Novi is a free command-line tool that launches the harness you already use—Claude Code, Codex, OpenCode, Pi, DeepSeek Harness, Command Code, or Kilo Code—on Novita models with one command. Install it once, add a Novita API key, and start novi claude instead of claude. Novi wires the provider settings, maps each model to the right tier, and adapts requests for the session; your harness, extensions, MCP servers, and workflow stay the same.
The setup work Novi removes
Pointing a harness at a new provider is mostly configuration work before it is coding work. Claude Code alone expects a base URL, auth token, default and small-model slots, subagent model, timeout, output cap, compaction window, and context size. Miss one and the session may still start, but degrade quietly: truncated output, broken tool calls, or a context window that fills early.
Novi sets 34 of those settings for every Claude Code session and removes them when the harness exits. It maps the main, mid, and background slots to Novita models, then adapts output limits, tool calls, and thinking behavior to the model in use. The other supported harnesses follow the same pattern with their own configuration.
The manual Codex and OpenCode guides show what Novi replaces.
Install Novi and launch your first session
Novi ships as a single, checksum-verified binary for macOS, Linux, and Windows. On macOS or Linux, install it with one line:
curl -fsSL https://novita.ai/novi/install.sh | shOn Windows 10 or 11, run the PowerShell installer:
irm https://novita.ai/novi/install.ps1 | iexThen sign in with novi login. It asks for a Novita API key, which you can create in Key Management. If you prefer environment variables, set NOVITA_API_KEY instead. Browser sign-in is on the roadmap, not available today.
Start a harness by putting novi in front of its usual command:
novi claude
novi codex
novi opencode
novi pi
novi dshnovi dsh launches DeepSeek Harness. If a harness is not installed yet, run novi with no arguments to open the interactive menu. It shows which harnesses are installed and up to date, offers to install missing ones, and opens the model catalog, API key management, and billing.
A 33-second look at the workflow
The launch film shows the workflow Novi is built for: several harnesses side by side, each on a Novita model, with a relay and no API key inside any harness. There is no voiceover.
The film opens on the line “Keep your harness” and seven harness windows—Claude Code, Codex, OpenCode, Pi, DeepSeek Harness, Command Code, and Kilo Code—running different Novita models. It closes on a diagram where Kimi K3, GLM 5.3, DeepSeek V4.1 Flash, MiniMax M3, and Qwen3.8 Max connect through Novi, and on the claim “0 API keys handed to your harness.”
The clip is an illustrative demo, not a benchmark or a test run. Command Code and Kilo Code are supported. The setup commands above are all you need, so watching is optional.
What Novi handles behind the launch command
Configuration and tier mapping
Claude Code expects different models for different jobs. Novi maps a main coding model, a mid model, and a background model for titles and quick calls. 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.
Settings are injected at launch and removed on exit, so a CLI session leaves no modified shell profile or provider block; mid-session saves are restored.
Model choice inside each harness
Every compatible Novita model appears in the harness’s own /model picker with its price and context window. Claude Code’s picker lists 40+ Novita models and Codex 20+. Compatibility is not universal: a model must support the tools and request shape a harness sends, and novi models lists which harnesses can run each model.
A local relay, not an API key in the harness
Each session starts a relay on 127.0.0.1 at a random port. The harness receives a one-time session token rather than your Novita key; the relay forwards requests with your key and refuses non-loopback addresses and browser origins. That keeps the real key out of shell environments, config files, and harness logs.
This is credential isolation, not data locality. Prompts and responses still travel to the Novita API. The relay does not make a cloud model run locally.
Desktop apps are a different path
Claude Desktop and Codex Desktop are in beta. They get a long-term, journaled provider profile that points at the relay and can be restored; this is not a no-configuration mode.
Which harnesses and models are supported
| Status | Harnesses |
|---|---|
| Generally available | Claude Code, Codex, OpenCode, Pi, DeepSeek Harness, Command Code, Kilo Code |
| Beta | Claude Desktop, Codex Desktop |
| Coming soon | Hermes, Cline |
Browser sign-in, auto routing, IDE integrations, team budgets, and Agent Sandbox integration are planned, not shipped.
The default lineup uses three models for different jobs:
| Role | Model | Context | Input / output per 1M tokens |
|---|---|---|---|
| Main | Kimi K3 | 1M | $3.00 / $15.00 |
| Mid | GLM 5.3 | 1M | $1.40 / $4.40 |
| Background | DeepSeek V4.1 Flash | 1M | $0.30 / $1.20 |
Pricing and context from the Novita AI model pages, checked October 21, 2026. novi models and the model catalog show the full set, including MiniMax M3 and Qwen3.8 Max.
To start with a different model, pass its Novita ID:
novi claude --model zai-org/glm-5.3Use an ID from novi models; the model must be compatible with that harness.
Cost and maintenance
Novi is free. Tokens are billed at standard Novita pay-as-you-go prices, and Novita AI Search is billed separately; turn search off with NOVI_NO_WEB_SEARCH=1. Each session prints a receipt, and novi usage --last 7d sums the week. Both are local estimates, not your Novita bill.
For maintenance, novi models browses the catalog, novi doctor checks the key, API, and harnesses, and novi update replaces the binary after a checksum. novi uninstall removes Novi and its saved key but leaves harness configs and history.
Conclusion
Keep your harness; run it on Novita. Novi removes the provider configuration work without moving you to a new editor or workflow. Install it, connect your Novita account, and launch your next session with the harness you already use. Get started with Novi.
FAQ
Is Novi free?
Novi is free to install and use. You pay standard Novita pay-as-you-go prices for tokens, and web search runs through Novita AI Search and is billed separately. Check Novita AI pricing for current rates.
Which harnesses does Novi support?
Seven are generally available: Claude Code, Codex, OpenCode, Pi, DeepSeek Harness, Command Code, and Kilo Code. Claude Desktop and Codex Desktop are in beta. Hermes and Cline are coming soon. The Novi page tracks the current list and launch commands.
Does Novi change my existing configuration?
For CLI sessions, no. Settings and provider files are injected at launch and removed when the harness exits. Desktop apps are different: in beta they write a journaled configuration that can be restored from the Novi menu.
Does my harness receive my Novita API key?
No. The harness talks to a loopback relay and receives a per-session token; only the relay forwards your Novita key to the Novita API. The relay is local, but model requests still leave your machine for Novita, so this is credential isolation rather than local inference.
How do I choose a different model?
Run novi models to see compatible models, then pass one to the launcher, for example novi claude --model zai-org/glm-5.3 for GLM 5.3. Inside a running Claude Code session, /model lists the models that harness can drive, with prices and context windows.
