Avi

A local desktop workspace for AI conversations, tools, and orchestration.

Website · Source code

Avi is an harness built from scratch which brings model conversations, cross-provider communication, project context, local tools, MCP servers, and multi-agent workflows into one desktop application. Conversation state is stored locally, while model requests are sent only to the providers you configure.

Features

• Light: small footprint compared to other harnesses
  • Low RAM and CPU consumption with multiple agents working
  • Fast startup, launches with the system
  • Few dependencies, easy to maintain
• Multiple providers: connect multiple AI providers and customize the models you’ll use for each provider.
  • OpenAI Subscription: your ChatGPT subscription – no need for Codex ACP
  • OpenAI-Compatible endpoints: /v1/responses and /v1/chat/completions
  • Model-specific settings: capabilities, reasoning supported
• Powerful sub-agents: sub-agents can actively communicate with the orchestrator and other sub-agents.
  • Define sub-agent levels: model + reasoning per sub-agent invocation level (low, medium, high), lets you choose which models and providers will run different task types
  • Active communication: sub-agents can send messages to the orchestrator and other sub-agents while working, and vice versa.
  • Sub-agent view panel: track sub-agent progress via the side panel.
• Powerful orchestration: chats have advanced reflection and orchestration tools.
  • Start, inspect, and converse with parallel threads: agents can view conversations, work folders, tasks, monitor and supervise other agents.
  • Remote MCP: persistent server that provides orchestration tools to connect to external services (Claude, ChatGPT, etc.)
  • Orchestration panel: view ongoing tasks, newly completed tasks, consumption insights
• MCP client: MCP client scoped per project
  • MCP control panel: view MCP tools, provided instructions
  • Isolation: separate MCP servers by folder or globally
  • Diagnostics: visually check servers that failed or are slow to start
• Context management and discovery: advanced discovery of skills, workflows, and instructions
  • Recursive context listing: searches for skills and workflows in the current folder and globally (in $HOME/.agents) without the agent having to search
  • Automatic contextualization: injects AGENTS.md, MEMORY.md, AGENTS.foobar.md... automatically into the agent’s context.
  • Slash commands: invoke workflows with /command and skills via $skill in the composer.
  • Context panel: manage skills, workflows and instructions findable by the agent.
• Advanced inference:
  • Very large tool results are truncated and written to files
  • Agent can query large tool outputs with tools.
  • Native tools for file reading, media reading (images, PDFs) and file writing (encoding‐sensitive).
  • Automatic retry for server and provider errors.
  • Queue and steer mechanism for advanced chat.
  • Execution permission level for potentially dangerous tools (ask for approval, allow for me, full access).
  • Automatic context compression on provider errors (context_length_exceeded) or when reaching user‐defined threshold.
  • Resume button on stopped or failed chats, which continue from the last assistant turn.
• Goals and targets:
  • Goals can be started with /goal or by the agent itself.
  • Helper model expands the goal with completion criteria, execution rules, and relevant meta‐information.
  • Model loops until the condition is met.
  • Infinite inference retry with long timeout for provider errors, rate‐limits, or server not interrupting the goal.
• Ultra mode:
  • Aggressive delegation mode of sub‐agents to different solution‐exploration fronts.
  • Has a rigid workflow of recognition, judgment, and refinement of the work done.
  • Can consume many more tokens.
  • Can be used together with goals.
• Planning mode:
  • Agent uses sub‐agents to create an execution plan for a task.
  • Delegates sub‐agents for exploration, research, and independent checks to refine the plan.
  • Instructs sub‐agents to talk actively with each other to reach a consensus.
• Quick chat and side chats:
  • Side chats: fork the current conversation into a quick side chat to ask about the agent’s work without interrupting the main chat. Side chats can direct, orchestrate, and assist the main agent and its sub‐agents.
  • Quick chats: minimalist quick chat for fast questions unrelated to any thread or folder.
• Side panel:
  • View files, git changes in the side panel
  • View tasks started by the agent during its threads
  • View provider limits and consumption (OpenAI Subscription only)
• Customizable:
  • Choose different personalities for the chat (friendly, candid, cynical, etc.)
  • Choose interface themes

Getting started

Prerequisites

• Bun installed and available in your terminal
• Git
• A supported Windows, macOS, or Linux desktop environment

Install and run

git clone https://github.com/aivaxlabs/avi.git
cd avi
bun install
bun run dev

To open the renderer developer tools:

bun run dev:devtools

During development only, pass --skip-single-instance when parallel Avi instances are required:

bun run dev --skip-single-instance

No environment variables are required for normal development. Providers and MCP servers are configured inside the application.

Provider setup

Open Settings → Providers, then choose one of the supported connection types:

OpenAI Subscription

Connect a ChatGPT account through the browser-based OAuth flow. Supported models are managed by Avi and become available after authorization.

OpenAI Compatible

Configure a provider that implements either:

• POST /v1/responses
• POST /v1/chat/completions

Provide the base URL, API key, and models exposed by the service. Model capabilities and reasoning behavior can be adjusted per model.

Contributing

1. Fork the repository.
2. Create a focused branch.
3. Install dependencies with bun install.
4. Make the smallest coherent change.
5. Run the relevant tests, syntax check, and build.
6. Open a pull request describing the behavior and validation performed.

Prefer concise changes that preserve existing behavior and keep provider-specific logic inside src/providers/.

Credits

• Created by AIVAX Labs
• File type icons: Microsoft Visual Studio Image Library