Skip to main content
An instance is one hosted OpenClaw runtime. Think of it as its own computer. Each instance has a filesystem, a browser, a terminal, web search, and an LLM. You can upload files to it, download files from it, give it API keys, and let it interact with any service a person at a computer could use. Give an instance a GITHUB_TOKEN and it can read repos, open PRs, and trigger workflows. Give it a DATABASE_URL and it can query your database. Give it Salesforce credentials and it can pull CRM data. The instance does not need a dedicated integration for every service. If there is an API and you provide the credentials, it can figure out the rest. During skill development, the manager creates multiple instances to run variants in parallel. Once the skill converges, it deploys the winning version to a production instance and cleans up the variants.

​
What an instance can do

Browse the web

Navigate pages, extract data, fill forms, take screenshots

Run code

Execute scripts, install packages, run shell commands

Read and write files

Upload, download, edit files on the instance filesystem

Call APIs

Any HTTP API with the right credentials

Search the web

Find information, datasets, documentation

Use an LLM

Managed model access for reasoning, analysis, generation
You do not need a pre-built integration for every service. If you can give the instance an API key or a URL, it can work with that service. Integrations are convenient shortcuts, not requirements.

​
Instance types

Includes Operator managed defaults:
  • Managed skills and integrations
  • Managed AGENTS.md prompt
  • Operator provider injection for model access
  • Managed search and runtime wiring
This is the fastest way to get started. When you ask the manager to build a skill, it creates prebuilt instances by default.
If a request is ambiguous, the manager asks whether you want blank or prebuilt before creating anything.

​
Instances during skill development

When the manager tests a skill, it creates variant instances:
PhaseNamingExample
First test round{skill}-v1, {skill}-v2, …lead-scoring-v1
Improvement rounds{skill}-gen{N}-v{M}lead-scoring-gen2-v3
Variant instances are temporary. After scoring, losing variants get deleted to free capacity. The winning skill gets installed on a production instance that persists.

​
Statuses

StatusIconMeaning
provisioningContainer, config, and storage are being set up
activeLive and accepting work
inactiveDeleted or stopped
suspendedBilling enforcement removed the container, data preserved

​
Instance detail page

Status

Runtime state and local cron assignments

Skills

Installed managed and custom skills

Channels

Telegram and Discord connections

Environment

Secret access and runtime settings

Server

Browse and download files, including skills, scripts, and workspace data
To download a skill, open the instance, go to Server, and browse .openclaw/skills/{skill-name}/. You can download the SKILL.md, scripts, references, or any file the agent created.

​
Secret access is separate from skill installation

A skill can be installed and still unusable. An instance may have the GitHub skill but cannot use it until a matching secret like GITHUB_TOKEN exists and is either global or granted to that instance.
Read Environment and Integrations together.

​
Capacity and plan limits

Your plan controls how many active or provisioning instances you can have at once.
PlanInstance limit
Basic1
Pro5
Max20
During skill testing, variant instances count against this limit. If you are at capacity, the manager adjusts the number of variants or asks if you want to run sequentially.

​
Security and persistence

  • Instance config is encrypted at rest (AES-256-GCM).
  • Secrets are managed separately from config.
  • Suspending an instance removes the running container but keeps data for later recovery.

Skills

Build, test, improve lifecycle

Channels

Connect Telegram or Discord

Environment

Secrets and access control

Integrations

External services and APIs