CPU3DAI tools for 3D, video and audio

Capri-AI: a local desktop agent with a cyberpunk-style HUD for Windows

On November 14, 2026, Capri-AI appeared on GitHub — a local agent environment for Windows with a Cyberpunk: Edgerunners-style interface. In six days the repository gained 345 stars. The authors describe the project as a "personal Jarvis for Windows": the agent manages files, applications, the terminal and the browser, and Claude Code, Codex, Antigravity or anything else can be used as the reasoning backend.

What it does

  • Custom tools and functions — the Plugin and tool API lets you register your own tools via a simple JSON schema, and the agent can call them.
  • MCP support — external tool servers connect via the Model Context Protocol, expanding the agent's arsenal without changing the core.
  • Agent memory — long-term context storage so Capri remembers projects, preferences and recurring workflows between sessions.
  • Human approval — destructive or irreversible actions go through a configurable policy layer and require explicit approval.
  • Tracing and debugging — every plan, tool call and result is written to a structured session log available for review.

Your own model instead of the cloud

The documentation does not name a specific way to connect a local model, but the README states directly: "The desktop shell, tools and memory work offline. Reasoning requires a connected backend unless you point Capri at a local model endpoint." That is, the shell, tools and memory work offline, while reasoning requires a connected backend — either cloud or a local model endpoint, if you point Capri at one. The implementation language is not listed in the fact sheet, and the installation method is not described. License — MIT.

How it differs from its neighbors

A feature comparison shows that Capri-AI has no capability that is claimed only for it and absent from Hermes Agent, DeerFlow and FrontierAgent. At the same time, all three neighbors claim multiple agents working together, sandboxed code execution, graphs and state persistence, as well as their own server and Docker — the Capri-AI documentation does not mention these items. Custom tools and MCP support are claimed by all four platforms. Capri-AI stands out more by positioning: it is a desktop shell with a cyberpunk HUD for Windows, not a server framework. Those who need a local personal agent with a visual interface may find this option closer to their needs; those building multi-agent systems with sandboxes and Docker deployment should look at the neighbors.

How alive the project is

The repository was created on November 14, 2026; as of November 21 it has 345 stars and one contributor. Over 90 days there is one stable release — 1.0.0 from November 14, 2026. The project is very young, and activity so far is limited to the first release and a single contributor.

Capri-AIHermes AgentDeerFlowFrontierAgent
Typeenvironmentenvironmentenvironmentenvironment
Language—PythonPythonPython
LicenseMITMITMITApache-2.0
GitHub stars345263 97285 5385 364
Releases in 90 days1212—
Multiple agents working together—yesyesyes
Custom tools and functionsyesyesyesyes
MCP supportyesyesyesyes
Agent memoryyesyesyes—
Knowledge base (RAG)——yes—
Sandboxed code execution——yesyes
Browser interaction——yes—
Human approvalyesyes—yes
Graphs and state persistence——yesyes
Tracing and debuggingyes—yesyes
Agent quality evaluation———yes
Voice and speech—yes——
Own server and Docker——yesyes
Images and documents——yes—
Own model: Ollama—yes——
Own model: llama.cpp—yes——
Own model: vLLM—yes——
Own model: SGLang———yes
Own model: OpenAI-compatible API——yesyes

Comparison based on the projects' READMEs and documentation as of 11/21/2026: "yes" means the capability is claimed there, a dash means it is not mentioned (which does not mean the capability is absent). Stars and releases are per GitHub data.

Capri-AI will suit those looking for a local desktop agent for Windows with a ready-made shell, system control and the ability to connect their own model endpoint. For new breakdowns of agent platforms and neural networks on your own hardware, check the CPU3D news feed.

See also