Skip to main content
This page lists API and execution bounds from the current source, not provider account quotas. MiB means 1048576 bytes; KiB means 1024 bytes. JS code units are JavaScript string-length units, not UTF-8 bytes; non-ASCII text may hit byte limits earlier. These values describe this source snapshot; check implementation when upgrading.

Requests and execution

Primary source: apps/control-plane/server.ts.

Session policy and initialization

Primary source: modules/execution/sessions.ts.

Input and exported files

Primary source: modules/execution/inputs.ts.

Preview and Harness distribution

Primary source: modules/sandboxes/previews.ts.

Cross-check the sources

The tables also use Run input/events, initialization, environment, artifacts, OCI registration, image resolution and source packages. Most schema bounds have no separate administration setting. Configurable Session timeouts do not make every implementation limit adjustable.

Registry and network restrictions

Image reference parsing accepts Docker Hub (docker.io, also the default when omitted) and GHCR (ghcr.io), not arbitrary registry URLs. The fixed parser rejects other origins or malformed references with invalid_image_reference (400). Authentication token responses are bounded to 65536 bytes; fetched manifest/configuration responses to 4 MiB. Overflow produces image_manifest_too_large (400). These are metadata-read bounds, not a 4 MiB limit on all image layers or the installed filesystem. Blob redirects also have an explicit destination policy; adding a credential does not bypass it. Provider-specific private-image support remains separate. The Session network schema permits unrestricted or allowlist, with unrestricted and an empty host list as defaults. Allowed-host entries are nonempty strings; this schema does not define a universal count or per-host length maximum. Do not infer unlimited provider support from that absence. The combination resolver rejects unsupported allowlists and missing credential-ingress access using network_policy_unsupported or credential_proxy_not_allowed_by_network_policy; creation reports the compatibility reason as 400. Select a capable Provider while preserving the required policy. These checks do not establish a live firewall enforcement result.

Tool, provider and account limits

Remote MCP currently bounds discovery to 100 tools, 20 pages and a 2 MiB response; see remote.ts. Named application response bodies are bounded to 1 MiB in the application transport. These protocol limits are not workspace file-size limits. Provider memory, CPU, creation concurrency, storage, image permissions and API quotas depend on connection configuration and the actual account. Consult the matching Sandbox guide; a trial account’s concurrency limit is not a global Hexagon quota. Private OCI, network allowlists and pause/resume require declared connector support. Unsupported policy must not be silently weakened. Some connectors impose tighter file or command bounds. A transport queue bound limits buffered data, not necessarily the total response length. Do not substitute it for an upload request, exported artifact or provider disk limit. See capacity and cost.

Calculate the size the API checks

For text, count both JavaScript UTF-16 code units and UTF-8 bytes. A task containing 41000 copies of 中 has 41000 code units but 123000 UTF-8 bytes: it fits the input field length and still exceeds the 120000-byte combined input/context limit. One 😀 occupies two code units and four UTF-8 bytes. In Python, len(text) counts Unicode code points, so it is not interchangeable with the schema’s JavaScript length. For an upload containing n binary bytes, standard Base64 uses 4 × ceil(n / 3) characters. A 1572864-byte file encodes to 2097152 characters and fits the field bound; one additional byte requires 2097156 characters and exceeds it. JSON paths, keys, quotes and separators add to the HTTP body size. Check the final serialized request as well as decoded bytes. Splitting six 1.5 MiB files into batches can address the 8 MiB request aggregate; splitting requests cannot make an oversized individual file valid. Upload batches must all finish before the first Run or Sandbox allocation. Later uploads fail with workspace_already_started; use project context to plan initial input and let subsequent Agent work modify the live workspace. Repeating a path within one request gives duplicate_input_path; before execution starts, a later request can replace an earlier path. Confirm returned size and SHA-256 rather than assuming every upload appends a new file. Environment limits count JSON.stringify(env) bytes, including keys and escaping. Initialization limits likewise include serialized structure and compiled background wrappers, so adding individually valid documents can still overflow the aggregate. The schema’s maximum for each field is not a promise that every field can be filled to its maximum simultaneously.

Locate the rejection

Artifact capture checks the manifest before its database transaction. A rejected manifest is not a partially accepted list; it does not erase previously committed artifacts. Do not equate a terminal Run with every intended file having been exported—query the recorded artifact list and verify the bytes you need.

Respond to a limit failure

Distinguish schema rejection from domain errors. Combined input/background overflow produces input_too_large (413), while initialization overflow produces initialization_too_large (413). Aggregate artifact size or duplicate names can produce invalid_artifact_manifest; single-file size, encoding or hash failures can produce artifact_integrity_failed. Repeating an identical request does not fix these conditions. Reduce input, batch uploads, export only needed files or select a suitable execution combination. Batching does not bypass per-file bounds. Do not change a required allowlist to unrestricted merely to pass validation. Network faults and provider quota refusals are separate from Hexagon input limits; reconcile an unknown create outcome before attempting another allocation. Values were source-reviewed; focused local encoding/schema evaluations are recorded in the limits review. This is not exhaustive boundary coverage or cloud load testing. Unlisted account capacity, total throughput or price is not unlimited and does not imply a service-level guarantee. Next: project context, artifacts, troubleshooting. For combination constraints, continue to Sandbox capabilities; for sizing, capacity and cost; for exact responses, error codes.