Run Claude Code, Codex and six more agent harnesses through one API
AgentSky is a cloud API that runs agent harnesses such as Claude Code and Codex against a choice of 16 models.
AI Panel Score
6 AI reviews
Reviewed
Facts checked against agentsky.dev
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.AgentSky is a cloud API that runs agent harnesses such as Claude Code, Codex and Hermes against a choice of eleven models through a single key. It suits teams that want agents inside their own product without operating the machines behind them, and developers who already run an agent locally and want it in the cloud. There is no subscription and no seats: prepaid credits are drawn on two meters, model tokens at provider list prices with no markup, from $0.44 to $10.00 per million input tokens, and compute from $0.021 to $0.071 per awake hour, with $3 of credit to start. Sessions persist between messages and suspend when idle, channels include Slack, Telegram, Discord and a plain web API, and one sky clone command lifts a local agent with its MCP servers into the cloud. TopReviewed's six-seat AI review panel scored it 8.1/10, praising the published rate card while noting the absence of a security or SLA page.
Getting started is three steps: create an account with no card, take the one API key that covers every agent and model, then name the agent and the model in your request. A playground lets you pick a harness and start working without writing the call yourself, and you can hand an agent real work - connect a GitHub repo, forward an email, upload files - rather than a toy prompt. Channels are parameters on the same call, so the agent can be reached from a web API, Slack, Telegram, Discord, WhatsApp, iMessage, a CLI, or the A2A protocol.
Eight harnesses are supported - Claude Code, Codex, Hermes, OpenClaw, pi, DeepSeek Harness, Kimi Code and opencode - against 16 models, and swapping either is a change to the request rather than a rebuild. Add-ons ride on the same call: search and web fetch, a browser, SEO and SERP data, image generation, background removal, video generation, image-to-video and transcription, plus connectors for Gmail, Notion, Slack and around 2,000 other apps. A "sky clone" command copies an agent you already run locally - instructions, model and MCP servers - into the cloud, and if you already pay for Claude Pro, Claude Max or ChatGPT you can connect that subscription so eligible agents run at $0 model usage. Agent Arena ranks harness-model pairs by Elo from head-to-head wins on 40,835 real tasks recorded between May and August 2026.
This is aimed at teams who want agents inside their product without running the machines: the site publishes three customer stories - Tycoon at 150,000 agents launched, HeyBoss AI at 250,000, WebJourney at 50,000-plus - each describing agents their own users interact with. The vendor positions itself as "the OpenRouter for agents". Pricing is prepaid credits with no subscription and no seats: model usage is itemised per turn at list price (from $0.042 per million input tokens on Jev 1.13 up to $10.00 on Claude Fable 5.1 and GPT-6 Astra), and compute runs $0.117 to $1.066 per awake hour by machine size, with the default 2 vCPU / 8 GB machine at $0.461. Suspended and parked agents cost nothing.
Enables search and web fetch, a browser, SEO and SERP data, image generation, background removal, video generation, image-to-video and transcription as parameters on the agent call.
Ranks harness and model pairs by Elo from head-to-head results on 40,835 real tasks recorded between May and August 2026.
Breaks billing down to the turn across two meters, model tokens and compute seconds, visible per agent.
Starts an agent from the browser by picking a harness and model, with GitHub repos, forwarded email or uploaded files as input.
Connects an existing Claude Pro, Claude Max or ChatGPT plan so eligible agents run at $0 model usage.
Runs Claude Code, Codex, Hermes, OpenClaw, pi, DeepSeek Harness, Kimi Code and opencode as cloud agents.
Keeps each agent on its own cloud machine so state survives between messages, with crashes and reconnects handled by the platform.
Pairs any supported harness with one of 16 models, including Claude Opus 5, GPT-5.6 Sol, DeepSeek V4 Pro and Kimi K3.
Suspends an idle agent so no compute is billed, then wakes it on the next message.
Exposes an agent through a web API, Slack, Telegram, Discord, WhatsApp, iMessage, a CLI or the A2A protocol.
Connects agents to Gmail, Notion, Slack and around 2,000 other applications.
Uses one key for every agent and model, with the choice named as parameters in the request.
Copies an agent already running locally into the cloud with its instructions, model and MCP servers in one command.
No subscription and no seats. Prepaid credits are drawn on two meters: model tokens at each provider's published rate, and compute per second while the agent machine is awake. Creating an account requires no card.
Connect an existing Claude Pro, Claude Max or ChatGPT plan once and eligible agents run at $0 model usage, leaving only the per-second compute meter.
One key, eight agent harnesses, and a rate card at list price. Easy pilot to approve.
“AgentSky rents the machine and the plumbing so your team stops maintaining either. The billing model is unusually legible for this category.”
Two meters, both published. Model tokens at provider list prices with zero markup, compute at $0.117 to $1.066 per awake hour, $0.461 on the default machine. No seats, no subscription, no minimum. That is a pilot finance can approve in a meeting.
What you're buying is optionality. Eight harnesses, 16 models, and switching either is an argument in a request rather than a rebuild. Three named customers describe agents their own users touch, at 150,000 and 250,000 agents launched. Vendor-stated, but specific enough to check.
The exposure is upstream. Every harness here is somebody else's open-source project and every model somebody else's API. If a provider reprices, so does your invoice. Run one production workload for a quarter and watch what the compute meter actually does before moving anything critical.
The vendor calls itself the OpenRouter for agents, and routing across eight harnesses rather than models is a genuinely different product.
No published SLA or certifications, and the security page routes zero-retention and compliance questions into a sales conversation.
Three steps to a running agent, no card at signup, and an existing local agent can be cloned up with one command.
Removes agent infrastructure from our roadmap entirely rather than shaving cost off something we already run.
Docs, a playground, a published benchmark over 40,835 tasks and three named case studies point to a shipping product, though the company itself is new.
Teams that want agents in production without running machines.
Regulated data has to pass through the agent.
Routing at the harness layer, not the model layer, is the bet worth understanding here.
“Most gateways abstract models. AgentSky abstracts the agent loop above them, which changes what your product depends on.”
The interesting decision is where the abstraction sits. A model gateway swaps weights behind one endpoint; this swaps the whole harness — Claude Code, Codex, Hermes, opencode — while keeping the session, the tools and the channel intact. If we build on it, our dependency becomes the request shape, not any single agent project.
Architecturally the state story carries the weight. Each agent gets a persistent machine that suspends when idle and wakes on the next message, so crash handling and reconnects sit on their side of the line. Add-ons ride the same call: browser, search, transcription, roughly 2,000 app connectors.
The three-year question is upstream drift. Harnesses evolve independently and a benchmark spanning May to August 2026 is a snapshot, not a trend.
Sits a layer above the model gateways it borrows its framing from, which is a defensible spot while harness choice keeps multiplying.
Persistent machines that suspend and wake match how agent workloads actually behave — long idle periods punctuated by bursts.
Web API, Slack, Telegram, Discord, WhatsApp, iMessage, CLI and A2A as call parameters, plus Gmail, Notion, Slack and around 2,000 connectors.
The dependency taken on is the request shape, but every harness underneath evolves on its own schedule and outside the vendor’s control.
Abstracting the harness rather than the model, with session, tools and channel preserved across a swap, is a deliberate and unusual architectural line.
Teams whose product embeds agents its users interact with.
You have already standardised on one harness and one model.
$0.461 an hour reads cheap until nothing suspends it: $331.78 a month of compute alone.
“Both meters are published, and the vendor's own estimator lands at $105.65 a month for one agent. Suspend discipline is what keeps compute from turning that into $331.78 of machine time by itself.”
The estimator on the pricing page doesn't flatter anyone. Default 2 vCPU / 8 GB, 20 turns a day, two hours awake: $105.65 a month. $78.00 of that is model tokens, $27.65 compute. Ten agents on that profile is $12,678 a year.
Compute is the meter that can surprise you. $0.461 an hour, never suspended, is $331.78 a month before a token. Sizes run $0.117 to $1.066 an hour. Bring your own subscription zeroes the model side on eligible agents, which flips the ratio entirely.
Token rates are provider list, $0.042 per million input on Jev 1.13 up to $10.00 on Claude Fable 5.1. Extras meter separately: browser $0.03 a minute, connectors $0.029 a call, images $0.422 an asset. Top-ups carry Stripe's 2.9% plus $0.30. But there's no published SLA behind any of it.
Self-serve prepaid credits start instantly but offer no invoicing or payment terms, which some procurement processes will not accept.
Prepaid credits with no subscription, seats, minimum or renewal window to negotiate out of.
Token rates, every machine size, per-use capability prices and the card processing fee are published without a sales call.
The pricing page's estimator returns $105.65 a month for the vendor's own usage profile, but no ROI model or self-hosting benchmark is published.
Both meters are computable per agent, but an always-on default machine is $331.78 a month before any model tokens.
Teams that can enforce a suspend policy on every agent.
Budget owners who need a fixed monthly cost per agent.
sky clone lifts a local agent into the cloud with its MCP servers intact
“The workflow is shaped like something the team actually uses. The gaps show up around what happens when a run goes wrong.”
One command clones the agent you already run locally — instructions, model, MCP servers — into the cloud. That's the tell. Whoever built it moves agents between machines often enough to hate doing it by hand.
Day-to-day the shape is good. Sessions persist between messages, suspend on idle and wake on the next one, so you're not writing reconnect logic. Usage is itemised to the turn across both meters, which is what you need at 2am when one agent's bill looks wrong. Channels are parameters, so putting the same agent in Slack and behind an HTTP endpoint isn't two integrations.
What's missing is failure detail. Restart behaviour is documented as managed recovery to the latest durable point, not zero-loss, but retry and timeout semantics are absent, and there is no changelog. You'll find those boundaries yourself.
Persistent sessions and automatic suspend remove the two chores — reconnects and idle cost — that usually land on the caller.
Docs, a playground and a published benchmark exist, but nothing published covers operational edges.
Failure semantics — retries, timeouts, what happens on restart — are not documented beyond a one-line FAQ on restart recovery, so early debugging is exploratory.
Eight harnesses, 16 models, built-in browser and transcription tools and A2A support leave room well past the first use case.
Cloning a local agent with its MCP servers in one command means the cloud version starts from the setup you already trust.
Engineers already running agent harnesses locally.
You need documented operational guarantees before shipping.
No card, no subscription, and a playground that starts an agent before you write any code.
“Trying it costs almost nothing and the first agent takes minutes. Where it gets murky is knowing what you have spent.”
No card is requested at signup, and the playground is enough to run something real rather than a hello-world. The playground lets you pick a harness and a model from a dropdown and start, and you can hand it a GitHub repo or a forwarded email instead of a toy prompt. That's a better first ten minutes than most developer tools manage.
The billing page does something I wish more products did: it tells you the card fee. $100 of credit costs $103.30. Nobody hides that well and this one just says it.
What nags is the two-meter model. Tokens and awake seconds are both perfectly reasonable and together they're two things to keep in your head. The estimator helps. It's still arithmetic you didn't have before.
The pricing page carries an estimator with model, machine size and turns per day as inputs, and discloses the card processing fee.
Starting is easy, but two billing meters and a choice between eight harnesses is more to hold than a single-model API.
No mobile app is described, though agents are reachable from WhatsApp, Telegram and iMessage, which covers the phone case differently.
No card at signup, a three-step flow and a playground that accepts a GitHub repo or forwarded email as the first task.
The platform states it handles crashes, reconnects and state, but no status page or uptime figure is published.
Curious people who want an agent running today.
You want one flat monthly number and no meters.
The pricing page answers everything to the cent; the security page hands the questions back to you.
“Every machine rate and metered extra is published down to the card fee, which is rarer than it should be. The security page exists, but it reads as a checklist of questions to take into a sales call rather than answers.”
The rate card and the security page read like they came from different companies. One publishes everything — $0.03 a minute for the browser, $0.029 a connector call, even the 2.9% + $0.30 Stripe fee on a credit top-up. Nobody planning to make money on opacity writes that down.
The other one exists — data boundaries, a pilot checklist, a pointer at the privacy policy for retention windows — but it is a list of questions to resolve, not answers. Bring your certification and service-level requirements to a sales call. No SLA. No status page. No changelog either.
Exit is fine, which surprised me. The harnesses are open-source and the models are public APIs, so leaving means running what you already run. Agent Arena's Elo comes from 40,835 tasks with dates attached. The 450,000+ agents launched is a banner number, unaudited.
Harness-level routing across eight harnesses and 16 models is a real gap in the market, but the vendor's own OpenRouter framing invites a comparison it has not yet earned.
The harnesses are open-source projects and the models are public APIs, so leaving means self-hosting what you already use.
Docs, a playground and an active benchmark show maintenance, but there is no changelog, no status page and no published SLA.
Every rate and metered extra is published to the cent, though the security page states no certifications or retention windows of its own.
Named harnesses, models and the 40,835-task Agent Arena run are checkable, but the 450,000+ agents launched banner has nothing behind it.
Teams who can run their own security review on a new vendor.
Buyers who need a signed SLA before any pilot starts.
Common questions answered by our AI research team
There is no subscription and no seats. You prepay credits and are billed on two meters: model tokens at provider list prices, from $0.042 to $10.00 per million input tokens, and compute from $0.117 to $1.066 per awake hour ($0.461 on the default 2 vCPU / 8 GB machine).
Eight harnesses: Claude Code, Codex, Hermes, OpenClaw, pi, DeepSeek Harness, Kimi Code and opencode. Each pairs with any of 16 models, including Claude Opus 5 and GPT-5.6 Sol.
Yes. Connecting a Claude Pro, Claude Max or ChatGPT plan lets eligible agents run at $0 model usage, leaving only the per-second compute meter to pay for.
Yes. Each agent runs on its own persistent cloud machine that holds state across messages, suspends automatically when idle and wakes on the next message.
Yes. The sky clone command copies an agent you already run locally, with its instructions, model and MCP servers, into the cloud on the same API.
Channels are parameters on the agent call. AgentSky supports a web API, Slack, Telegram, Discord, WhatsApp, iMessage, a CLI and the A2A protocol.
You can create an account with no card required, but there is no permanently free tier: usage draws on prepaid credits.
Who AgentSky competes with, and how each one differs
Warp's Oz platform is the closest functional match: it also runs Claude Code, Codex and its own agent inside sandboxed cloud containers across any model, driven from CLI or API. It sells to enterprise engineering teams as a per-seat subscription with concurrency caps (free/4 concurrent up to $180/mo/40 concurrent) wrapped in audit trails and access control, where AgentSky meters per-second compute plus pass-through model tokens on one key with no seat commitment.
OpenRouter is the model router AgentSky names itself after ("the OpenRouter for agents"): one key across hundreds of models at provider list prices with no markup on tokens. It stops at the inference call, with no agent harnesses, no persistent machines that suspend and wake, and no chat-channel surface, so the caller still supplies the runtime AgentSky sells.
Vercel's AI SDK HarnessAgent API plus Vercel Sandbox and AI Gateway cover the same "one interface, many harnesses" idea (Claude Code, Codex, Pi today), but ship as an SDK the developer wires into their own application and deployment, billed as Vercel platform usage. AgentSky exposes the equivalent as a hosted endpoint where session lifecycle, suspension and eight harnesses are managed server-side.
Anthropic's Claude Managed Agents host an agent server-side with a managed sandbox and a filesystem that persists across interactions, the same shape as AgentSky's suspended machines, but restricted to Claude harnesses on Claude models. AgentSky resells those same Claude models at list price while also offering seven other harnesses and rival models (GPT, Gemini, DeepSeek, Kimi, GLM, Grok), which makes Anthropic both its largest supplier and a single-vendor substitute.
Modal sells the substrate rather than the agent: sandboxes with gVisor isolation whose compute charges drop to zero while idle, the same suspend-when-idle economics AgentSky bills on. The customer brings their own harness code, model API keys and session plumbing, so Modal is the build-it-yourself alternative to AgentSky's managed harness catalogue.
Devin is one proprietary coding agent on its own persistent cloud workspace, reachable from Slack and sold to engineering teams on seats and ACU credits. AgentSky sells the runtime instead of the agent: the caller names the harness and model per request across eight harnesses and eleven models, and pays only for model tokens and awake compute.
AgentSky is a US-based cloud platform, operated by HeyMall, Inc., that runs eight AI agent harnesses and a dozen models through a single API with persistent sessions.
