OpenClaw Cost: Software, API and Hosting Pricing

The short answer

OpenClaw itself is open-source software, but a useful deployment is rarely cost-free. Your real bill can include a model provider, hosting, web search, media processing, embeddings, third-party APIs and operator time.

For most personal installations:

• Software licence: $0
• Gateway host: $0 on an existing machine, or a monthly server charge
• Model: subscription quota, pay-per-token API usage or local compute
• Tools: free, metered or subscription-based depending on provider
• Operations: your time for updates, security and recovery

There is no single “OpenClaw monthly price” because OpenClaw routes work to the services you choose.

The five cost buckets

1. OpenClaw software cost

The public OpenClaw repository is open source. You do not buy a per-seat OpenClaw licence to install the Gateway.

That does not make the deployment economically free. Open-source software exchanges a licence bill for infrastructure choice and operational responsibility.

2. Model cost

Every model-backed reply consumes provider capacity. How you pay depends on authentication:

• direct API keys may be billed per input, output and cached token;
• hosted subscriptions or coding plans may expose quota windows rather than a dollar amount per message;
• local models avoid hosted token billing but require hardware, electricity and maintenance;
• routing providers may add their own price and markup structure.

Model cost grows with:

• long conversation histories;
• large files or retrieved context;
• high-output requests;
• repeated tool loops and retries;
• background jobs;
• multiple agents working on the same task;
• premium reasoning or multimodal models.

Use current provider pricing pages for arithmetic. Do not copy an old token price from a blog and build a budget around it.

3. Hosting cost

You can run the Gateway on:

• an existing laptop or desktop;
• a home server;
• a VPS or cloud instance;
• a container platform;
• a dedicated machine for a sensitive trust boundary.

An existing always-on machine may have no new monthly invoice, but it still consumes electricity and shares a security boundary with other workloads.

A small VPS creates a clean, always-on deployment surface but adds compute, storage, backup and possibly bandwidth charges. Current entry-level cloud prices change, so verify the provider’s own pricing page before publishing or approving a budget.

The cheapest host is not automatically the cheapest operation. A server that is difficult to back up, patch or access safely can generate more labour than it saves.

4. Tools and third-party API cost

The official API-cost reference identifies several features that may call paid services:

• web search;
• Firecrawl or other web extraction;
• transcription and media understanding;
• image and video generation;
• remote memory embeddings;
• speech and voice providers;
• provider usage probes;
• model scans;
• skills that call external APIs.

Some alternatives are key-free or local. For example, local embedding providers can avoid hosted embedding bills, while a managed provider may be easier to operate.

A cheap model paired with uncontrolled search, transcription and image-generation calls can still create an expensive agent.

5. Operational cost

Budget time for:

• upgrades and compatibility checks;
• security audits and credential rotation;
• backups and restore testing;
• monitoring failed jobs and channel disconnects;
• reviewing new skills and plugins;
• investigating unexpected usage;
• recovering from a tool or model mistake.

This is the cost most “free AI agent” comparisons ignore.

How to see OpenClaw usage

Per-session status

In chat, use:

/status

When usage metadata and local model pricing are available, OpenClaw can show an estimated cost for the last reply.

Per-message usage footer

/usage full

This appends token usage and, where supported, estimated cost to replies.

Other modes:

/usage tokens
/usage cost
/usage off

Subscription or CLI-backed providers may report tokens or quota without a per-message dollar estimate.

Cross-session analysis

The Control UI’s Usage area derives totals from available local transcripts and can break activity down by provider, model, agent, channel and token type.

Those figures are estimates from local history—not a provider invoice or guaranteed lifetime total. Missing price metadata creates gaps.

Provider quota windows

openclaw status --usage
openclaw channels list

For supported providers, openclaw status --usage is the canonical command for quota windows such as a percentage remaining. openclaw channels list is primarily channel and authentication-profile inventory; it may include usage snapshots where supported, but do not treat it as a billing report. Quota and cost are different concepts: a plan can have no per-message dollar estimate while still enforcing a usage limit.

A practical monthly budget model

Use this formula:

monthly cost = hosting
             + model/API usage
             + paid tools
             + storage/backups
             + operator time

For model APIs:

model cost = input tokens × input rate
           + output tokens × output rate
           + cache writes/reads × applicable rates

Then add a contingency for retries, background jobs and unusually long sessions. Agentic work is bursty; an average chat message is not a reliable proxy for a browser-heavy research task.

Three realistic operating patterns

Existing-machine personal assistant

• Gateway on a computer you already own
• one hosted model provider
• mostly key-free or bundled tools
• no always-on cloud host

This minimises direct infrastructure spend, but availability depends on that machine and its network.

Small always-on VPS

• monthly server and backup charge
• hosted model usage
• one or two metered search/media providers
• private remote access

This creates a cleaner 24/7 service boundary. Monitor model and tool usage, because the fixed server is often not the largest bill.

Multi-agent operator stack

• larger or multiple hosts
• several model providers
• parallel agents and scheduled jobs
• search, extraction, media and embedding services
• stronger monitoring and recovery requirements

Parallelism can multiply token and tool usage. Set budgets per provider and design jobs to stay quiet when nothing changed.

How to keep costs under control

1. Use the smallest model that reliably completes the task. Cheap failures and retries are not cheap.
2. Separate routine and high-reasoning agents. Do not run every reminder on the premium model.
3. Limit context growth. Long transcripts and huge tool outputs increase input usage.
4. Constrain loops. Set clear turn, retry and time limits for autonomous work.
5. Use caching where supported. Stable instructions and repeated context may benefit.
6. Prefer local/key-free tools when their reliability is adequate. Free but flaky can still waste model tokens.
7. Set provider-side hard limits. A dashboard warning is weaker than a billing cap.
8. Track background jobs separately. Scheduled automation can spend while nobody is watching.
9. Review skills before giving them keys. A skill can create a new billing surface.
10. Compare local estimates with invoices. OpenClaw’s usage UI is analytical, not the billing authority.

Hidden cost traps

Runaway tool loops

An agent repeatedly browsing, retrying or delegating can turn one request into many model calls. Bound the workflow and surface failures.

Context-heavy scheduled jobs

A recurring task that reloads large documents every run may spend far more on input than expected. Pre-filter data and use change detection.

Media defaults

Automatic transcription or image understanding can invoke a separate provider before the main model answers.

“Free” local models

Local inference has no token invoice but uses hardware, memory, electricity and engineering time. It can still be the right choice for privacy or predictable volume.

Missing price metadata

If OpenClaw knows token counts but not model pricing, the UI may show tokens without dollars. Zero displayed cost does not prove zero provider cost.

Frequently asked questions

Is OpenClaw free?

The software is open source. Model providers, hosting and optional APIs may cost money.

Can I use a ChatGPT or Claude subscription?

Supported authentication paths and provider policies vary. Some subscription-style or CLI integrations expose quota rather than per-message pricing. Follow the current provider documentation and OpenClaw’s model setup guide.

Can I run OpenClaw without paid APIs?

You can reduce hosted spend with local models and key-free tools, but the hardware and operational costs remain. Capability and reliability may differ from hosted services.

What usually costs the most?

For active personal use, the model often dominates variable cost. For a lightly used always-on agent, hosting may be the clearest fixed cost. For media or research workflows, specialist APIs can become material.

Are OpenClaw’s displayed costs exact?

No. They are estimates based on available transcript usage and configured local pricing. Use provider invoices as the billing record.

Bottom line

Do not ask only “What does OpenClaw cost?” Ask which authority, availability and automation you need, then price the model, host and tools that provide it.

Start with one agent and one model, enable /usage full, add provider-side limits and observe a normal week before scaling. That produces a defensible budget instead of a guess dressed as a monthly plan.