Skip to main content
OpenClaw is an open-source (MIT) personal AI assistant that runs on your own computer or server and answers you in Telegram, WhatsApp, Discord, Slack, iMessage, Teams and 20+ more services. It ships no models of its own: RuAPI plugs in as a custom provider — Claude in the native Anthropic format, DeepSeek, GPT and GLM in the OpenAI-compatible format. One key covers every model, and requests go to www.ruapi.ai without a VPN. For a short explanation of why OpenClaw needs a provider like RuAPI, what the models cost and how to pay without a foreign card, see www.ruapi.ai/openclaw-v-rossii; the setup steps are below.
This guide was checked against the OpenClaw source code and docs (2 Nov 2026). If a menu item or field has a different name in a newer OpenClaw release, check docs.openclaw.ai.

What you need

1

RuAPI account and API key

Console → API Keys → Create API key. OpenClaw replies in messengers and runs scheduled jobs on its own, so set a Credit limit on the key — the agent cannot spend more than that. See API keys & security.New accounts usually get $0.05 right away: on that balance claude-haiku-5-5, deepseek-flash and gpt-6-luna work. Other models become available after the first top-up — in USDT from 5 USDT, no foreign card needed (how to buy USDT with rubles via SBP). Foreign-issued bank cards are accepted too.
2

Operating system

macOS, Linux or Windows. On Windows you can install the Windows Hub app, install OpenClaw with the PowerShell script, or work in WSL2 — all options are below.
3

Node.js — only for npm installs

The official installer sets up Node.js itself if it is missing. For an npm install you need Node.js 24 (24.16+) or 26 (26.1+); the developers recommend Node 26, and 24.22.1+ or 26.12.1+ for security fixes. Check with:

Install OpenClaw

The installer checks Node.js, installs OpenClaw and opens the setup wizard right away: you can connect RuAPI there (method 2 below) or exit and add the provider to the config file (method 1). Check the install:

Connect RuAPI

To OpenClaw, RuAPI is two custom providers that share one key: Connect Claude in the Anthropic format: that is what makes Claude caching work (see below). Models are referenced as provider/model, e.g. ruapi/claude-sonnet-5-5 or ruapi-openai/deepseek-flash.
1

Open the config file

It is usually ~/.openclaw/openclaw.json; this command prints the exact path:
If the file doesn’t exist yet, create it. If it does, don’t remove anything: add the fields from the next step inside the existing models and agents sections (don’t create a second copy of either section).
2

Add the providers and the primary model

  • id is the exact model name from the pricing page, case-sensitive. List every model you want to call under models.
  • input: ["text", "image"] marks the model as image-capable, so OpenClaw passes photos from messengers to it as images. Without this field the model is treated as text-only; glm-5.3 has none because it works with text only.
  • fallbacks are backup models: if the primary one fails, OpenClaw tries them in order.
Don’t add /v1 to the ruapi provider URL: OpenClaw’s Anthropic client appends it itself.
On the $0.05 starter balance, make ruapi/claude-haiku-5-5 or ruapi-openai/deepseek-flash the primary model — claude-sonnet-5-5 works after the first top-up.
3

Check the settings

The Gateway watches the file and applies changes on its own. Validate the config and check the selected model:
models status should show ruapi/claude-sonnet-5-5 as the primary model and ruapi-openai/deepseek-flash as the fallback. If the changes weren’t picked up, run openclaw gateway restart.
Put the key in ~/.openclaw/.env:
and reference it in both providers: apiKey: "${RUAPI_API_KEY}". You can also set variables inside openclaw.json itself, under env.vars, e.g. env: { vars: { RUAPI_API_KEY: "sk-…" } }. Keys placed directly under env (without vars), as in older examples, are not accepted by the current OpenClaw.

Method 2: the openclaw onboard wizard

The wizard is handy when you need a single provider. The installer opens it automatically; to run it by hand (the same command also switches providers on an existing install):
1

Find Custom Provider

Choose Quick start. OpenClaw shows any model access it detected and the provider list: open More… → Custom Provider.
2

Fill in the fields

For models OpenClaw doesn’t recognize, the wizard also asks Does this model support image input? — answer no for GLM models (e.g. glm-5.3): they work with text only. The wizard treats deepseek-flash as text-only on its own; to send it images, add input: ["text", "image"] to it in openclaw.json afterwards.Choose the format explicitly rather than Unknown (detect automatically): auto-detection may connect Claude in the OpenAI format, where OpenClaw sets no cache markers and only RuAPI auto-caching applies.
3

Wait for the check

Before saving, the wizard sends a real request to the model. If the check fails, it offers to fix the URL (Change base URL) or the model (Change model).
4

Second provider — same way

To add both Claude and DeepSeek/GPT, run openclaw onboard again and add a second Custom Provider with its own Endpoint ID. You can switch the primary model later with openclaw models set ruapi/claude-sonnet-5-5.

First request

Open a chat right in the terminal and ask for a one-sentence hello:
Or open the web dashboard with openclaw dashboard and send a message there. You should get a reply, and RuAPI Console → Usage Logs should show an entry for the selected model. Check real charges there: OpenClaw’s own stats count a model without a cost field as free. To keep the Gateway running in the background and after reboots, install it as a service: openclaw gateway install (or openclaw onboard --install-daemon). Status: openclaw gateway status.

Choosing another model

Or change agents.defaults.model.primary in openclaw.json. The model must be listed under models of its provider. Full list and prices: pricing page. OpenClaw hands short internal tasks — session titles, recaps, transcript summaries — to a utility model. You can make it a cheap one:

Claude caching

  • Anthropic format (ruapi provider). OpenClaw sets cache_control markers on tool definitions, system instructions and recent messages by itself — the cache lives 5 minutes. The agent’s repeated context is billed at the cache-read price.
  • One-hour cache is turned on with cacheRetention: "long" for a specific model. Writing to that cache costs 2 × the input price (see prompt caching).
    "short" is the regular 5-minute cache, "none" sends no markers.
  • OpenAI format. If you connect Claude through ruapi-openai, OpenClaw sets no markers for it, and only RuAPI auto-caching applies: on the first turn of a new conversation at most the system prompt and tools are cached, and after a gap of more than 5 minutes the cache expires. DeepSeek and GPT cache automatically, no markers needed.
To check caching, open Usage Logs: the entry for a repeated request should show a “Cache Tokens” line. If it doesn’t, set cacheRetention: "short" for the model explicitly.

Troubleshooting

  • apiKey must be a RuAPI key (sk-...), not an Anthropic or OpenAI key.
  • If the key is set as ${RUAPI_API_KEY}, make sure the variable exists in ~/.openclaw/.env or env.vars.
  • In RuAPI Console → API Keys, confirm the key is enabled and the model is in the key’s Allowed models (if that list is set).
  • The model ID needs the provider in front: ruapi/claude-sonnet-5-5, ruapi-openai/deepseek-flash.
  • The model must be in that provider’s models, and its id must match the pricing page exactly, including case (GLM-5.2 but glm-5.3).
  • baseUrl: https://www.ruapi.ai for anthropic-messages, https://www.ruapi.ai/v1 for openai-completions.
  • Before the first top-up, only claude-haiku-5-5, deepseek-flash and gpt-6-luna work.
OpenClaw only accepts a config that fully matches its schema: an unknown key or invalid value stops the Gateway from starting. Check with:
Common mistakes from older examples: api: "openai" instead of "openai-completions", and models as a list of strings instead of { id, name } objects.
This page used to suggest setting ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY and calling models as anthropic/claude-.... The built-in anthropic provider is designed for Anthropic’s official API, so switch to a custom provider:
  1. Remove ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY from env (or env.vars) in openclaw.json and from ~/.zshrc / ~/.bashrc.
  2. Add the ruapi provider as in method 1 above.
  3. Replace anthropic/... with ruapi/... in agents.defaults.model.
  • Check connectivity: curl -I https://www.ruapi.ai.
  • Exit the wizard (Ctrl+C) and add the provider to openclaw.json by hand (method 1 above).
For unknown OpenAI-compatible endpoints, OpenClaw doesn’t request token usage in streaming mode by default. This doesn’t affect RuAPI billing. To let OpenClaw see usage itself, add compat: { supportsUsageInStreaming: true } to the ruapi-openai models — RuAPI supports it.
During setup OpenClaw creates a recurring scheduled job (proactive check). List jobs with openclaw automations list --all and turn one off with openclaw automations disable <id>.
Ollama is a built-in OpenClaw provider (models like ollama/...). Add the RuAPI providers alongside it: for example, ruapi/claude-sonnet-5-5 as the primary model and a local model in agents.defaults.model.fallbacks, or the other way round. Details: OpenClaw docs.

Connecting messengers

OpenClaw’s main value is talking to a model through a messenger. Once RuAPI is set up, connect Telegram, WhatsApp, Discord and other channels following the OpenClaw channel docs — model calls go through RuAPI automatically.

See also

Prompt caching

How repeated context gets at least 10× cheaper.

Hermes Agent

OpenClaw’s closest alternative — also on RuAPI.

How to buy USDT with rubles

Crypto Bot, Bybit P2P or an exchanger — three ways.