Novi lets developers run Claude Code and Codex on Novita models

What Novi does

Novi works as a wrapper around the harness a developer already uses. After installing it and adding a Novita API key, the user types novi claude instead of claude. Novita lists seven generally available harnesses: Claude Code, Codex, OpenCode, Pi, DeepSeek Harness, Command Code and Kilo Code.

The company says pointing a harness at a new provider is mostly configuration work. Claude Code alone expects a base URL, auth token, model slots, a subagent model, timeout, output cap, compaction window and context size. If one is missed, the session may still start but degrade quietly, with truncated output, broken tool calls or a context window that fills early. According to Novita, Novi sets 34 of these settings for every Claude Code session and removes them when the harness exits.

For Claude Code, Novi maps a main, a mid and a background model. The defaults are Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash. Subagents inherit the session model, and web search runs through Novita AI Search.

Setup, models and the local relay

Novi ships as a single, checksum-verified binary for macOS, Linux and Windows. macOS and Linux users install it with a one-line curl command, while Windows 10 and 11 users run a PowerShell installer. Users then run novi login and enter a Novita API key, or set the NOVITA_API_KEY environment variable. Browser sign-in is on the roadmap but not available today.

Each session starts a relay on 127.0.0.1 at a random port. The harness receives a one-time session token instead of the Novita key, and the relay forwards requests with the real key. According to Novita, this keeps the key out of shell environments, config files and harness logs. The company stresses that this is credential isolation, not local inference: prompts and responses still travel to the Novita API.

Compatible models appear in each harness's own /model picker with price and context window. Novita says Claude Code's picker lists 40+ Novita models and Codex 20+. Not every model works with every harness, and novi models shows which harnesses can run each one. A different model can be chosen by passing its ID, for example novi claude --model zai-org/glm-5.3.

The default lineup, with pricing and context taken from Novita model pages as checked on October 21, 2026:

• Main: Kimi K3, 1M context, $3.00 input / $15.00 output per 1M tokens
• Mid: GLM 5.3, 1M context, $1.40 / $4.40
• Background: DeepSeek V4.1 Flash, 1M context, $0.30 / $1.20

Cost, status and what it means for developers

Novi itself is free. Tokens are billed at standard Novita pay-as-you-go prices, and Novita AI Search is billed separately and can be turned off with NOVI_NO_WEB_SEARCH=1. Each session prints a receipt and novi usage --last 7d sums the week, though Novita notes both are local estimates and not the actual bill.

For CLI sessions, settings are injected at launch and removed on exit, so no shell profile or provider block is left modified. Claude Desktop and Codex Desktop are in beta and use a journaled provider profile that can be restored, which Novita says is not a no-configuration mode. Hermes and Cline are listed as coming soon. Browser sign-in, auto routing, IDE integrations, team budgets and Agent Sandbox integration are planned, not shipped.

For developers, the practical effect is a way to try open-weight models from Novita inside the tools they already know, without rebuilding their setup by hand. Maintenance commands include novi doctor to check the key, API and harnesses, novi update to replace the binary after a checksum, and novi uninstall, which removes Novi and its saved key but leaves harness configs and history.