. Automa has also been submitted to APKPure. Once APKPure is done reviewing it, it will be released there too, so updates can arrive on their own. Until then, this page is the download. There is no Automa cloud account to create, ever. Follow @corxlabs on X to hear when it goes live there.
automa update in Termux
after updating the app so the server matches it.1.9.15, 22 November 2026. IMY comes to Terminals. IMY (Imagenery) is a strict
compiler for code you write your own way: the first line says what you are building, in a header such as
# ("adds two numbers and shows the sum"), and under it you write each step in a syntax you make
up. IMY checks it, turns it into another programming language and runs it, and never changes your file.
It converts as you type: each line you enter is checked at once, and the result appears above the line
you are typing. A second menu on the IMY tile picks the language your code becomes: Python, JavaScript,
Ruby, Rust, C, Go, Assembly and thirteen more. IMY knows 48 languages, and builds and runs 42 of them when
the language’s own tool is installed. :import takes in a project you already have, at
most 25 files of up to 40 kB each, so your code can use its functions and variables by name; Python files
are read on the device, other languages by the model with every answer checked against the file, and none
of the project’s code is run. A program its own build tool refuses goes back to the model to be
corrected before anything runs. IMY stops and names the lines at an English request in place of code, at
anything used before the line that defines it, and at a line it cannot be sure about. The first tap on
the IMY tile installs it (inside the Linux layer on a phone); the next opens its console, with nine ready
programs to start from. It works with Automa’s free models with no key, and the model you pick is
the one used: one that is not being served says so instead of quietly using another. Your own Anthropic,
OpenAI or Gemini key works too, and anyone can install IMY on a computer with pip. A Check button on
Terminals and in the chat’s model picker asks every free model for one word and marks the ones
closed at their provider; one that was only busy is counted apart and left unmarked. One more free model,
LFM3 1.7B. Fixed: a terminal starts at the width your screen shows, so a tool’s first lines no
longer break mid-word on a phone; Running now lists only terminals still running (one whose program had
ended stayed there for a minute); a tool that finishes installing in the background turns Ready by
itself; the installer’s messages fold between words on a narrow screen. It installs over 1.9.14 and
keeps your data; the APK is 25.3 MB, up from 25.1 MB. Checks before release: IMY’s 570 tests, the
board’s 7,399, and the APK’s own server bundle installed the way a phone installs it and
walked through at phone size from an account that never had IMY. The free models’ key is readable
in none of the APK’s 16,853 files.
1.9.14, 18 November 2026. Automa in your language. Account, Language offers Auto and 40 languages. Auto follows your phone’s or browser’s language and finds your country from its time zone, with nothing looked up online; dates, numbers and times follow your language and country, and 23 formats stay in US English on purpose, such as dollar amounts and chart axes. The main screens’ 210 strings are built in for every language; the rest of the board is translated by Automa’s free models the first time a screen opens, and kept on your server for everyone on it. Only Automa’s own wording is sent for that, never your chats, tasks, names or files, and your own words are never translated. Arabic, Persian, Hebrew and Urdu read right to left. Agents write to you in your language, and answer in another one if you write in it; Moan keeps Patois for English. The app follows the choice too: the whole connect screen in every language, and on Android 13 and newer the app’s own notices, shortcuts and notification channel. The choice is saved to your account, so every phone and browser you sign in on uses it. Fixed on the way: dates already on screen change format as soon as you switch language, and App Studio no longer leaves errors behind for a project that is gone or a dropped connection. It installs over 1.9.13 and keeps your data; the APK is 25.1 MB, up from 23.9 MB. Before release every test in the repository ran, 28,950 in 2,024 files; each failure was fixed, or came from the test machine (an older Node or Rust, or a busy moment) and passes without it.
1.9.13, 17 November 2026. Model lists that keep themselves current. An agent’s model choice, App Studio’s providers, the Terminals menus and the chat’s Newest models now follow a public catalog of the major makers’ models, newest first, so a model released after this version shows up without an update, and App Studio marks the last 30 days’ as New. Added now: Claude Fable 5.2, Gemini 4 Carbon and Gemini 4 Hydrogen, and Qwen 4 Max on OpenRouter. Three new free models, Ling 3.2 Flash, ERNIE 5.5 Flash Preview and Particle 1, each answered and called a tool in a live test. A free agent with no model picked now runs on Nemotron 3 Ultra: it was trying DeepSeek V4.1 Flash, whose free version is offline. Security: the key Automa uses for the free models is stored scrambled (1.9.12 kept it in readable pieces), and every build checks that it is readable nowhere in the app; a phone running Automa can still recover it. Your phone’s database password and app key are locked to Termux on every install; sign-in tries are counted per real caller, so a made-up address for each guess no longer gets around the limit; and other websites can no longer show the board inside a frame. On the phone, PostgreSQL 18 runs 4 helper processes instead of 7 and the server no longer runs git every few seconds, so Android has less reason to close Termux. Fixes from a review of 1.9.12: Troubleshoot and Safe start each get their own answer from Termux (two at once used to share one), Troubleshoot waits while a Safe start runs, and after Restart normally the safe mode banner keeps checking past its 3-minute wait. It installs over 1.9.12 and keeps your data. The release checks ran 8,047 tests in the board and shared projects and 225 in the server files it touches; one test failed once on a busy machine and now waits for what it checks.
1.9.12, 17 November 2026. Staying up on the phone. When Termux closes with
“[Process completed (signal 9)]”, Android has ended its processes: Android 12 and newer do
that once all apps together run more than 32 in the background, and every Android does it when memory
runs low. The database on the phone now runs fewer processes and uses less memory, and the server
keeps 6 database connections at most. The new automa keep-alive lifts Android’s
process limit over Wireless debugging, once (it stays after a restart), and takes Termux off battery
optimisation. When Automa stops, Termux says why and keeps a record (automa logs). After
three stops in a row within five minutes of starting, Automa starts in safe mode by itself: the board
opens with agents, schedules and plugins off, and a banner says why, with Restart normally and
Troubleshoot. automa troubleshoot, or Troubleshoot in the app, checks Termux, storage,
memory, the database and Automa, fixes what it safely can (a lock left behind, an update cut short, a
stopped database) and never touches your data or backups. A Refresh button in the header loads the
page again and reconnects live updates. It installs over 1.9.11 and keeps your data. 61 tests are
new; of 28,099 in the full run, 3 failed the first time and pass after their fixes.
1.9.11, 16 November 2026. Fixes from a scan of the whole app. Saving your name or photo while the connection drops now says “Can’t reach Automa right now”, and the App Studio page an invited collaborator uses says it can’t reach the session; both showed “Failed to fetch”, and the change is not sent twice. An agent’s settings page loads the extra settings some AI connections add (Codex, Gemini and others) by itself once Automa answers again, where they stayed missing until the app was reopened. After a reconnect, pages you had already left stop asking the server every 10 seconds. The command-line tool packs a 64 MB company folder for import in milliseconds, where that step took several seconds. It installs over 1.9.10 and keeps your data. Checked by opening 112 pages in a browser at phone size and running 28,210 tests; 10 of them failed on the first run and pass after their fixes.
1.9.10, 12 November 2026. Starting up and reconnecting. While Automa is still
starting on the phone, the connect screen now shows which step it is on, how long it has waited and
“step x of y”, instead of waiting with no end in sight. When the connection drops and
comes back (the network returns, or you reopen the app), the board reconnects by itself, where
before it showed “Failed to fetch” until the app was reopened, and a dropped connection
no longer sends you back to sign-in. In Termux, automa run follows the start step by
step and opens the app when it is ready. It installs over 1.9.9 and keeps your data. Checked on the
server and the board; not yet on a phone.
1.9.9, 9 November 2026. Fixes from a scan of every page and a live run of the app. When the free AI service asks Automa to slow down, Automa now waits the minute it asks for, once for all agents, and carries on, instead of retrying four times in 25 seconds and failing the reply. Home no longer shows “Task unavailable” cards for chats and tasks you deleted, and deleting a chat or task also cancels the retries still waiting for it, so its agent isn’t left “still working” for hours. The bring-your-own and Vercel connect pages no longer open with a sidebar for a connection that doesn’t exist.
| Version | Date | Highlights |
|---|---|---|
| 1.9.15 | 22 Nov 2026 | IMY in Terminals: code in your own syntax becomes one of 48 languages as you type, and 42 of them build and run where their tool is installed; a check of which free models answer; one more free model, LFM3 1.7B |
| 1.9.14 | 18 Nov 2026 | Automa in your language: Auto and 40 languages for the board, the connect screen and the app; right to left for Arabic, Persian, Hebrew and Urdu; agents write to you in your language |
| 1.9.13 | 17 Nov 2026 | Model lists that keep themselves current, with Claude Fable 5.2, Gemini 4 Carbon and Hydrogen and Qwen 4 Max; three new free models; the free models’ key stored scrambled and sign-in limits per real caller; fewer processes on the phone |
| 1.9.12 | 17 Nov 2026 | Fewer database processes on the phone and automa keep-alive for Android’s process limit; safe start after three quick stops; automa troubleshoot; Refresh in the header |
| 1.9.11 | 16 Nov 2026 | Two more places say “Can’t reach Automa” instead of “Failed to fetch”; an agent’s settings load by themselves after a reconnect; company import packs in milliseconds |
| 1.9.10 | 12 Nov 2026 | The connect screen shows each startup step; the board reconnects by itself instead of “Failed to fetch”; a dropped connection no longer signs you out |
| 1.9.9 | 9 Nov 2026 | Waits out the free AI service’s slow-down instead of failing the reply; no dead cards on Home; deleting a chat cancels its waiting retries |
| 1.9.8 | 8 Nov 2026 | Starts even when an old queued run can’t be verified; no 30-second stall; web search falls back to Bing |
| 1.9.7 | 3 Nov 2026 | Stays on your page when you come back; deleting a chat stops a reply at once |
| 1.9.6 | 2 Nov 2026 | A full audit of 16 areas with every confirmed finding fixed: the app proves the server holds its key, agents never see server secrets, connectors sign in inside the app, agent budgets enforced |
| 1.9.5 | 30 Oct 2026 | Notifications when a reply or App Studio job finishes while you’re away; live Stats tab; the AI looks at the page like you do |
| 1.9.4 | 30 Oct 2026 | App Studio teams with a shared team chat and per-model token shares |
| 1.9.3 | 30 Oct 2026 | A growing pool spreads free-model requests, so many people asking at once all get answers |
| 1.9.2 | 29 Oct 2026 | Server updates on the phone switch over only when complete; a failed update leaves Automa running |
| 1.9.1 | 28 Oct 2026 | Usage limits on Costs; failed replies say why; Moan’s facts brought up to date |
| 1.9.0 | 28 Oct 2026 | The Que engine under every free-model call; App Studio with no key |
| 1.8.8 | 27 Oct 2026 | Moan, the Jamaican model; free agents make spreadsheets, decks and documents |
| 1.8.7 | 27 Oct 2026 | Free agents keep their model and recover in place; maker logos on agents |
| 1.8.6 | 26 Oct 2026 | Free chat agents paired with SDXL Lightning, and draw when asked |
| 1.8.5 | 26 Oct 2026 | Start on the free models without a key; a database backup before every update |
| 1.8.4 | 26 Oct 2026 | Free models in chat that stream, think, search, use tools and draw |
| 1.8.2–1.8.3 | 26 Oct 2026 | Ask Claude from any app; automa update; one signing key from here on |
| 1.8.0–1.8.1 | 24 Oct 2026 | Terminals, App Inspector and your API balance |
| 1.7.0 | 24 Oct 2026 | Claude agents with Claude Code’s tools on phones; thinking modes |
| 1.6.x | 22 Oct 2026 | Live replies with visible thinking; token and cost counts everywhere |
| 1.5.0 | 21 Oct 2026 | Slash commands, app shortcuts, the offline sheet |
| 1.4.x | 10 Oct 2026 | App Studio; the app locked to its own Automa; the tour |
| 1.3.0 | 29 Sep 2026 | Run Automa on the phone; drag-to-connect org chart |
| 1.1.0–1.2.0 | 28 Sep 2026 | Continue with Google, push notifications, the org chart on phones |
| 1.0.0 | 27 Sep 2026 | The first Automa for Android |
Automa is an Android app, made in Jamaica by Nathan at CorX Labs, that runs a whole company of AI agents for you. You hire agents, give them roles, chat with them like teammates, hand them tasks, let them build and run apps, draw pictures, search the web and work with your files, and you approve anything that matters. The agents work on an Automa server you control: your computer, a cloud machine, or the phone itself inside the free Termux app.
Automa is one of the first AI agent apps built in Jamaica. It was made by Nathan, the 14-year-old founder of CorX Labs, the lab behind CorX3.8-27B: Jamaica’s first large open-weight language model. Jamaica’s AI scene grew quickly in 2026, with other platforms launching too, so CorX Labs does not claim Automa was the very first; it is the one that puts a whole company of AI agents, a Patois-speaking model and an app studio on one Android phone.
Yes, a Jamaican AI agent platform that fits on a phone. Where most AI apps give you one assistant, Automa gives you the pieces of a platform: agents with roles and budgets, tasks and approvals, projects and goals, an app studio, coding terminals, files, skills and connectors, all running on an Automa server you control.
An AI agent is an AI model given a job, tools and the freedom to take steps on its own (searching, reading, writing files, running code) instead of only answering one message. In Automa every agent is an employee with a role, a title, a manager, a model and a budget, and anything that matters waits for your approval.
A chatbot is one assistant in one conversation. Automa is a team: several agents with roles, an org chart, tasks with owners and plans, approvals, budgets, schedules and files, all on a server you control. You can still just chat with any agent, and the chat keeps its history and cost.
Small businesses that want help with research, marketing, logos and paperwork; students who want research, documents and a place to learn to code without a card; creators who need captions, posts and pictures; and developers who want Claude Code, Codex and Gemini CLI on a phone.
Yes, with Automa’s Terminals: Claude Code, Codex, Gemini CLI and OpenCode run in their full terminal UI on a Debian layer the app sets up for you, with Automa running on the phone in Termux. On a phone, Claude agents answer through the Anthropic API with Claude Code’s own tools.
No. Pick a free model and start: Automa’s free models need no key, no account and no card. Add keys later under Settings → API keys if you want Claude, GPT, Gemini or another paid model, and set a monthly spending limit so nothing overspends.
No. Tap Run Automa on this phone and the server runs inside Termux on the phone. A computer, a virtual machine or a cloud machine works too, and is what you want if the agents should keep working while the phone sleeps.
Yes. Moan is Automa’s own Jamaican model, free with no key. It answers in Patois first and in Standard Jamaican English for formal documents and code, carries a primer on Jamaica’s history, government, courts, statutes, economy and culture, and keeps a living index of Jamaican sources (the Gleaner, the Observer, JIS, the Office of the Prime Minister, Parliament, the Bank of Jamaica and the Laws of Jamaica) so its replies cite source, date and link.
Automa itself is not tied to Jamaica: you install the APK yourself and the agents run on a server you control. Free agents search and read the web with sources, so research about Trinidad and Tobago, Barbados, Guyana, the Bahamas or anywhere else works the same way. Moan’s built-in knowledge index is the Jamaica-specific part.
Twenty-plus free chat models and two free image models with no key, including Moan, Nemotron 3 Ultra, Gemma 4, Qwen 3.8 and Ling 3.2 Flash, plus SDXL Lightning and SDXL Base for pictures. With your own keys or subscriptions it runs Claude, GPT, Gemini, DeepSeek, Grok and Kimi, and engines such as Claude Code, Codex, Cursor, Gemini CLI, OpenCode and any OpenAI-compatible endpoint.
Yes. A free chat model paired with an image model draws pictures and logos, and redraws an attached photo with the change you ask for. Free agents also make real Excel spreadsheets with formulas, PowerPoint decks with charts and pictures, and Word documents, and read the files you attach.
Yes, in App Studio: 69 languages, an AI pair or a whole AI team that writes the code and fixes what breaks, a sandboxed preview, a public link in one tap, undo for every AI change, and friends invited by code or link. It works on the free models with no key.
With the server on the phone, Automa itself keeps working offline. The AI features need internet. When the phone loses the network, a sheet says so and counts down to the next try, and what you typed is kept.
There is no Automa cloud account. Your organizations, tasks, chats and files live on your own Automa server, and sign-in exists only on that server, when you turn it on. Free-model requests go from your server to the free AI service Automa uses.
Yes. Invite members with roles; an invite can’t give more than the person who made it has. App Studio projects can be shared with friends by join code or link.
No. Automa is for Android 8.0 or newer, on phones, tablets, foldables and Chromebooks.
Tap Download Automa 1.9.15 (APK) on this page, open the file on an Android phone, and allow Install unknown apps if Android asks. The APK is 25.3 MB and signed. Automa has also been submitted to APKPure, and once APKPure finishes its review it will be released there too, so updates can arrive on their own. Every release from 1.8.2 on installs over the previous one and keeps your data.
Nathan, the 14-year-old founder of CorX Labs in Jamaica, who also built the CorX language models and NatureV1. Automa builds on an MIT-licensed open-source control plane for AI agents.
Automa 1.9.15
Hire agents, hand them work and approve what matters, on free models with no key, from one Android phone. Made in Jamaica by CorX Labs.