AgentSky
AgentSky is a managed agent-as-a-service platform that launches long-horizon AI agents in one click. Pick any harness and any model, reach the same agent across eight channels, and keep its history through restarts, model swaps and snapshots.
What is AgentSky?
AgentSky is a managed agent-as-a-service platform: one click launches a long-lived, always-on AI agent that runs in a managed sandbox, keeps its history and can be reached wherever work already happens. The homepage puts it bluntly — no Mac mini, no setup, no ops.
The defining choice is neutrality. You pick a harness — Claude Code, Codex, Hermes or OpenClaw — and a model among the eleven references priced publicly: Claude Fable 5, Claude Opus 5 and Sonnet 4.6, the GPT-5.6 Sol, Terra and Luna tiers, DeepSeek V4 Pro and Flash, Gemini 3.5 Flash, GLM-5.2 and Kimi K3. Either can be swapped later, even mid-task, and the agent keeps its full history across the change.
That same agent answers on eight channels: WhatsApp, iMessage, Telegram, Slack, Discord, web chat, the A2A protocol and the command line. Its history, tools and state follow it across every one of them.
Underneath sits a durable runtime. Each agent runs as a managed worker with persistent state, a full timeline, snapshots, backups and restore handled by the vendor, inside an isolated sandbox that boots in seconds with only the tools you attach. Idle agents suspend automatically and parked agents cost nothing. A clone to cloud path recreates a Claude Code, Codex or OpenClaw instance you already run locally, copying configuration, instructions, model and MCP servers while leaving secrets, API keys and session history on your machine.
Nine capabilities ship built in, each with a named provider: web search and page fetch through Exa, browser sessions through TinyFish, structured SERP data through DataForSEO, image generation through GPT Image, background removal through rembg, video generation through Seedance, image-to-video through MiniMax and voice transcription through Fish Audio. Per-call spend caps of USD 1 for images and USD 5 for videos ask for approval before going further. Managed OAuth connectors reach Gmail, Notion, Slack and a claimed 2,000+ further applications.
Everything the interface does is also a REST call on the versioned /api/v1 namespace, with a public OpenAPI spec, or a command in the sky CLI, alongside a declarative agent.toml file and a standing server-sent event stream. A self-evolving mode is announced as coming soon.
What it does
- Launch a long-horizon AI agent in one click, with no API key, SDK or infrastructure to set up first
- Pick a harness among Claude Code, Codex, Hermes and OpenClaw, and a model among eleven priced references, then swap either mid-task
- Reach the same agent on WhatsApp, iMessage, Telegram, Slack, Discord, web chat, the A2A protocol and the CLI
- Clone a Claude Code, Codex or OpenClaw agent already running locally into an always-on cloud agent with sky clone
- Call built-in capabilities: web search and fetch, browser sessions, SERP data, image and video generation, background removal, transcription
- Connect Gmail, Notion, Slack and a claimed 2,000+ further applications through managed OAuth connectors
- Drive the whole fleet from the sky CLI and the REST API, and park idle agents at no cost
When to use AgentSky / When not to
A quick filter to help you decide if AgentSky is the right fit.
When to use AgentSky
- Developers who want a permanent, scriptable agent without standing up sandboxes, queues or channel integrations themselves
- Teams benchmarking harnesses and models against their own workload, using identical agents and the public leaderboard
- Small teams hiring a virtual teammate that lives in Slack, Telegram, iMessage or web chat and remembers weeks of context
- Skill and tooling authors who need live, long-running agents to iterate against without resetting the conversation
- Hackathon and prototype builders who need agents in seconds and want to park them for free once the weekend is over
When not to use AgentSky
- Organisations with a GDPR or procurement checklist: there is no privacy policy, no DPA, no subprocessor list and no named legal entity
- Buyers who need contractual data residency, since no hosting country or region is disclosed anywhere on the site
- Anyone looking for a permanently free tool, as the USD 3 sign-up credit is the only free usage before prepaid credits apply
- Users who want a dedicated mobile application rather than reaching an agent through WhatsApp, iMessage or Telegram
- Risk-averse buyers who avoid very young vendors: the domain was registered on 31/07/2026 and first archived on 08/08/2026
How to use AgentSky
A typical end-to-end flow, from setup to results.
- Sign up at agentsky.dev: a USD 3 credit is granted with no card required
- In the composer, pick the harness for your agent — Claude Code, Codex, Hermes or OpenClaw
- Pick the model, then grant the built-in capabilities the agent should be allowed to call
- Press Launch agent; the sandbox boots in seconds and the agent starts its first session
- Connect the channels you use — Slack, Telegram, Discord, WhatsApp, iMessage or web chat — and approve the OAuth connectors you need
- Alternatively, install the CLI with npm i -g @agentsky/sky and authenticate with sky auth login (use --no-browser on headless machines, or set SKY_API_TOKEN in CI)
- Run sky launch from a project directory: it scaffolds ./agent.toml, creates the agent and prints its address
- Talk to it with sky agent message for a single streamed turn, or sky agent chat for an interactive session
- Over HTTP, POST to /api/v1/sessions/{id}/messages for a bare 202, then read /api/v1/sessions/{id}/stream until turn.status_idle; replay missed events from the events endpoint and deduplicate by event id
- Edit agent.toml and re-apply with sky deploy, then park the agent when work stops — parked agents are free
Pros & Cons
Pros
- No lock-in on harness or model: either can be swapped mid-task and the agent keeps its full history
- Eight channels out of the box, Slack included, where the alternatives named on the vendor's own comparison page offer fewer or none
- Durability is genuinely managed: continuous history, snapshots, backups and restore, not a promise left to the customer
- Transparent metering, with per-million-token model rates, per-second compute rates and a monthly cost simulator published openly
- Starting a project costs nothing: a USD 3 credit with no card, free harnesses and free parking for idle agents
- A complete developer surface — REST API, OpenAPI spec, CLI, agent.toml, SSE — rather than a UI with an API bolted on
- Bring your own Claude Pro/Max or ChatGPT subscription and eligible agents run at USD 0 model usage
Cons
- No privacy policy at all: agentsky.dev/privacy returns 404 and the footer carries no privacy link
- No GDPR mention, no Data Processing Agreement, no subprocessor list and no Article 27 representative
- No legal entity, no postal address, no about page and no contact page; support@agentsky.dev is the only published address
- No data hosting country or region is disclosed, so residency cannot be assessed
- No permanent free tier: past the USD 3 welcome credit, every awake minute and every token is billed
- A very young product — domain registered 31/07/2026, first archived 08/08/2026 — with the self-evolving feature still unreleased
- No official social presence found (no LinkedIn company page, Instagram or YouTube), only a Discord server
Pricing & Plans
There is no permanent free plan. Signing up grants a one-off USD 3 credit with no card required, and parked agents remain free indefinitely, but any active agent is billed. The lowest entry point is USD 3.00 per month per active agent, to which usage-based charges are added: model tokens at each vendor's published rate and compute billed per second while the agent's machine is awake. Harnesses themselves are free.
- the pricing page publishes meters rather than plans
- Harness (Claude Code
- Codex
- Hermes
- OpenClaw) — free
- Active agent — from USD 3.00 per month per agent
- plus usage
- Parked or suspended agent — USD 0
- free indefinitely
- Model usage — billed per turn at each vendor's published per-million-token rate
- from USD 0.21 input / USD 0.42 output (DeepSeek V4 Flash) to USD 15.00 input / USD 75.00 output (Claude Fable 5)
- 1 GB at USD 0.021 per hour
- 2 GB at USD 0.038 per hour
- 4 GB (default) at USD 0.071 per hour
- Welcome credit — USD 3
- no card required
Data, GDPR & hosting
A consolidated view of how AgentSky handles your data.
GDPR overview
There is no mention of the GDPR whatsoever. A full-text search of every page collected on 10 August 2026 — homepage, terms, pricing, developer docs, capabilities reference and use cases — returns no occurrence of GDPR, of data protection or of any equivalent phrase. No privacy policy exists: agentsky.dev/privacy returns HTTP 404 and no privacy link appears in the footer, whose only legal entry is the Terms of Service. No Article 27 representative is designated, no data protection officer is named, no Data Processing Agreement is published or offered on request, and no EU establishment is disclosed. The only published contact address is support@agentsky.dev, given in section 13 of the terms. European buyers therefore have no documented basis on which to assess compliance: the terms, effective 10 July 2026, are the only document that addresses data at all.
Who owns the data?
The Terms of Service state that you retain ownership of the content you provide and of your agents' configurations, history and outputs, to the extent the underlying model providers permit. The licence granted to the vendor is operational only: it may store, process, back up and transmit that data on your behalf in order to run the service. Connector access tokens are held by a managed OAuth broker rather than stored by AgentSky. The vendor retains every right in the service itself, including its software, design and branding. Deleting an agent removes its runtime and stored state, although the retention practices governing that deletion are never published.
Reuse rights
Nothing in the Terms of Service restricts what you may do with your own content, your agents' configurations or their outputs: ownership stays with you and no further permission is required to reuse them, subject only to whatever the underlying model providers allow for generated output. The rights you grant the vendor run the other way and are narrow — store, process, back up and transmit, on your behalf, to operate the service. Training is never mentioned, in either direction. Zero Data Retention is supported and passed through where the model provider offers it. You remain responsible for how your agents act on third-party platforms such as Slack or Telegram, and for complying with those platforms' terms; outputs are AI-generated and must be reviewed before you rely on them.
Data retention & training
Hosting summary
Nothing about hosting jurisdiction is published. No country, no region and no data centre is named on any page collected, and there is no privacy policy or trust page in which such information would normally sit. What the site does describe is architectural rather than geographic: agents run in isolated, managed environments that boot in seconds, with snapshots, backups and restore handled by the vendor, and connector access tokens held by a managed OAuth broker rather than stored by AgentSky. Named third-party providers process data per capability — Exa, TinyFish, DataForSEO, GPT Image, rembg, Seedance, MiniMax and Fish Audio — alongside the model vendors, but their own hosting is not disclosed either. The website itself is served behind Cloudflare, resolving to 104.21.58.6 in the United States, which reflects an anycast edge node and says nothing about where customer data is stored. Buyers with a residency requirement have no documented answer here and should ask before committing.
Things to keep in mind
Risks and trade-offs to weigh before adopting AgentSky.
- Data governance is undocumented: with no privacy policy, no retention schedule and no hosting location, you cannot tell where your agent's memory lives or how long it survives
- Costs scale with the agent's appetite, not with a flat plan; the vendor's own simulator reaches USD 121.25 per month for twenty turns a day and two waking hours on its most expensive model
- An always-on agent with connector access to Gmail, Notion and Slack is a standing permission: review the per-call spend caps and the capabilities you grant before you automate anything
- You are responsible for what your agents do on third-party platforms and for complying with those platforms' terms, while the vendor's liability is capped at the fees paid in the previous twelve months
- Outputs are model-generated and may be inaccurate; the terms say plainly that you must review them before relying on them, which is easy to forget once an agent has been trustworthy for weeks
- A teammate that remembers everything invites delegation without review, and long-horizon autonomy makes it harder to notice when judgement has quietly been outsourced
- The vendor depends on third-party harnesses and models whose availability it does not control, and no minimum age is stated anywhere
Setup & Integrations
Technical difficulty
Very low for everyday use. The interface asks three questions — harness, model, capabilities — then launches the agent; there is no API key, SDK or infrastructure to prepare, and connectors are authorised once through managed OAuth. Developers who prefer the terminal install the CLI and authenticate in two commands, then run sky launch, with the agent.toml file optional. Only product integration is genuine engineering work: the REST API, the event stream and the replay semantics assume a developer. Nothing needs installing locally to use an agent day to day.
Deployment
Integrations
Behind AgentSky
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement AgentSky.
Frequently asked questions
What is AgentSky?
Which harnesses and models can I run?
How much does it cost?
Is there a free plan?
Which channels can I reach my agent on?
Can I move an agent I already run locally to the cloud?
Does the agent lose its memory when it restarts?
Is there an API?
What does AgentSky publish about privacy and data protection?
Should you pick AgentSky?
AgentSky occupies a genuinely useful gap. Running an always-on agent yourself is cheap to start and expensive to keep, while developer platforms hand you a toolkit and leave the channels, the sandboxing and the uptime to you. AgentSky is what those platforms look like once someone has finished assembling them: harness, runtime, channels and connector auth already wired, reachable in one click.
The engineering choices are consistent with that promise. Neutrality on harness and model is real and reversible mid-task; eight channels including Slack is more than any alternative the vendor names on its own comparison page; durability is managed rather than described; and the developer surface — REST API, OpenAPI spec, CLI, agent.toml, event stream — is complete rather than decorative. Pricing is unusually candid, with per-token and per-second rates published in full and a simulator that shows what a working agent actually costs.
The reservation is not technical, it is documentary. There is no privacy policy, no GDPR mention, no Data Processing Agreement, no subprocessor list, no disclosed hosting location and no named legal entity or postal address. A single email address is published. For a product that runs autonomous agents against your Gmail, your Notion and your Slack, that gap matters far more than any missing feature, and it will stop most procurement processes before the demo.
The explanation is probably the calendar: the domain was registered on 31 July 2026 and first archived on 8 August 2026. This is a product days old, and its legal scaffolding has not caught up with its engineering. Try it on a low-stakes agent with the free credit, judge the runtime on its own merits, and re-evaluate the governance question once the missing documents appear.
- Choosing a selection results in a full page refresh.
- Opens in a new window.