Ion

Agent Teams that ship

Comment

Ion - AI Agent Orchestration & Workflow Automation Platform

Project Description

Ion is a high-performance native desktop AI work IDE that enables seamless orchestration of AI agents across multiple platforms and unified workflow automation. It is engineered for developers, power users, and teams who need structured control over AI tasks—from quick queries to complex, multi-agent workflows with human-in-the-loop approval gates.

Ion bridges the gap between disparate AI platforms (Nebius, Codex, OpenCode, DeepSeek, Pi, Oh My Pi, and more) into a single, unified workspace. With Nebius Token Factory and NVIDIA Nemotron at its center, Ion provides a private, durable environment where users can define specialized agent teams, construct automated multi-stage pipelines, manage credentials securely in the OS vault, and inspect exact usage and costs.

Core Features

1. Nebius Token Factory & Multi-Provider AI Orchestration

  • Nebius Token Factory as core engine: Direct integration streaming text, reasoning tokens, and tool proposals from NVIDIA Nemotron 3.5 Lightning and DeepSeek models.
  • Support for 13+ AI platforms: Seamlessly switch between Nebius, Codex, OpenCode, DeepSeek, Pi, Oh My Pi, Antigravity, Fx, Grok, Kimi, and local models.
  • Automatic harness discovery: Dynamic detection of installed coding harnesses and live model catalog discovery directly from your provider keys.
  • Strict separation of intent and execution: The model proposes actions; the local daemon verifies, budgets, and executes them.

2. Team-Based Agent Execution

  • Hierarchical agent teams: Define agents with their own specialized instructions, model, environment, permissions, and tool access.
  • Lead & specialist delegation: Leads delegate subtasks to specialists, review evidence, and accept or reject child results.
  • @team:{name} invocation: Call an entire team from any task with standard syntax.
  • Strict safety budgets: Hard caps on model calls, token count, execution time, and a universal Stop signal that reaches every child subagent.
  • Bounded parallelism & recovery: Recovery rules that never blindly replay an action whose outcome was uncertain.

3. Workflow Automation & Multi-Stage Pipelines

  • Pipeline builder: Build multi-stage workflows manually or generate draft workflows from natural language prompts using live Nebius models.
  • Isolated code worktrees: Hand code and repository state between stages using isolated Git worktrees without workspace pollution.
  • Conditional branching & retries: Branch on structured JSON outputs, enforce output format contracts, and configure bounded failure recovery.
  • Human approval gates: Place human review gates between stages before passing artifacts forward.

4. Human-in-the-Loop Decisions Hub

  • Cross-project approval queue: A centralized Decisions view collects every pending command, file edit, and connected-app action across all active projects.
  • Interactive approval controls: One-click Approve, Decline, Open task, and Open run.
  • Diff & command inspection: Inspect full syntax-highlighted diffs and shell commands before execution.

5. Integrated & Durable Task Management

  • Native persistence: Powered by SQLite; tasks retain their full transcript, tool activity, diffs, and artifacts across app restarts.
  • Ultra-responsive virtualization: Smooth virtualized list scrolling with instant text search, tested on long transcripts of 1,000+ messages.
  • Forking & checkpoints: Session rewind, branch points, and attachment support (images, documents).

6. Composio Integration for 500+ Connected Apps

  • Zero implicit trust: Connecting an account (Gmail, Slack, GitHub, Linear, etc.) grants zero authority to execute actions automatically.
  • Granular per-agent grants: Explicitly assign specific actions to individual agents.
  • Exact action receipts: Every external API invocation presents its account, version, and arguments for approval, returning an immutable execution receipt.

7. Document & Artifact Management

  • Multi-format library: Native handling of HTML, Markdown, Plain Text, CSV, JSON, SVG, XML, YAML, and PDF.
  • Safe large-document rendering: Built-in in-memory private URL serving for large HTML artifacts to bypass OS webview limits.
  • Full-fidelity inspection: Reading view, syntax-highlighted source view, search, filtering, and export of original raw bytes.
  • Digest & provenance tracking: Every artifact links directly back to the task, model, and prompt that generated it.

9. Secure Credential Management

  • OS Credential Vault integration: Keys are stored directly in your OS-native secure enclave (never logged in plaintext or exposed to LLMs).
  • One-click connection testing: Validate endpoints and retrieve authorized model catalogs before saving.
  • Masked inputs & lifecycle control: Easy key rotation, replacement, and permission management.

10. Local Speech & Voice Dictation

  • On-device speech recognition: Integrated with NeMo-Speech.cpp running NVIDIA Nemotron ASR models completely offline and locally.
  • Multi-language support: Auto-detect, English, Hindi, and more.
  • Non-intrusive insertion: Audio transcribes directly into an editable composer and is never auto-sent without review.

11. Routines & Scheduled Automation

  • Deterministic schedules: Run workflows on recurring schedules with durable firing identities to prevent duplicate triggers.
  • Execution controls: Pause, resume, trigger immediate manual runs, preview next-fire timestamps, and view execution history.

12. Mobile Companion & Personal Gateway

  • Mobile app (Expo / React Native): Monitor running tasks, review artifacts, and answer approvals on the go.
  • Owner-hosted Personal Gateway: A lightweight, Bun-based gateway using scoped tokens to pair mobile clients securely with your desktop daemon.

What's Next (Incoming Roadmap)

  • [ ] macOS and Linux builds: Expand from verified Windows builds to official native macOS (Apple Silicon / Intel) and Linux packages.
  • [ ] Signed production installers: Provide signed Windows MSI/EXE installers and macOS notarized packages.
  • [ ] Remote agent execution: Offload heavy builds and long-running agent workflows to registered remote worker targets and Nebius cloud instances with local fallback.
  • [ ] Enterprise billing & team budgets: Cloud billing export imports, shared team budgets, and decimal-exact cost allocation across shared projects.
  • [ ] Hosted speech fallback: Provide an optional, trusted hosted speech endpoint for systems lacking GPU acceleration for local NeMo-Speech.
  • [ ] Wider harness support: Add bounded team adapters for Amp, Cursor, and ordinary-task delegation for Fx.
  • [ ] Native provider goals: Incorporate native autonomous goal state for model providers that expose high-level planning primitives.

Getting Started

  1. Download & Install: Install Ion on your workstation.
  2. Add Your Key: Open Settings > Credentials, paste your Nebius Token Factory API key, and click Test connection (your model catalog will populate immediately).
  3. Open a Project: Click Open Folder to select any workspace or repository.
  4. Run a Task or Team: Create a task, select a Nemotron or DeepSeek model, or summon an agent team with @team:{name} to begin.

Built With

Share this project:

Updates

Submission history