https://github.com/user-attachments/assets/cba7775e-f61e-4625-aefd-76417e72ca33

DecisionsAI

The last agent you'll ever need.

A voice-first development agent for your computer, your projects, and your phone.

Website · Start here · Install · Voice & dictation · Phone · Workflows

---

DecisionsAI is an agent

DecisionsAI is an agent that helps you develop faster and get everyday work done on your computer. Talk to it, dictate into any app, ask it to control your desktop, or give it work inside a real project. It can use your files, tools, skills, coding agents, and connected services to carry the request through instead of only telling you what to do.

On the desktop, your main point of entry is the Oracle—or, more simply, your avatar. It is a small animated character that stays on top of your other windows, giving you a persistent place to speak to DecisionsAI or open its controls without hunting for another app window. It shows when the agent is idle, listening, thinking, or working. You can configure how it looks and choose the Oracle skin or another avatar; the character can change, but the DecisionsAI agent underneath remains the same.

Its local web interface is the control deck. That is where you chat, organize projects, track tickets, see what is running, review results, and step into a workflow when you want more detail. You do not have to keep it open while the agent works.

Telegram is the remote control. From your phone, you can send the same agent a message, voice note, screenshot, or document; approve a decision; steer work already in progress; and receive the result when it is done.

WhatsApp can be a source of client work. Link one or more contacts or groups to an existing ticket board, then ask the agent from Telegram to check what came in. DecisionsAI can collect the relevant messages, media, and voice transcriptions, show you what it found, and wait for approval before creating tracked work. When approved work reaches QA, it can prepare a WhatsApp reply and return it to Telegram for you to send, revise, or leave as a draft. It does not send merely because a worker finished, and only you move the ticket from QA to Complete.

DecisionsAI is model-agnostic, provider-agnostic, and CLI-agnostic. It does not lock your projects or memory to one AI company. Start with a local model, connect a cloud provider, use Codex, Cursor, Claude Code, Pi, or another coding CLI, and swap them later without rebuilding the way you work.

Start with the voice agent. Add projects, tickets, Telegram, and workflows when you need them.

Start simple

At its simplest, DecisionsAI gives you two voice hotkeys:

• Hold Option + Command (the default macOS push-to-talk shortcut) to speak to the agent, ask for something, and hear its reply.
• Hold Control + Command (the default macOS dictation shortcut) to turn your speech into text inside whichever app you are using.

“Open Safari.”
“Reply to this message.”
“Take what I’m saying and clean it up.”
“Summarize what is on my screen.”

The agent can talk back, control your computer, run an action, or answer in Chat. Dictation simply writes what you say at the cursor. Both modes can run locally, so you do not need to begin with projects, workflows, or a collection of cloud accounts.

That is the first layer: ask for something and the agent does it.

Use your voice in three different ways

Voice is not a single mode in DecisionsAI. You can use it according to what you are trying to do:

The default macOS shortcuts are Option + Command for push-to-talk and Control + Command for dictation. Shortcuts are editable in Preferences → Shortcut Keys, so they can fit around the tools you already use.

You can begin here and never touch a workflow. Voice, dictation, shortcuts, snippets, Chat, and actions are useful on their own.

Use it from your phone

Telegram gives you the quickest way to reach the same agent away from your desk. Send a text, voice note, screenshot, or document and receive the answer in the same conversation.

When you need more than messaging, send remote to the Telegram bot. DecisionsAI returns a secure link to its mobile web interface, where you can talk or type to the agent, use snippets, view the screen, click, scroll, and transfer files. It is an encrypted web remote for your own DecisionsAI instance—not a separate native mobile app or a second agent.

At this stage it can still be simple: ask a question, send something to your computer, or tell the agent to perform one action. The project and workflow layers become useful only when the request needs to be tracked or carried through several steps.

Then give it a project

Link a real project folder and continue speaking naturally:

“On my website, make the green order button black.”

DecisionsAI uses the active project to understand what “the green button” refers to. It can create a ticket so the change is trackable, send the work to your coding agent or CLI, check the result, record the time, and tell you what changed.

You do not need to name a model, write a technical prompt, or manually assemble a workflow for a small request.

You ask  →  DecisionsAI does the work  →  You get the result
                         │
                         └─ A project ticket keeps the history

Projects give the agent a home for the work. Tickets give each request a visible record. Memory keeps useful facts, decisions, files, failures, and next actions available when you return—even if you change models later.

Let it handle a larger job

Now ask for something bigger:

“Rebuild the checkout flow, fix the mobile layout, test it, and report back when it is ready.”

This is where the orchestrator and workflows become useful. DecisionsAI can split the request into tickets, use the existing Development workflow, and move the work through six clear stages:

Understand → Plan → Build → Review and test → Fix what failed → Report

The workflow is not the product you have to operate. It is the method DecisionsAI uses when a request is too large or risky for a single action. The Workflows screen is there when you want to inspect progress, see which worker is active, review evidence, or steer the run.

DecisionsAI reuses a suitable workflow before considering a new one. The Development workflow already covers project context, implementation, independent review, testing, correction, reporting, and memory. Specialist workflows are only useful when they add something genuinely different.

Grow into it

The product becomes more capable as your work becomes more demanding:

Start with one model if that is all you need. Later, you can give planning, coding, vision, Computer Use, and review to different workers. DecisionsAI keeps the project and memory stable while Codex, Cursor, Claude Code, Pi, Ollama, OpenRouter, or another configured provider does a particular part of the job.

As the work grows, Telegram grows with it. The same conversation that handled a quick voice note can receive progress, present approval buttons, accept a correction, steer a running workflow, and return the final report. The web interface becomes mission control rather than another inbox you must watch.

Ollama runs models. DecisionsAI runs work.

What you get

Ready to try it? Jump to Installation. You can start with the voice agent and add the rest later.

---

The Local Web Interface

DecisionsAI spins up a local-only web UI (not exposed to the internet). Open it from the Oracle menu or the in-app tray.

---

System Requirements

Offline / local mode

DecisionsAI detects your system RAM at first launch and picks models that fit. Local models use your machine's memory. Cloud models (marked :cloud) run remotely and do not load their weights into local RAM.

Recommended local setup (12 GB+ RAM):

Role	Model	RAM needed
Chat	ornith:9b	~6 GB
Coding	ornith:9b	~6 GB
Vision	qwen3-vl:2b	~1.9 GB
Image	x/flux2-klein:latest	local only

Local-only fallbacks (10 GB+ RAM):

RAM	Chat model	Coding model	Approx. VRAM
10–11 GB	qwen3:4b	qwen2.5-coder:3b	~3.5 GB
12+ GB	ornith:9b	ornith:9b	~6 GB

Online / cloud mode

No large model downloads. Only Whisper.cpp and Kokoro install locally. Use local models for sensitive work and cloud models when you need bigger weights.

---

Installation

One-liner

curl -fsSL https://decisionsai.net/install.sh | bash

Quick start

git clone https://github.com/tensology/decisionsai.git
cd decisionsai

The launcher handles dependency checks, Python setup, model downloads, and launch automatically.

When Codex, Cursor, or the Claude-compatible harness surface are available on the machine, setup and every bin/start.py run recalibrate the harness stack: repair IDE plugins, re-project skills (so plugin reinstall does not wipe them), refresh MCP recommendations, and merge safe MCP servers into Cursor/Codex configs. Third-party keys such as Composio and Cursor API tokens are stored in Preferences → API Keys and injected at recalibrate time — you should not need to edit ~/.cursor/mcp.json by hand.

Manual installation

# 1. Python environment
python3 -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# 2. Dependencies
pip install -r requirements.txt

# 3. Download AI models
python bin/setup.py

# 4. Start
python bin/start.py

System deps: brew install portaudio ffmpeg (macOS) · sudo apt-get install portaudio19-dev ffmpeg (Linux) · winget/Chocolatey/Scoop (Windows)

Optional components

---

Quality checks

Run the test suite with:

pytest

This runs the unit and property-based tests, skipping optional heavy dependencies by default.

Keyboard & Voice Commands

Voice commands

Exact phrasing can vary.

Global shortcuts

Defaults are editable in Preferences → Shortcut Keys.

---

Integrations

Telegram

Connect in Preferences → Advanced → Telegram. Telegram is the fastest way to hand DecisionsAI work from anywhere: send text, a voice note, screenshot, document, correction, or follow-up. The same orchestrator can answer directly, create/update a project ticket, start a workflow, pause for a decision, or steer an existing run. Inline buttons and typed or spoken replies support approve, reject, continue, stop, and feedback. Progress and final reports return to the same conversation. Type remote for an HMAC-encrypted link to the full remote-control UI: screen stream, click, type, scroll, and file transfer.

WhatsApp

Connect WhatsApp in Preferences → Advanced, then link a contact or group feed to a Ticket Board. DecisionsAI can sync incoming text, voice/media metadata, captions, and conversation context; configured boards can auto-snapshot new messages into durable tickets. From there the normal approval and orchestration policy can route the ticket to a project, direct action, or workflow. This deliberately keeps ordinary conversation from becoming unintended execution: only the feeds and autonomy boundaries you configure become work.

Google Workspace

Connect via Preferences → Advanced → Google (OAuth 2.0).

IDE Integration

Codex, Cursor, the Claude-compatible harness surface, and other coding backends attach to project context. Ticket-board work, IDE chats, workflow runs, and automation runs can report back to the orchestrator when the project link is set.

Setup installs or repairs the local Codex and Cursor plugins when those tools are present. They emit session events into Orchestrator so the desktop agent can see IDE progress instead of losing it inside the editor.

Harness stack (orchestrated from distr/core/harness_stack.py, run on bin/setup.py and quietly on every bin/start.py):

Workflow runs can push a pre_chain of skills into the active project harness (browser, design, agent-reach, composio, yt-dlp, etc.) based on ticket text and project surface.

The Decisions agent exposes ide_thread to list, read, and prompt Codex/Cursor sessions. Composio API keys live under Preferences → API Keys; saving recalibrates MCP headers automatically.

Orchestrator holds tickets, boards, automations, workflows, browser runs, project folders, and IDE sessions in one ledger. Local skills live under skills/; vendor packs under plugins/*-pack/.

Models are roles, not a lock-in

Preferences separates the conversational/orchestrator LLM, coding LLM, vision model, image model, Computer Use model, Step Runner, and board agent. Each workflow step can override the backend, provider, model, tools, skills, cost tier, and independence requirement.

With opt-in Auto routing, DecisionsAI evaluates the step role, complexity, risk, capability requirements, and recorded failures. It can plan with Codex, implement bounded work with a free/local Ornith model, choose an independent reviewer such as HY3, and escalate to more expensive providers only when the lower tier has produced evidence that it is stuck. The chosen route and reason appear immediately in Runs/chat activity.

Ornith 1.0 is a particularly strong local fit: the open-source agentic-coding family includes 9B and 35B variants with tool calling. DecisionsAI supports ornith:9b for accessible local work and configured 35B variants for more capable machines, while keeping the project memory portable if you swap it out later.

---

Workflows

Multi-step workflows run on the Step Runner. Loops are reusable development and operational playbooks on top of that runner. A group of project tickets can move through planning, implementation, independent review, tests, browser QA, approval, and reporting while each step uses a different model or CLI. Import a loop preset, steer a waiting harness from Telegram or the web UI, and inspect validation, routing history, executor context, evidence, heartbeats, and reports in the existing Runs/chat surfaces.

---

Technology Stack

You do not need to understand this stack to use DecisionsAI. These are the main components for people who want to extend or audit it.

Offline core:

Optional cloud services:

---

Project Structure

bin/                 # Launchers and setup scripts
distr/
├── core/            # Agent pipeline, LLM, STT, TTS, actions, workflows
├── gui/
│   ├── dialogs/     # About, preferences windows
│   ├── oracle/      # Oracle overlay and tray
│   └── web/         # Local web UI (templates, static, API)
plugins/             # IDE plugins (codex-ide, cursor-ide), ECC, and vendor harness packs (competition, agent-reach, community-skills, yt-dlp)
skills/              # Local Decisions skills (harness stack, playwright, composio, design references, …)
sidecar/             # Go binary: machine control (macOS/Windows)
assets/
├── avatars/         # Skin packs (Clippy, Nugget, Rusty, Masko, etc.)
└── readme/          # README images
docs/                # Orchestrator docs (local planning notes stay gitignored)
.artifacts/          # Gitignored local runtime output (tickets, pi skills, cursor handoffs)
tests/               # Property-based and unit tests

---

Legal References

Connected services can include WhatsApp, Telegram, Gmail, Jira, Trello, shared chat rooms, uploaded files, voice notes, images, project folders, CLI/IDE logs, model-provider requests, and workflow audit trails. Review the public policies before connecting external accounts or data sources.

---

Contributing

Suggestions, bug reports, and pull requests are welcome. Open an issue to discuss a change before sending a PR. Check existing issues first to avoid overlap.

License

Licensed under the TENSOLOGY COMMUNITY LICENSE AGREEMENT. See LICENSE.md for details.

---

Note: This project has no cryptocurrency or token associated with it. Any coin using the DecisionsAI name is not affiliated with us.