# Limits and workarounds

> Limits that apply to every Nevia SDK in early access, and where to find each adapter's own limits

> Nevia is in early access for selected teams. The CLI installs with `curl -fsSL https://get.nevia.cloud | sh`, and the MCP server is at https://api.nevia.cloud/mcp. A person signs in once in the browser.
> Setup, sign-in and requesting access: [SKILL.md](https://nevia.cloud/SKILL.md). When Nevia fits and when it doesn't: [why.md](https://nevia.cloud/why.md).

> For the complete documentation index, see [llms.txt](https://nevia.cloud/llms.txt). This page's HTML version: https://nevia.cloud/docs/reference/sdk/limits

These limits apply to all four SDKs. Most come from the API the SDKs call, so
they also apply to the CLI and MCP. Each adapter has further limits, listed on
its own page.

## Exec

* **About 50 seconds per command through the API.** Exec runs through the API
  by default, which ends a request after about 50 seconds. Use direct exec for
  longer commands. It sends the command straight to the computer. Each SDK page
  shows how to turn it on.
* **About 700 KB of output through the API.** The API returns the whole output
  in one answer, capped at 1 MiB, so output above about 700 KB can fail. Direct
  exec keeps the last 64 KiB of each stream instead and marks the result
  `truncated`. For large output, write it to a file on the computer and read it
  with the files methods.
* **Background children keep exec open.** Exec returns when the command's
  output streams close, not when the command exits. A command that starts a
  background job must redirect that job's output, for example
  `server > /tmp/server.log 2>&1 &`.
* **A timeout ends a few seconds late.** The SDK stops a timed-out command with
  TERM, then KILL five seconds later. Allow a few seconds beyond the timeout.
* **No reattach.** Once a call ends, there is no way to reconnect to a command
  that is still running. Cancelling or interrupting the call does not stop the
  command on the computer.
* **Commands run as the image's user.** On most public images that is root.

## Files

* **Files go through exec.** The files methods read and write in chunks over
  exec, so a large file takes many requests. Moving 100 MB takes hundreds of
  requests.
* **Only some paths survive a branch.** Files outside the persistent disk are
  not copied to a branch.
  [Computer lifecycle states](https://nevia.cloud/docs/reference/computer-lifecycle.md#persistence)
  says what persists.

## Computers

* **A checkpoint needs its source computer.** Deleting a computer also
  deletes its checkpoints. Keep the source computer for as long as you need
  them.
* **No restart.** The current build has no restart. A restart call answers
  `not_supported`.
* **Reserved environment names.** The API refuses environment variable names
  that start with `NEVIA_`.

## Errors

* **New error codes can appear.** Treat a code you don't recognise as a
  generic failure. A capability the computer doesn't have answers
  `not_supported`.

## Adapters

Each adapter page lists its own limitations: [LangGraph](https://nevia.cloud/docs/agent-frameworks/langgraph.md#limitations), [Pydantic AI](https://nevia.cloud/docs/agent-frameworks/pydantic-ai.md#limitations), [Agno](https://nevia.cloud/docs/agent-frameworks/agno.md#limitations), [Mastra](https://nevia.cloud/docs/agent-frameworks/mastra.md#limitations), [Eve](https://nevia.cloud/docs/agent-frameworks/eve.md#limitations), [Flue](https://nevia.cloud/docs/agent-frameworks/flue.md#limitations), [Rig](https://nevia.cloud/docs/agent-frameworks/rig.md#limitations) and [Alloy](https://nevia.cloud/docs/agent-frameworks/alloy.md#limitations).
