yana-rt

yana-rt command [options]yana-rt -h | --helpyana-rt -V | --version

yana-rt is the Rust runtime for Yana AI. It owns the turn loop and the authority check used by the CLI, Yana Studio, and optional MCP, so those clients do not each reimplement permission logic. Install from crates.io with `cargo install yana-rt` (crate and binary name yana-rt, default features include cli). Version 1.5.1 as of the 2026-10-05 crates.io release; check `yana-rt --version` for the installed build.The same source is also built as yana-ai-rt, which inserts the chat subcommand from `argv[0]`. A `pip install yana-ai` entry point named yana-rt is a Python shim: it searches YANA_RT_BIN, then PATH (skipping its own script), then a packaged native binary, then `target/release/yana-rt`. Point YANA_RT_BIN at a compiled ELF/Mach-O, never at `which yana-rt` after a pip install.Providers in the runtime catalog include cloud (Anthropic, OpenAI, Gemini, Groq, DeepSeek, OpenRouter, xAI, Novita, NVIDIA, MiniMax, GLM, Hugging Face, 9Router, Kimi) and local (Ollama, LM Studio, llama.cpp, TurboFieldfare, AirLLM), plus a custom OpenAI-compatible endpoint. Failover, a SQLite session store, shadow-git checkpoints, file read/patch, gated command execution, optional MCP/LSP, and `web.search` over a checked HTTPS layer live here.Several subcommands are not routed through yana-ai on purpose: trust (a person at a terminal must confirm repo-supplied config, like direnv allow), compact (hook fast path for Bash output), capability (one-shot tools for the desktop app), and guard (in-process PreToolUse checks). `yana-ai guard` is a different command (Python hook installer).Optional Cargo features: mcp, lsp, pty-bridge, discord. They are off in the default build. Apache-2.0. No Python runtime required for this binary.

-h, --help

Print clap help for the binary or a subcommand

Print the crate version

Machine-readable output on scan, doctor, task, map, score, hunt, and others

Chat/eval provider from the runtime catalog (`src/model/catalog.rs`)

Override the provider default. For Ollama, the runtime tries `/api/tags` before a static fallback

Exit non-zero on findings at this severity or above

With chat, run approved run_command calls without `sandbox-exec.sh`

.yana-ai/

Per-project runtime state: config, leases, cost policy, trusted integration files, session database

Finding IDs scan should suppress

scan [target] [--json] [--markdown file] [--sarif file] [--fail-on level] [--only category] [--ignore id] [--diff base] [--no-color] [--quiet] [--scanner-dir dir] [--include-skills]

Audit the AI-agent setup. Default target is `.`. yana-ai audit calls this. --include-skills also scans `core/skills/**` (off by default; high false-positive rate)

Interactive REPL. read_file and run_command are available; run_command always asks a human first. --no-sandbox skips `sandbox-exec.sh` but still requires approval. Default Ollama endpoint is `127.0.0.1:11434`

Health checks (Python, git, ignore files, Claude/MCP config, keys, hooks). doctor dispatch compares clap commands with `bin/yana`

Task lifecycle. create takes a name and optional --scope. done requires --evidence. depend edges: `blocks` (default), `related`, `parent-child`, `discovered-from`

Evidence checks. judge takes --provider / --model (default Ollama)

Confirm `.yana-ai/web-search.json`, `mcp-servers.json`, or `lsp-servers.json`. allow needs a TTY and a typed `yes`. Not exposed through yana-ai

Knowledge graph. search has --expand and --limit (default 15). diff default base is `origin/main`

Categories: `secrets`, `code`, `deps`, `supply-chain`, `all` (default)

What the agent can reach (Claude settings, MCP, workflows)

Numeric score and risk band (LOW / MEDIUM / HIGH / CRITICAL)

Flag-driven project init. init dry prints what would be written. Prefer yana-ai init for the interactive wizard

In-process PreToolUse ports (destructive-command guard, token budget). Called as `exec yana-rt guard` from hook scripts

Opt-in Bash-output compaction (`YANA_COMPACT=1`). --detect classifies without running. Hook path; no yana-ai compact

One-shot capability tools for non-Rust clients (desktop). Not routed through yana-ai

VirusTotal hash lookup. Needs VT_API_KEY. Does not upload file contents

Further runtime surfaces; yana-ai forwards matching names. mcp and remote discord need optional features

YANA_RT_BIN

Used by the Python/npm shims to locate this binary. Must be a compiled executable

Set by the Python shim to stop recursive re-entry. Do not set by hand

Cloud chat providers

Enable or force-bypass Bash-output compaction

filescan check

Discord adapter (--features discord), allowlisted plain chat only

The pip entry point yana-rt is a wrapper, not this binary. A stale runtime on PATH rejects new subcommands; upgrade with `cargo install yana-rt --force`. trust allow is interactive on purpose so an agent driving yana-ai cannot confirm config. chat is a separate process and is invisible to Claude Code hooks; it builds its own approval gate. Destructive guards match command strings. Optional mcp is a prototype stdio server, not wired into clients by default. There is no Linux ARM Studio build; this CLI still builds with Cargo on supported Rust hosts.

The runtime crate yana-rt is the native engine of Yana AI, started 17 May 2026 by Vũ Văn Tâm. It is versioned independently of the yana-ai Python package. crates.io 1.5.1 (2026-10-05) added provider failover, the SQLite session store, and capability tools.

yana-ai(1), cargo(1), claude(1), codex(1), aider(1), aichat(1), ollama(1), direnv(1), git(1)

Source code · Homepage · Documentation