Create computer
Provision a new computer in a workspace.
template_ref create restores the template’s golden snapshot, which boots far faster than the plain base image. Without one the computer cold-boots from the base image.Request
400. The older name project_id is also accepted; workspace_id wins when you send both.409. Omitting the field, or a longer name, returns 400.linux, windows, macos, android, or ios. Any other value returns 400. windows needs a Scale plan or a Windows computer bought on your plan, otherwise it returns 403 with code WINDOWS_REQUIRES_SCALE. ios is a physical handset rather than a virtual computer. Every value other than linux only succeeds where the fleet has a host that supports it, otherwise create returns 422 with code OS_UNAVAILABLE. With a gpu, the GPU decides which os works, and a V710 or B50 on Windows has the license in its own price instead: see GPU computers.4, 8, 16, 32, or 64. 0 is treated as 4. Any other value returns 400. A create from an API key, the SDKs, the CLI or MCP that names neither ram nor cpu (and no template_ref) gets 4 GB and 1 vCPU. The same create from the dashboard gets a Medium (8 GB, 2 vCPU) when the plan has room for one. A computer launched by the hour with no size is a Medium, as described under billing_mode. A macOS computer always has 10 GB: omit ram or send 10. Any other value with os: "macos" returns 400 with code MACOS_RAM_FIXED. One computer can use up to your account’s total RAM, 32 GB at most on your own. 64 returns 403 with code CONTACT_US_FOR_SIZE unless Orgo has set your account up for computers past XL. A value above your account’s total RAM, or above what your other computers leave of it, is rejected with 403, not clamped. Ignored for os: "ios", which always records 4.With a V710 or B50 gpu, ram picks one of the GPU sizes instead: 8, 16, 32 or 64, and a Medium (16) when you send neither ram nor cpu. 64 is a GPU XL, and it is never refused with CONTACT_US_FOR_SIZE. The sizes are in V710 and B50.0.5, 1, 2, 4, 6, 8, or 16. 0 is treated as 1. Any other value returns 400. Omitted means the vCPU that goes with the RAM: 1 for 4 GB, 2 for 8 GB, 4 for 16 GB, 6 for 32 GB, and 8 for 64 GB. With a template_ref, omitted means the template’s own vCPU, never below that RAM default. 8 returns 403 with code CONTACT_US_FOR_SIZE unless Orgo has set your account up for computers past XL. 16 is above the 8 vCPU per-computer cap and returns 403 with code PER_COMPUTER_CPU_CAP. A value above what your account’s vCPU pool has left returns 403, not clamped. Ignored for os: "ios", which always records 1.With a V710 or B50 gpu, cpu is 2, 4, 6 or 8, the vCPU of one of the GPU sizes, and 8 is never refused with CONTACT_US_FOR_SIZE. The sizes are in V710 and B50.40 GB on current paid plans). 0 counts as omitted. A negative or non-numeric value, or one above your plan’s per-computer ceiling, is rejected with 400 and code disk_exceeds_quota, not clamped.WIDTHxHEIGHTxDEPTH format (e.g. 1024x768x24, 1920x1080x24). Omitted means 1280x720x24, or 1920x1080x24 for macos. With a template_ref, omitting it keeps the template’s own resolution.us (Auto (US), the default): any US city with room. eu: any city in Europe, London today. apac: any city in Asia Pacific, Sydney today. A city code runs it in that city only: nyc New York, ash Ashburn, mia Miami, chi Chicago, dal Dallas, lax Los Angeles, sjc San Jose, lon London or syd Sydney. Case does not matter, and full names such as london or sydney work too. Omitted, null or "" is the same as us.The cities with a server now, and their room, are in locations from GET /billing/compute-limits?workspace_id=. The computer keeps its location: a start after a stop, a restore and a wake bring it back there, and copies of it keep it. With no room there, the create returns 503 with code REGION_UNAVAILABLE and never runs it somewhere else. A word Orgo does not know returns 400 with code INVALID_REGION, and so does any location with os macos, android or ios. An older three letter code that is not a city, such as dfw, or a comma separated list, is read as no choice. Everything about locations is in Choose where a computer runs.os, and how each is billed, is in GPU computers.{"model": "v710"}or{"model": "b50"}: one GPU slice, on Linux, or on Windows where offered."slices": 1is accepted and can be left out.{"model": "rtx-pro-4500"}or{"model": "rtx-pro-6000"}: a whole NVIDIA card, Linux only, with"billing_mode": "hourly"only. The bare strings"rtx-pro-4500"and"rtx-pro-6000"work too."2q"(2 GB VRAM) or"4q"(4 GB VRAM): the legacy vGPU slice, Linux only, on the plan only. If no GPU host can take it, the create either returns503GPU_HOST_UNAVAILABLEor makes a CPU computer, and the201then carriesgpu_downgradedandwarning.
"v710", returns 400. "none", "", false, and "false" mean no GPU on the plan. By the hour, leave gpu out instead: "none" and "false" return 400 there. Omitted means a computer with no GPU.plan counts it against the workspace owner’s plan. monthly is the same as plan, except for a V710 or B50, where it buys that GPU computer for the month. hourly runs it by the hour from credits. Any other value returns 400 with code INVALID_BILLING_MODE. Omitted, the create uses the plan, except in the cases below that run it by the hour.A V710, B50 or RTX PRO computer needs a billing_mode, and is not moved to the hour on its own. The one exception is an account an agent made with Sign up an agent: there, a V710 or B50 with no billing_mode runs by the hour from the account’s credits. What each GPU takes is in GPU computers.Free accounts run hourly computers, a Medium only. On a paid plan the plan comes first: a computer that fits the plan’s vCPU and RAM uses the plan, and hourly for it returns 409 with code FITS_PLAN. A computer past them runs by the hour instead: a create with no billing_mode from a key with full access or Billing write is launched by the hour from credits when your credit covers it, and the 201 has billing_mode: "hourly" in instance_details. Send billing_mode: "plan" to get the refusal instead of spending credit. With too little credit the answer is 402 with code insufficient_credit. A key that may not spend keeps the plan’s refusal.Only the workspace owner can launch an hourly computer, and only with os linux or windows, no gpu other than a GPU object, and no template_ref other than a starter template; anything else returns 400. A V710 or B50 by the hour is never refused with FITS_PLAN. On a paid account the owner must be signed in, or use a key with full access or Billing write: an MCP sign-in or another key gets 403 with code HOURLY_NEEDS_BILLING. You do not send a quote. Orgo quotes the size your ram and cpu name at its listed hourly price, or a Medium when you name no size; a free account always gets a Medium. On a paid account, a size no hourly offer has returns 400 with code HOURLY_SIZE_UNAVAILABLE. Send an Idempotency-Key header so a retry launches the computer once: a UUID, or any printable key of 1 to 255 characters such as job-42. A key that is not a UUID names one launch per UTC day. The same key with another name, size, OS or workspace returns 409 with code HOURLY_REQUEST_CONFLICT. Without enough credit the create returns 402 with code insufficient_credit.Omitting billing_mode means hourly for an account an agent made that no person has claimed yet, and for an account Orgo has set to pay for every computer from credits (when the create is one hourly can take). On a Free account, a create from an API key or an MCP sign-in that omits it, and that hourly can take, runs by the hour too, instead of returning 403 UPGRADE_REQUIRED.At the account’s computer limit (50 on every plan, plan and hourly computers together), a create with no billing_mode is launched by the hour from credits when your credit covers an hour of every hourly computer you would then have, the new one included, and hourly can take the computer. The 201 then has billing_mode: "hourly" in instance_details. plan there returns 403 with code VM_SLOT_ADDON instead. Computers per account has the rules.1 to 720 (30 days). Omitted or null, it never ends. Linux and Windows only: with any other os it returns 400 with code ENDS_NOT_SUPPORTED. Any other value, such as 0, 721, 1.5, or "4", returns 400 with code INVALID_ENDS. When the time is up the computer is deleted, with everything on it. To keep a computer, leave out the end. An hourly computer with an end prepays its whole run from credits. What happens at the end, and what is prepaid, is described in Computers with an end."delete" is accepted and changes nothing, with an end or without. Omitted or null is the same. Any other value, "stop" included, returns 400 with code INVALID_ENDS: A computer with an end is deleted when its time is up. To keep a computer, leave out the end. Nothing is made.namespace/name@version, for example system/claude-code@1.0.0 (curated) or default/my-template@1.0.0 (your own). A malformed ref returns 400; a template whose build is not ready returns 409 with code TEMPLATE_NOT_READY. Explicit cpu, ram, and disk_size_gb override the template’s own hardware. The template’s own cpu and ram are brought down to what your account can give one computer rather than rejected; its disk is not, so a template disk above your ceiling returns 400 with code disk_exceeds_quota. Omitted means a plain base-image computer.billing_mode: "hourly": a quote from POST /api/billing/compute. Omitted, Orgo makes one at the listed price. Sent without hourly, it returns 400 with code BILLING_MODE_REQUIRED.billing_mode: "hourly": a UUID naming this launch, so a retry with it reads the first outcome. Omitted, the Idempotency-Key names the launch, else it gets a new one. Sent without hourly, it returns 400 with code BILLING_MODE_REQUIRED, except on a monthly V710 or B50 purchase, which it names the same way. On a monthly V710 or B50 purchase or an RTX PRO, a value that is not a UUID returns 400 with code COMPUTE_REQUEST_ID_INVALID.true makes every check (credit, the GPU limit, stock) and rents nothing: the answer is 200, with the fields described in RTX PRO by the hour. Any value other than true, false or null returns 400 with code invalid_dry_run. Ignored on any other create.Common configurations
403 with code CONTACT_US_FOR_SIZE, and Orgo sets it up on request. A V710 or B50 GPU computer is the exception: it has four sizes of its own, and its XL (8 vCPU, 64 GB) is never refused with CONTACT_US_FOR_SIZE, as described in V710 and B50. No computer goes above 8 vCPU, 64 GB RAM, or 300 GB of disk, on any plan, unless Orgo raised the disk limit for your account, up to 1,000 GB. GET /billing/compute-limits?workspace_id= says the limit in disk_ceiling_gb. cpu: 16 passes the value check and then returns 403 with code PER_COMPUTER_CPU_CAP.
Below those limits, a computer can use up to your account’s total RAM, and its vCPU comes from your account’s vCPU pool. Disk has a per-computer ceiling that depends on your plan. See pricing.
Response
Returns the created computer, with201. An RTX PRO create answers 202 instead, or 200 for a dry run, as described in RTX PRO by the hour.
creating, running, restarting, updating, suspended, frozen, stopped, error, deleted. A successful create returns running. A retry of an hourly launch that is still being made returns creating. See Get computer for what each value means.http://{host}:{port}. This is an internal fleet address, not a dashboard link and not the endpoint you connect to. Use connection_url to connect. It is rewritten whenever the computer restarts, is started again, or has its RAM resized.us, eu or apac. From the server it runs on, or from the region you asked for when that is not known. null before it is placed.chi, lon or syd. null before it is placed, and for a computer in the US whose city Orgo does not record.id (the UUID) as your handle and read this fresh from Get computer.https://www.orgo.ai/desktops/{instance_id}). Append /ws/websockify, /ws/terminal, or /ws/audio for the WebSocket endpoints; HTTP Desktop API calls go to https://www.orgo.ai/api/desktops/{instance_id}/proxy/{endpoint}.?token=, and Bearer token for its Desktop API proxy. Rotates on every restart. Do not persist it. Take a fresh value from POST /computers or GET /computers/{id}. null when the caller may not read it (an MCP sign-in, a computer’s identity token, or a key without Control write), when the password is waiting to change, or when the stored credential cannot be decrypted.provider, id, name, webUrl, vncHost, vncPort, apiPort, serverAddress, resolution, and fromPool (always false), plus hypervisor and templateTerminals when they apply.hardware. The host’s launch record for the new computer, without the host’s name or the source its state was restored from.true only when you asked for a legacy "2q" or "4q" and no GPU host was available, so a CPU computer was created instead. A V710 or B50 is never swapped for a CPU computer.model, slices (always 1 for a new computer), and vram_gb_total.gpu. What the computer costs: monthly_usd, yearly_usd, hourly_usd (running, by the hour), and lines, one per part (the GPU slice, vCPU, memory, disk, and the Windows license on Windows), each with label, monthly_usd and hourly_usd.mode is "monthly", with what the first month charged.null when it never ends, including when you sent ends_in_hours and the end could not be saved (then ends_warning says so).ends_in_hours or on_end. Read only: "delete" when the computer has an end, since it is deleted then, or null when it has none.ends_in_hours and the computer was made without an end: This computer can’t have an end, so it keeps running until you stop it. or The computer was made, but its end could not be saved, so it keeps running until you stop it. The computer exists either way, so do not retry the create. When the end could not be saved, set it with Update computer.ends_in_hours. When the prepaid run ends, which is ends_at. null when the run is not prepaid.prepaid_until. What is held from your credits for the run, in cents. null when prepaid_until is.Example
Response
Computers with an end
By default a computer never ends. Sendends_in_hours to give a Linux or Windows computer an end, from 1 hour to 720 hours (30 days) after the create. You can add time with Extend computer, move or clear it with Update computer, and give one at start with Start computer.
What happens at the end
Whenends_at passes, Orgo deletes the computer, running or stopped, as a delete does: its disk and everything on it are gone. It is never stopped first. Until then Get computer reports on_end: "delete", and deletes_at equal to ends_at.
- To keep a computer, leave out the end. To keep one that has an end, take the end off with Update computer and
"ends_at": nullbefore it comes. - Switching an hourly computer to always on takes the end off, so it is kept. A computer Orgo moves onto your plan by itself, once your plan has room, keeps its end.
- A computer you stop before its end keeps that end, and is deleted when it comes.
- An hourly computer’s billing ends at the delete. What its prepay did not use comes back to your credits.
on_end: "stop". At its end it is stopped and marked ended (ended_at), then deleted 7 days later (deletes_at) unless it is started or its end is taken off. Any new end you give it deletes it.
What an end costs
A computer on the plan is paid for by the plan, with or without an end. Nothing is prepaid. An hourly computer with an end prepays its whole run when it is created: its running rate timesends_in_hours, from your credits. It draws on that prepay until its end, so it is never stopped for lack of credit before then. Whatever it does not use comes back once: when it stops, is deleted, or reaches its end, or when its end is moved or taken off. The prepay is the run and nothing after it: at its end the computer is deleted, and its billing ends there.
A prepay always leaves an hour of your other hourly computers that are not prepaid: running ones at their running rate, stopped ones at their stopped rate. When your credit cannot cover the prepay plus that hour, the create returns 402 and makes nothing:
needed_cents includes keep_cents, the hour they keep, and the message names both parts:
needed_cents is rounded up and available_cents down.
Example: an hourly computer for 4 hours
This works on a free account, or on a paid plan for a computer past the plan’s vCPU and RAM or past the account’s computer baseline. A computer that fits the plan returns409 FITS_PLAN: send it without billing_mode and it uses the plan, end included.
201 carries the fields shown in Response, plus the end and the prepay. For a Medium at $0.055 an hour:
GPU computers
Addgpu to the create body to get a computer with a GPU. The GPU decides which operating system the computer can run:
"windows": true for that GPU.
These rules hold for every V710, B50 and RTX PRO computer:
- It spends credits or charges a card. Only the workspace owner can create one, signed in to Orgo or with a key that has full access or Billing write. An MCP sign-in, or a key without Billing write, gets
403with codeHOURLY_NEEDS_BILLING, orGPU_MONTHLY_NEEDS_BILLINGfor a monthly purchase. - It counts toward your account’s GPU limit. One limit covers every V710, B50 and RTX PRO computer, by the hour or monthly. At the limit the create returns
409with codeGPU_LIMIT, carryingusedandmax. Delete one, or ask Orgo to raise it. - It starts blank. A
template_refis refused.
--gpu flag takes only the legacy 2q and 4q. To create a V710, B50 or RTX PRO computer, call this endpoint.
V710 and B50
A V710 or B50 computer has one GPU slice: a sixth of an AMD Radeon PRO V710 card, with 4.2 GB of VRAM, or a quarter of an Intel Arc Pro B50 card, with 4 GB. The computer sees the slice as a GPU of its own. They are for an account that Get GPU options answers. For any other account the create returns403 with code GPU_NOT_ENABLED.
Size. One GPU in every size. Pick the size with ram:
CONTACT_US_FOR_SIZE.
With neither ram nor cpu you get a Medium. cpu is optional. When you send it, it must be that size’s vCPU. Leave out disk_size_gb, or send the disk in the table for your OS. Anything else, or slices other than 1, returns 400 with code GPU_SIZE_OUT_OF_RANGE.
OS. "os": "linux" is the default. "os": "windows", exactly in lowercase, works where Get GPU options answers "windows": true for that GPU. A Windows GPU computer has the Windows license added to its price, and at least 64 GB of disk. Any other os, "Windows" included, returns 400 with code GPU_REQUIRES_LINUX. So does "windows" where Windows GPU computers are off. Where they are on, but that GPU can’t take a Windows computer for you now (its windows_free_slices in Get GPU options is 0), the create returns 409 with code GPU_POOL_FULL.
Billing. Send billing_mode:
"hourly": runs by the hour from your credits. Running, it costs its monthly price times 1.35, spread over 730 hours: 35% more than monthly. Stopped, only its disk is charged, at the Linux stopped rate. Send anIdempotency-Keyheader so a retry launches it once.ends_in_hoursworks as on any hourly computer, and prepays the run."monthly": only where Get GPU options answers"monthly_billed": true. It buys the computer for the month at its monthly price, paid from your credits first, then the card on file, with one month minimum. It has no end, soends_in_hoursreturns400with codeGPU_MONTHLY_NO_END. To end it, cancel it in Account → Usage: it keeps running until its paid month ends, then it is stopped and deleted 7 days later. Send acompute_request_idor anIdempotency-Keyso a retry buys it once. Wheremonthly_billedisfalse,"monthly"returns400with codeGPU_HOURLY_ONLY.- Omitted, or
"plan": refused, and nothing is charged. Where monthly is billed,400with codeGPU_BILLING_MODE_REQUIRED, carryingmonthly_usdandhourly_usd. Elsewhere,400with codeGPU_HOURLY_ONLY. The one exception is an account an agent made with Sign up an agent: there, omitted means"hourly".
201 carries the computer’s own price.
RTX PRO by the hour
An RTX PRO 4500 or 6000 is a whole NVIDIA card, rented for you when you create it. It runs Linux only.- Body.
workspace_id,name,"gpu": {"model": "rtx-pro-4500"}or{"model": "rtx-pro-6000"}, and"billing_mode": "hourly". You must own the workspace inworkspace_id. - Where it lands. Not in
workspace_id, which is only checked to be yours. The computer goes into your account’s GPU workspace: a workspace of its own that holds your GPU hardware, made for you when you have none. Find the computer by thedesktop_idthe poll returns. - OS. Leave out
os, or send"linux". Anything else returns400with codeGPU_REQUIRES_LINUX:An RTX PRO computer runs Linux.For Windows with a GPU, use a V710 or a B50. - Billing.
"hourly"only. By the month an RTX PRO is a Get one purchase rather than a create, so"monthly"returns400with codeRTX_MONTHLY_IS_GET_ONE, and itsnextnames the call. Any otherbilling_mode, or none, returns400with codeGPU_BILLING_MODE_REQUIRED. - No end.
ends_in_hoursoron_endreturns400with codeRTX_NO_END. Delete the computer to stop paying. - Account key. A key for one workspace returns
403with codeRTX_NEEDS_ACCOUNT_KEY. - Price. $0.97 an hour for an RTX PRO 4500, $3.36 an hour for an RTX PRO 6000. It is charged by the minute from credits you bought, from when it is ready until you delete it. Starting needs an hour of bought credit, or the create returns
402with codeINSUFFICIENT_COMPUTE_CREDIT. If your credits run out, the card is given back and the computer is deleted, with everything on it.GET /api/computers/rtx-hourlylists the models you can rent, withhourly_microsandstock, and says"enabled": falsewhere RTX by the hour is not open to your account.
202, not 201: the card is being set up. The body has request_id, status: "starting" and id: null, and the Location header and next name the call to poll. Poll it until phase is ready, when desktop_id is the computer’s UUID, or failed, when error says why. Setup takes a few minutes. A rental polled later can also be ended: it ran, then was given back, such as after a delete or when your credit ran out. ended_reason says why.
Send a compute_request_id (a UUID) or an Idempotency-Key header. The same id again answers with what the first request did, with replayed: true, and never rents a second card.
Send "dry_run": true to make every check (credit, the GPU limit, stock) and rent nothing. The answer is 200, with ok, dry_run: true, request_id, product_id, hourly_micros, stock (low, medium or high, or null when it can’t be read), available_micros (your bought credit, in millionths of a dollar), gpus (your GPU computers: used and max), and name when you sent one. A dry run with a request id that already rented answers 200 with that rental’s poll body and replayed: true.
Example: a Windows V710 by the hour
201 carries the fields shown in Response, plus the GPU and its price:
os. For a B50, send {"model": "b50"}.
Example: a Linux RTX PRO 4500 by the hour
202:
next URL, where $REQUEST_ID is the request_id above:
phase is starting and step is renting, starting or connecting. Once it is ready, use desktop_id with Get computer.
GPU errors
Refunds when a monthly GPU computer does not start
A monthly V710 or B50 is paid for before its computer starts. When the computer then can’t start, the payment is refunded in full and the answer carriesrefunded: true: 409 with code GPU_POOL_FULL or name_taken, or 503 with the launch’s own code, such as BACKEND_UNAVAILABLE. When the refund has not gone through yet, the answer is 502 with code gpu_monthly_refund_pending instead. A retry of a refunded purchase with the same request id returns 410 with code gpu_monthly_launch_failed: send a new request id.
Errors
Every error body carrieserror. Most also carry a machine-readable code, and some carry extra fields, listed below.
UPGRADE_REQUIRED, VM_SLOT_ADDON, RAM_ADDON, VCPU_ADDON, CONTACT_US_FOR_SIZE, PER_COMPUTER_RAM_CAP, PER_COMPUTER_CPU_CAP, CHANGE_PLAN, PLAN_LIMIT, and the code-less grandfathered case) also carry canManageCapacity, which tells you whether you are the workspace owner and can raise the limit yourself. Where a plan change would fix it, they also carry upgradeTier. WINDOWS_REQUIRES_SCALE carries upgradeTier alone. RAM_ADDON and VCPU_ADDON also carry hourly_available: true when the workspace owner, with a credential that may spend credits, can launch the same computer by the hour instead with "billing_mode": "hourly". So does VM_SLOT_ADDON, past the baseline, when the owner’s credit covers launching it by the hour.
A 402, 403, 409 or 503 for credit, the computer limit, the pools, the plan, a size past XL, a taken name, an unbuilt template or FREE_HOURLY_BUSY also carries retryable (whether the same request can work later as it is) and docs_url. Most also carry next, the call that fixes it, as { "method", "url", "why" } with a body when the call needs one. When that call adds credit, the body also has add_credit_url, the page where a person adds it.
Unavailable operating systems
An accepted OS value can still be unavailable in your workspace. When no eligible host supports it, create returns422 with code OS_UNAVAILABLE. Failed provisioning removes its placeholder record, so you can retry the name. Temporary capacity failures remain retryable service errors.
The 300 GB disk cap is a product maximum, not an included allowance. Orgo can raise it for one account, up to 1,000 GB; disk past your plan’s own is still bought storage. Current Scale includes a 150 GB per-computer ceiling; unused storage add-ons can raise it to 300 GB. Complimentary current Scale has the same resource limits as paid current Scale.Authorizations
Headers
For an hourly launch: a UUID, or any printable key of 1 to 255 characters such as job-42. A key that is not a UUID names one launch per UTC day. A retry with the same key launches nothing new and returns the first outcome. The same key with another name, size, OS or workspace returns 409 (HOURLY_REQUEST_CONFLICT).
1 - 255Body
Operating system. Omitted, you get linux. Any other value returns 400. windows needs a Scale plan or a Windows computer bought on your plan, otherwise it returns 403 with code WINDOWS_REQUIRES_SCALE. ios is a physical handset rather than a virtual computer. Every value other than linux only succeeds where the fleet has a host that supports it, otherwise create returns 422 with code OS_UNAVAILABLE. With a gpu, the GPU decides which os works: a V710 or B50 runs linux, or windows where GET /computers/gpu-options says so, with the license in its own price; an RTX PRO and the legacy 2q/4q run linux only.
linux, windows, macos, android, ios vCPU cores. Omitted, it follows the RAM: 4 GB gets 1, 8 GB gets 2, 16 GB gets 4, 32 GB gets 6, 64 GB gets 8. With a template_ref, the template's vCPU, never below that. 8 returns 403 (CONTACT_US_FOR_SIZE) unless Orgo has set your account up for computers past XL. 16 is above the 8 vCPU per-computer cap and returns 403 (PER_COMPUTER_CPU_CAP). Above what the account vCPU pool has left returns 403 (VCPU_ADDON). 0 is read as 1. Ignored for ios. With a V710 or B50 gpu, cpu is 2, 4, 6 or 8, the vCPU of one of the GPU sizes, and 8 is never refused with CONTACT_US_FOR_SIZE.
0.5, 1, 2, 4, 6, 8, 16 RAM in GB. Omitted, you get 4, or 10 for macos: a create from an API key, the SDKs, the CLI or MCP that names neither ram nor cpu (and no template_ref) gets 4 GB and 1 vCPU. The same create from the dashboard gets a Medium (8 GB, 2 vCPU) when the plan has room for one, and a computer launched by the hour with no size is a Medium. A macos computer is always 10: any other value returns 400 with code MACOS_RAM_FIXED. 10 with any other os returns 400. 64 returns 403 (CONTACT_US_FOR_SIZE) unless Orgo has set your account up for computers past XL. Capped by the workspace owner's plan. With a V710 or B50 gpu, ram picks one of the GPU sizes instead: 8, 16, 32 or 64, and a Medium (16) when neither ram nor cpu is sent. 64 is a GPU XL, never refused with CONTACT_US_FOR_SIZE.
4, 8, 10, 16, 32, 64 Disk size in GB. Omit to use template sizing or the owner plan birth size (40 GB on current paid plans). Your plan and unused storage add-ons determine the ceiling, up to 300 GB on current plans, unless Orgo raised the account's disk limit (at most 1000 GB; GET /billing/compute-limits disk_ceiling_gb). A 150 GB plan ceiling does not include a free 300 GB disk.
Where a Linux or Windows computer runs. us (Auto (US), the default): any US city with room. eu: any city in Europe, London today. apac: any city in Asia Pacific, Sydney today. A city code runs it in that city only: nyc New York, ash Ashburn, mia Miami, chi Chicago, dal Dallas, lax Los Angeles, sjc San Jose, lon London or syd Sydney. Case does not matter, and full names such as london or sydney work too. Omitted, null or "" is the same as us. The cities with a server now, and their room, are in locations from GET /billing/compute-limits?workspace_id=. The computer keeps its location: a start after a stop, a restore and a wake bring it back there, and copies of it keep it. With no room there, the create returns 503 with code REGION_UNAVAILABLE and never runs it somewhere else. A word Orgo does not know returns 400 with code INVALID_REGION, and so does any location with os macos, android or ios. An older three letter code that is not a city, such as dfw, or a comma separated list, is read as no choice.
us, eu, apac, nyc, ash, mia, chi, dal, lax, sjc, lon, syd, null Adds a GPU. {"model": "v710"} (AMD Radeon PRO V710, one slice, 4.2 GB of VRAM) or {"model": "b50"} (Intel Arc Pro B50, one slice, 4 GB): Linux, or os "windows" (exactly lowercase) where GET /computers/gpu-options answers windows: true for that GPU, with the Windows license added and at least 64 GB of disk. One GPU in every size, picked by ram: 8 Small (2 vCPU, 40 GB disk), 16 Medium (4 vCPU, 80 GB, the default), 32 Large (6 vCPU, 150 GB) or 64 XL (8 vCPU, 300 GB). slices is 1 or left out. Needs billing_mode "hourly", or "monthly" where gpu-options answers monthly_billed: true. No template_ref. For an account gpu-options refuses, 403 (GPU_NOT_ENABLED). {"model": "rtx-pro-4500"} (32 GB of VRAM) or {"model": "rtx-pro-6000"} (96 GB), or the bare string: a whole NVIDIA card, Linux only, with billing_mode "hourly" only, no end, from an account key. It answers 202 with request_id: poll GET /computers/rtx-hourly?request_id= until phase is ready. For Windows with a GPU, use a V710 or a B50. The legacy "2q" (2 GB VRAM) and "4q" (4 GB VRAM) are Linux only and never by the hour; if no GPU host can take one, the create returns 503 (GPU_HOST_UNAVAILABLE) or makes a CPU computer, and the 201 carries gpu_downgraded and warning. "none", "", false, and "false" mean no GPU on the plan. Any other value returns 400. A non-Linux os with an RTX PRO or a legacy string, or an os other than "linux" or "windows" with a V710 or B50, returns 400 (GPU_REQUIRES_LINUX). Every V710, B50 and RTX PRO counts toward the account's GPU limit: 409 (GPU_LIMIT, with used and max).
Launch from a template's golden snapshot instead of a base image. Format namespace/name@version, e.g. system/claude-code@1.0.0 (curated) or default/my-template@1.0.0 (your own). The hardware fields above override the template's defaults. The template's build must be ready.
"system/claude-code@1.0.0"
How the computer is paid for. plan counts against the owner's plan, and monthly is the same as plan, except for a V710 or B50 gpu, where it buys that computer for the month (where GET /computers/gpu-options answers monthly_billed: true, else 400 GPU_HOURLY_ONLY): credits first, then the card on file, one month minimum, no end. A V710, B50 or RTX PRO computer needs a billing_mode and is not moved to the hour on its own: omitted or plan returns 400 (GPU_BILLING_MODE_REQUIRED or GPU_HOURLY_ONLY). The one exception is an account an agent made (POST /v1/agents), where a V710 or B50 with no billing_mode runs by the hour from its credits. An RTX PRO always needs one, and takes hourly only (monthly returns 400 RTX_MONTHLY_IS_GET_ONE). hourly runs it by the hour from credits. Any other value returns 400 (INVALID_BILLING_MODE). Omitted, the create uses the plan, except as below. Free accounts run hourly computers, a Medium only. On a paid plan the plan comes first: a computer that fits the plan's vCPU and RAM uses the plan, and hourly for it returns 409 (FITS_PLAN); one past them can run by the hour, and a plan create refused for that room carries hourly_available: true. Hourly is for the workspace owner only, Linux or Windows, with no legacy gpu string, image or snapshot, and only a starter template_ref. A V710 or B50 by the hour is never refused with FITS_PLAN. On a paid account it needs a person signed in or a key with full access or Billing write, else 403 (HOURLY_NEEDS_BILLING). Without a quote, Orgo quotes the size ram and cpu name at its listed hourly price, or a Medium when they name none; on a paid account a size no hourly offer has returns 400 (HOURLY_SIZE_UNAVAILABLE). Send an Idempotency-Key header (a UUID, or any printable key of 1 to 255 characters such as job-42, which names one launch per UTC day) so a retry returns the same computer. Without enough credit it returns 402 (insufficient_credit). Omitted means hourly for an unclaimed agent account, for an account Orgo has set to pay for every computer from credits when the create is one hourly can take, and for a Free account's create from an API key or an MCP sign-in that hourly can take (instead of 403 UPGRADE_REQUIRED). At the account's computer limit (50 on every plan, plan and hourly computers together), a create with no billing_mode is launched by the hour when the credit covers an hour of every hourly computer, the new one included, and hourly can take the computer; plan there returns 403 (VM_SLOT_ADDON).
plan, monthly, hourly When the computer ends, in whole hours after the create: 1 to 720 (30 days). Omitted or null: it never ends. Linux and Windows only, else 400 (ENDS_NOT_SUPPORTED); any other value returns 400 (INVALID_ENDS). When the time is up Orgo deletes the computer, running or not, with everything on it. To keep a computer, leave it out. A computer on the plan prepays nothing. An hourly computer with an end prepays its whole run from credits at launch, always leaving an hour of your other hourly computers that are not prepaid; when the credit is short the create returns 402 with needed_cents and available_cents and makes nothing.
1 <= x <= 7204
Not a choice: a computer with an end is deleted when its time is up, running or not, with everything on it. "delete" is accepted and changes nothing, with an end or without; omitted or null is the same. Any other value, "stop" included, returns 400 (INVALID_ENDS): A computer with an end is deleted when its time is up. To keep a computer, leave out the end.
delete, null With billing_mode: "hourly": a UUID naming this launch, so a retry with it reads the first outcome. Omitted, the Idempotency-Key names the launch, else it gets a new one. Sent without hourly, returns 400 (BILLING_MODE_REQUIRED). A monthly V710 or B50 purchase takes it too, and so does an RTX PRO rental. There, a value that is not a UUID returns 400 (COMPUTE_REQUEST_ID_INVALID).
Response
An RTX PRO create with "dry_run": true: every check (credit, the GPU limit, stock) was made, and nothing was rented. A dry run with a request id that already rented answers with that rental's poll body and replayed: true instead.