Rizo

An AI coding assistant for VS Code — chat, file edits, and terminal/git commands, with cost-optimized, company-locked model selection. BYOK: bring your own OpenRouter API key, billed directly by OpenRouter for what you actually use.

Features

Chat

• Pick a provider, once per chat — Claude, Gemini, OpenAI, DeepSeek, Kimi, or Free. That choice locks in for the chat's whole life; the in-chat switcher only ever offers that same company's own variants (e.g. Claude's Haiku 4.5 / Sonnet 5 / Opus 5), never a different provider — one tool-calling convention and context window for the life of the conversation, want a different company, start a new chat
• Effort dial — a second pill next to the model switcher (Low / Medium / High, adjustable any time) controls how hard the current model thinks, independent of which model's answering — maps to OpenRouter's unified reasoning.effort field
• Usage you can see — a running today's-tokens / this-month's-cost readout at the top of the panel, aggregated across every chat and every provider, resetting itself automatically at midnight and on the 1st; per-chat totals sit under the composer too
• Live turn status — while a turn's running, see which variant is actually answering, what it's doing right now, elapsed time, and a running token count, not just a static "Thinking..." bubble
• Streaming replies with a live tool-call transcript — see what's actually happening (which file, which command) as it happens
• Markdown rendering — code blocks, lists, bold/italic render properly in replies
• Stop button — cancel an in-flight reply any time
• File & image attachments (pick from disk, mention a workspace file, or paste an image straight from the clipboard), multiple named threads (renameable, deletable with a confirm dialog — and undoable, via an "Undo" toast or the Reopen Closed Session command), and automatic summarization once a chat gets long
• Message queueing — type a follow-up while Rizo's still working; it sends once the current reply finishes instead of the input locking

Smart routing

• Smart Starting Variant — a chat's first message can start on a stronger variant than the cheapest default when it actually calls for it (a real bug report, a big refactor ask), based on a cheap heuristic, not an extra model call — and never overrides a variant you picked yourself
• Effort auto-suggestion — the same heuristic also picks Effort per turn, backing off permanently the moment you set it yourself
• Auto-escalation — a turn that visibly struggles (hits its step limit, keeps failing the same tool call, or stalls across turns without making progress) offers a one-click "Retry with a stronger variant" — still within the same company, only when a stronger one exists
• Free, your way — Auto by default, or pick one of 5 specific free models yourself (from whichever companies actually publish one — no free Claude/Gemini exists) with its own Effort control where it supports one

Safety & control

• File edits with approval — every write shows a native diff view before anything touches disk; nothing happens without your click
• Terminal & git tools — with an elevated, non-bypassable warning on destructive commands (force-push, hard reset, branch deletion, rm -rf) or ones that hide what they actually run (piping a remote download into a shell, bash -c "...", eval)
• Security pattern-scan — a proposed change is checked against dozens of known-dangerous patterns (eval, hardcoded secrets, disabled TLS checks, unsafe deserialization, etc.) and flagged right in the approval dialog, before you click
• Command-output redaction — API keys, tokens, and private keys that happen to show up in a command's output (an .env dump, env, aws configure list) are redacted before they ever reach the model or get stored in thread history
• Configurable permissions — auto-approve specific commands by pattern, or disable a tool entirely, via VS Code settings
• Work outside the open folder, on purpose — every folder in a multi-root workspace is usable, and naming a real path in your own message (never the model's own idea) grants access to it for the rest of that chat — same trust boundary file attachments already use
• Truncated writes fail informatively — a large file cut off by the response size limit gets told to split into smaller calls, not a generic error that just triggers an identical, costly retry

Shortcuts

• Slash commands — /commit, /review, /test expand into a full guided prompt tied to house engineering discipline
• Project instructions — drop a .rizo/instructions.md in your workspace for project-specific rules, no fork required

Getting started

See ONBOARDING.md for the full setup walkthrough (get an OpenRouter key, first run, how approvals work).

Short version: install → Cmd+Shift+P / Ctrl+Shift+P → "Rizo: Open Chat" → paste your OpenRouter key when prompted → pick a provider for the chat.

License

Apache License 2.0, modified by the Commons Clause — see LICENSE. Free to use, modify, and share; the Commons Clause withholds the right to sell it. Source-available, not OSI-certified "open source."