What is Yana AI?
Yana AI is a local control runtime that sits between an AI model and tools that can read or change files, run commands, or access other capabilities. It checks each requested action against configured permissions and policies, and can pause for human approval before execution. The project includes a Rust runtime, a Python CLI, governance rules and hooks, with adapters for Claude Code, Codex, Cursor and Antigravity. It is intended for developers who want AI agents to work within defined boundaries and leave an inspectable record.
Yana AI Features:
Permission Checks: The runtime checks an agent’s authority, requested capability and applicable policy before allowing an action to reach an executor.
Human Approval: Actions that require a person can pause for approval, while users can block requests that should not proceed.
Scoped Capabilities: Capability manifests define which tools and powers an agent can request, keeping permissions separate from the model’s responses.
Audit Trail: Decisions are recorded in a hash-chained log, making later edits to the record detectable.
Tool Execution: Available capabilities include file reading and patching, validated command execution, an MCP client, checked web search and a read-only LSP client.
Model Providers: The runtime connects to listed cloud providers and local options including Ollama, LM Studio and llama.cpp through a shared control path.
Host Adapters: Adapters connect Yana to Claude Code, Codex, Cursor and Antigravity, with enforcement strength varying by host.
Yana AI Benefits:
Bounded Agent Access: Keeping permissions outside the model means changing models does not automatically change which actions an agent is allowed to perform.
Human Oversight: Approval pauses give users a chance to review sensitive actions before they affect files or run commands.
Reviewable Decisions: A record of requests, decisions and execution results helps teams investigate what an agent did and why.
Consistent Controls: A shared runtime gives supported clients a common place to check permissions instead of relying only on agent instructions.
Local Model Choice: Support for local and cloud models lets users choose where model processing happens without discarding Yana’s permission layer.
Use Cases:
Code Changes: Let an agent inspect a project and propose file edits, while requiring approval before writes or shell commands.
Command Review: Check an agent’s requested terminal operation against configured rules before allowing it to run.
Agent Rollout: Add governance hooks to a supported coding host when introducing AI agents into an existing development workflow.
Local AI Work: Connect a local model to file and command capabilities while keeping permission checks in the runtime.
Action Audits: Review the recorded decisions and results after an agent has worked through a project task.

