How to Use RunInfra: The Beginner's Guide (First Setup, the One-Prompt Connect, and the Flat-Rate Arc)

CodexFuse Team · 2026-11-03T03:02:11.952Z · 13 min read min read

RunInfra puts fast open models (GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8) inside Claude Code, Codex and more for a flat $10/mo: one-prompt setup, Coding Plan limits, model picker, and the cache that makes the plan go further.

How to Use RunInfra: The Beginner's Guide (First Setup, the One-Prompt Connect, and the Flat-Rate Arc)

Learning how to use RunInfra starts with one sentence: it is a GPU-native inference platform whose Coding Plan puts fast open models inside the coding agent you already use for a flat $10 a month. RunInfra runs the models on its own GPUs and routes your agent's requests to them — you keep Claude Code or Codex as your interface and swap the expensive frontier model for a cheap, fast open one. Because repeated repo context mostly hits RunInfra's session cache — cached input starts at $0.01 per million tokens on select models — the same codebase context costs almost nothing the second time around (every figure in this paragraph is RunInfra's own published claim, checked today).

One important note before we start: the RunInfra brand covers two products. The original, launched in July 2026, is an inference-optimization platform — pipeline builder, vLLM and TensorRT engines, and the runinfra pull command you will still see on PyPI and in directory listings. This guide is about the new one: Coding Plans, launched October 20, 2026, which put fast open models inside coding agents. That launch is two weeks old, so this is a genuine first-wave beginner's guide — page one of the search results mixes the old product's docs with the new product's launch pages, and no independent walkthrough of the current setup flow exists yet.

If you run models on your own machine instead, our Ollama beginner guide covers that lane — Ollama runs models locally, RunInfra is flat-rate cloud models inside your agents. If you build agent-driven apps, our Dify beginner guide covers the agent-app side; RunInfra is the model layer underneath. And once your API key is live, you can poke the endpoints with a REST API tester as you follow along. This RunInfra tutorial walks the full arc in order: the mental model, first setup, the one-prompt connect, the limits system, and the model picker.

What Is RunInfra, Exactly?

RunInfra is a GPU-native inference platform: it runs open models on its own GPUs and sells you access two ways — pay-as-you-go credits, or a flat-rate Coding Plan that meters your agent's model usage. One credit is one US dollar; you get $1 free on signup, once per account, never expiring, with top-ups from $5 to $1,000,000 in whole dollars under Settings, then Billing. Most beginners start with the Coding Plan, because it is the product built specifically for coding agents.

A Coding Plan gives your coding agent one API key that works with both the OpenAI and Anthropic SDKs at api.runinfra.ai/v1, and session-aware routing keeps repeated context cheap: RunInfra says cached input starts at $0.01 per million tokens on select models, and it publishes last-24-hour cache-hit figures up to 99.9% on its homepage — the numbers, and their caveats, get their own section below. The company also states zero data retention, no training on your prompts, and SOC 2 Type II on its Product Hunt launch page; those are vendor claims, stated as claimed, not independently audited here.

What Do You Need Before You Set Up RunInfra?

Three things: a RunInfra account (signup is free and drops that $1 credit in), Node.js 22 or newer if you use the npx path, and a coding agent from the supported list. That is the entire prerequisite list — no credit card detail is published as required for the free credit.

The documented setup path is npx -y @runinfra/cli@latest (Node 22+), with a global install (npm install -g @runinfra/cli@latest) as the alternative; the docs also list a Python path (pip install --upgrade runinfra-cli or pipx install runinfra-cli) requiring CLI 0.5.0 or newer. One trap to know about: PyPI's runinfra-cli 0.2.1 is the original inference product's CLI — if the version you get does not match the doc's 0.5.0+, use the npx path, which is the one the agent-setup doc leads with.

How Do You Connect Your Coding Agent in One Prompt?

You let the agent wire itself: paste one prompt, approve a browser sign-in, say yes to the setup plan it shows you. This is the flow RunInfra leads with on its homepage, and it is the fastest honest path to a working connection.

1. Paste the prompt below into your coding agent's chat.
2. Approve the sign-in in your browser when the agent asks.
3. Read the setup plan the agent proposes — before it installs anything, changes settings, or spends money — and say yes.

The prompt, verbatim from RunInfra's homepage:

Set up RunInfra's open models in this coding agent for me. Read https://runinfra.ai/docs/tools-sdks/agent-setup.md and follow it step by step. If you cannot open links, run `npx -y @runinfra/cli@latest --help` and follow its FOR CODING AGENTS section. I approve sign-in in my browser. Before you install anything, change this agent's settings or spend money, show me the plan and wait for my yes.

Agent-led setup works with Claude Code, Codex, OpenCode, Pi, Aider, Qwen Code, Kilo Code, Goose, Continue, Cline, Droid, Zed, Grok Build, Hermes, OpenClaw, and Crush — sixteen agents in the doc's supported-agents table, with RightNow Agent getting its own guide. If your agent is Cursor, Warp, Trae, JetBrains AI Assistant, Twinny, Fx, or dsh, there is no agent-led flow for you: the docs say those connect with an API key at api.runinfra.ai/v1 instead. That list is the full set of RunInfra coding agents the doc supports today — everything else connects by API key. Our Claude Code beginner guide covers the agent you will most likely paste this into.

Which Coding Plan Should You Pick?

Start with Starter at $10 a month — RunInfra pricing is just three flat tiers, each covering roughly 20% more usage than its price, and the other two are bigger versions of the same shape. The whole decision is one row of this table:

Three mechanics matter more than the table. First, the 5-hour window starts with your first plan request, and the weekly limit resets on a fixed schedule counted from activation — there is no "daily" plan limit; "daily" appears only in the Standby refill terms. Second, at a limit you pick the fallback: keep full speed on credits (with a cap or no limit), then free slower Standby, and only then does a stopped request get a 402 plan_limit_reached response carrying its reset time. Third, Standby is best effort: it refills at midnight UTC with daily amounts up to $0.20 / $0.58 / $1.98 per tier, may be reduced or paused at any time, and is switched off while a payment is due.

Plans cover the chat, messages, and responses endpoints (/v1/chat/completions, /v1/messages, /v1/responses, Standby when available), with embeddings and rerank covered when available; images and audio are always pay-as-you-go, and plan usage is counted at standard per-token rates with no rollover — the plan is not a credit deposit. One honest counterpoint, reported as posted: a Product Hunt commenter (Umberto) argued that $12 of usage for $10 does not justify the limits, and asked how the plan beats the GLM, CommandCode, or Ollama paid plans — the paid-plan comparison gets its own section below.

Which Model Should Your Agent Use?

For most beginners: DeepSeek V4.1 Flash for cheap everyday coding, GLM 5.3 when you need the strongest open model, Qwen3.8 27B for the cheapest input pricing. That is the speed-vs-price arc in one line; the full menu is the table below.

RunInfra publishes a model library with per-million-token list prices. Here are the eight models in its Coding Plans library, as published:

* GLM 5.3 Flash is listed at 51% off — $0.11 down to $0.05 input, $0.03 to $0.01 cached, $0.45 to $0.22 output — until November 4, 2026, 7:59 AM UTC. On the speed side, the independent comparison table subconscious.dev (updated October 29, 2026) lists about 541 tokens per second for Nemotron 3.5 Lightning 30B, marked as a vendor figure in their row. Every price above is a vendor-published list price, checked today; what you actually pay moves with the cache, which is the next section.

How Does the Cache Make the Plan Go Further?

Session-aware routing means your agent's repeated context — the same repo files sent turn after turn — is served from cache at the cached-input price instead of the full input price. That is the whole economic trick of the Coding Plan: a coding session is mostly re-sending context the model has already seen.

RunInfra publishes different cache-hit figures in different places, so read them as marketing ranges, not promises: the homepage says up to 99.9% over the last 24 hours, the Product Hunt page says up to 99.5%, and the LinkedIn showcase says up to 99.3%. The concrete example RunInfra publishes on its homepage is DeepSeek V4.1 Flash 1017 at 98.9% cache over 24 hours, which it prices at an effective $0.012 per million input tokens against a $0.15 list — 12.5 times lower. Treat the percentages as vendor claims from pages checked today; the mechanism (cached context is cheap) is the durable part.

RunInfra vs OpenRouter vs Ollama: Which Lane Is Which?

These are three different products that keep getting compared because they all touch "cheap models": RunInfra is flat-rate open models inside your coding agents, OpenRouter is a bring-your-own-key router across many providers, and Ollama runs models on your own machine. If you searched runinfra vs openrouter, the short version is: RunInfra sells you flat-rate usage on its own GPUs; OpenRouter sells you per-token access to other providers' models.

The Product Hunt discussion names the real alternatives people weigh against RunInfra: the GLM and CommandCode paid plans and Ollama's own paid plans — same flat-rate shape, different model menus. If your priority is keeping everything on your own hardware, the local lane wins and our Ollama beginner guide is the right starting point; if you want cheap cloud models inside the agents you already use, that is exactly the lane this guide's product was built for.

7 Beginner Mistakes to Avoid

Every mistake below comes from a real, checkable constraint in RunInfra's docs — avoid them and the first week goes smoothly.

Sources

Every figure and claim above was read from a page fetched on 2026-11-03:

• RunInfra homepage (runinfra.ai) — Coding Plans from $10/mo, the verbatim one-prompt setup prompt, model library with per-model prices and cache figures
• RunInfra docs: Coding Plan (runinfra.ai/docs/introduction/coding-plan) — Starter/Pro/Team tiers, 5-hour and weekly limits, reset windows, at-limit behavior, Standby terms, endpoint coverage
• RunInfra docs: Agent Setup (runinfra.ai/docs/tools-sdks/agent-setup) — npx -y @runinfra/cli@latest, Node 22+ requirement, supported-agents table, API-key agents
• RunInfra docs: Plans (runinfra.ai/docs/introduction/plans) — 1 credit = $1, $1 free on signup, $5–$1,000,000 top-ups
• Product Hunt launch page (producthunt.com/p/runinfra/runinfra-2) — October 20, 2026 launch, model list, zero-data-retention and SOC 2 Type II claims, Umberto's pricing comment
• RunInfra LinkedIn showcase (linkedin.com/showcase/runinfra) — YC F26, cached input from $0.01/1M, 99.3% cache figure
• Subconscious.dev provider table (subconscious.dev/providers/moonshot-ai, updated October 29, 2026) — independent comparison row, ~541 tok/s for Nemotron 3.5 Lightning 30B (marked vendor)

Frequently Asked Questions

What is RunInfra used for?

RunInfra runs open AI models on its own GPUs and sells access to them. Its Coding Plans product exists for one job: putting fast, cheap open models inside the coding agents you already use — Claude Code, Codex, OpenCode, Cline, and a dozen more — for a flat monthly price instead of per-token frontier-model bills.

How much does RunInfra cost?

Coding Plans start at $10 a month (Starter, roughly $12 of usage), then $29 (Pro, about $35) and $99 (Team, about $119). Signup also gives a one-time $1 credit that never expires, and credits top up from $5. Images and audio are always pay-as-you-go, and plan usage does not roll over.

How do I connect RunInfra to Claude Code?

Paste RunInfra's setup prompt into Claude Code, approve the sign-in in your browser, and say yes to the setup plan the agent proposes. The agent wires itself using the documented npx -y @runinfra/cli@latest flow — you do not hand-edit config files.

Is RunInfra the same as OpenRouter or Ollama?

No. RunInfra is flat-rate open models inside your coding agents; OpenRouter is a bring-your-own-key router across many providers; Ollama runs models on your own machine. They overlap on "cheap models" but sit in different lanes — the comparison table above lays out which fits which need.

What happens when I hit my RunInfra limit?

Your choice, in order: keep full speed on credits (with a cap or no limit), fall back to free slower Standby, and only then do requests stop with a 402 plan_limit_reached response that includes the reset time. The 5-hour window starts with your first plan request, and the weekly limit resets on a fixed schedule from activation.