Our take: Choose AgentSky if you want to run Claude Code, Codex or other coding agents in a browser or through one API and pay only for compute seconds and model tokens, or $0 model cost on a connected Claude or ChatGPT plan.
Main limitation: Prepaid credits are required, so even connected-plan usage still pays for compute.
By the AI Tool Atlas editors.
Based on the vendor's official documentation; we have not tested this product hands-on. What “documented overview” means
Who AgentSky fits
Choose AgentSky if you want to run Claude Code, Codex or other coding agents in a browser or through one API and pay only for compute seconds and model tokens, or $0 model cost on a connected Claude or ChatGPT plan. Based on the vendor's documentation, not on a hands-on test.
Best for: Developers
Also a good fit for
Teams comparing coding agents
Consider another tool if prepaid credits required
Or if
Compute billed about $0.461/hour on default machine
Connectors and paid capabilities cost extra
AgentSky pricing and usage limits
Prices checked Oct 1, 2026 on the official US pricing page. Official pricing ↗ Prices and limits change often.
The official AgentSky pricing page as we saw it. agentsky.dev, captured Oct 1, 2026
Free access and trial conditions
No free tier confirmed; model usage is $0 on a connected Claude or ChatGPT plan, but credits are still needed for compute.
Connector calls and paid capabilities charged per use, e.g. $4 per connector (up to 4 apps, up to $16).
What is AgentSky, and how does it work?
AgentSky lets users run agents such as Claude Code, Codex, OpenCode and Hermes in the browser or through an API, and switch agent or model mid-session.
Billing is prepaid credits on two meters, compute per second and model tokens, or free model usage if a Claude or ChatGPT subscription is connected.
AgentSky's official site: how the product presents itself. agentsky.dev, captured Oct 1, 2026
What you provide and what you get back
What you give AgentSky and what comes back, from its documentation
You provide
Task prompts and repo files
Connected Claude Pro/Max or ChatGPT Plus/Pro plan, or prepaid credits
API calls such as POST /v1/agents
You get back
Agent session output, logs and execution traces
Itemized per-turn billing
Agent events streamed over SSE
Where you can use it
Platforms
Web, API
Account
Required
Default machine
2 vCPU, 8 GB
Compute rate
$0.461/hour, per second
Model router
16+ models, one key, 0 markup
Interfaces
Browser, REST API, sky CLI
Connectors
2,000+ apps
What people use AgentSky for
Scenarios are worked out by our editors from the documentation; the actual result depends on your input and plan.
Running coding agents in the browser
Who
Developers who want a hosted coding agent without local setup
You provide
A repo and a task prompt
Expected result
An agent session running in a 2 vCPU, 8 GB machine in the browser
Feature overview on the official site. agentsky.dev, captured Oct 1, 2026
40+ agents
Lists 40+ agent harnesses such as Claude Code, Codex, Hermes, OpenClaw and opencode to launch in the browser or by API Lets you swap agents without leaving the platform
Access: Any account with creditsLimit: Agent availability is per the marketplace listingsource (opens in a new tab)
Model router with one key
Routes requests to 16+ models through one key at 0 markup Avoids managing separate provider keys
Access: Credits balanceLimit: Token rates differ per model and are billed per turnsource (opens in a new tab)
Connect Claude or ChatGPT plan
Connects a Claude Pro/Max or ChatGPT Plus/Pro subscription so model usage bills $0 Cuts token cost if you already pay for a plan
Access: Set up in SettingsLimit: Compute is still metered, and Claude plans cover only Claude Code agentssource (opens in a new tab)
Per-second compute metering
Bills compute per second at $0.461/hour for the default 2 vCPU, 8 GB machine Parked agents cost nothing
Access: Prepaid creditsLimit: Connectors and paid capabilities are charged separatelysource (opens in a new tab)
AgentSky pros and cons
Where AgentSky is strong
No seats or minimums stated You pay nothing while an agent is parked, so idle sessions are freedocumentedsource (opens in a new tab)
Model usage billed at $0 with connected Claude/ChatGPT plan There is no seat fee or minimum spend, so small tests are cheapdocumentedsource (opens in a new tab)
Itemized per-turn billing Connecting a Claude or ChatGPT plan removes model token chargesdocumentedsource (opens in a new tab)
Published per-model token rates Itemized per-turn billing makes costs easy to auditdocumentedsource (opens in a new tab)
Compute billed about $0.461/hour on default machine Compute adds about $0.461 per hour on the default machine even with a plan connecteddocumentedsource (opens in a new tab)
Connectors and paid capabilities cost extra Connectors and paid capabilities add per-use charges, for example $4 per connectordocumentedsource (opens in a new tab)
How to use AgentSky for running a coding agent in the cloud
Set up AgentSky
Sign up on agentsky.dev and add prepaid credits
Optionally connect your Claude Pro/Max or ChatGPT Plus/Pro plan in Settings
Install the CLI with npm i -g @agentsky/sky and run sky auth login
Run sky launch in your project folder to create the agent
Example: running a coding agent in the cloud
Run sky launch in a repo and accept the default agent
Send a task with sky agent message "Summarize today's signups"
Switch the harness from Claude Code to Codex if you hit a limit
Open the per-turn billing to see compute and token cost
Review and export the result
Confirm the response streamed until session.status_idle
Compare per-turn cost against the published model rate
Check that a parked agent adds no compute charge
Example prompt suggested by our editors (not a recorded output of the product)
Read the repo, list failing tests, and fix the smallest one.
AgentSky API
Plan comparison on the official pricing page. agentsky.dev, captured Oct 1, 2026
API access and separate billing
AgentSky has no public API in its official documentation. REST /v1 with SSE streams; sky CLI and agent.toml. API documentation ↗
Privacy, data use and commercial rights
Summarized from the vendor's current policies; this is not legal advice and a vendor's badge is not an independent audit.
How your inputs and outputs are handled
Stores prompts, files, connector data and agent outputs; trace history is a Workspace setting.
Trace contribution is off by default; if enabled, content may be used to improve the service and licensed in anonymized form. Some model providers may train on requests unless provider training routing is restricted.
Content held as long as needed to run a request; trace history stored while enabled.
Commercial use and plan restrictions
Terms section 7 sets licensing; not detailed in pages.
Enterprise Workspace mentioned for zero-retention needs.
Controls for business users
Separate paid/free model provider-training routing settings
Prepaid credits on two meters: compute per second (about $0.461/hour for 2 vCPU, 8 GB) and model tokens at each model's published rate. source (opens in a new tab)
Can I use my Claude or ChatGPT subscription?
Yes. Connect Claude Pro/Max for Claude Code agents or ChatGPT Plus/Pro for Codex agents and AgentSky bills $0 for model usage; compute is still metered. source (opens in a new tab)
Which agents does it support?
The marketplace lists 40+ agents, including Claude Code, Codex, Hermes, OpenClaw, pi, Kimi Code and opencode. source (opens in a new tab)
Trace contribution is off by default. Some model providers may train on requests, so the Workspace has settings to restrict routing to providers without a training opt-out. source (opens in a new tab)
Not verified: Free tier or trial; Exact credit top-up minimums; Data retention period; Refund policy.
How we documented and maintain this page
Every fact about AgentSky above comes from the 2 official pages listed here, last compared on Oct 1, 2026; we have not run AgentSky ourselves, so it has no editorial score. How we check and re-check pages.
Article updated Oct 10, 2026; prices checked Oct 1, 2026. Outbound links to vendors are plain links; we do not earn commissions. How we review · Report an error