Policy and Limits

policy bounds how a capability runs: how long it may take, how much it may read and return, and how often it may be called. Set common values once in the top-level defaults section and override them per capability:


defaults:
  timeout: 10s
  max_output_bytes: 1MB
  rate_limit: {requests: 60, per: 1m, burst: 10}
 
capabilities:
  get_customer:
    policy:
      timeout: 3s
      max_rows: 1

A value in defaults only applies to capabilities where it makes sense; max_rows in defaults does not affect an HTTP capability. Setting a key directly in a capability’s policy where it does not apply is a configuration error.

Settings

Key	Default	Applies to	Meaning
`timeout`	`5s`	all	Time budget for execution and output validation, up to `300s`
`readonly`	`true`	DB	`true` allows only `SELECT`. Set `false` for `INSERT`, `UPDATE`, or `DELETE`
`max_rows`	`100`	DB `SELECT`, CSV	DB results are truncated at this count. CSV fails with `LIMIT_EXCEEDED` if more rows match
`max_affected_rows`	none	DB mutations	If more rows are affected, the transaction is rolled back and the call fails with `LIMIT_EXCEEDED`
`max_input_bytes`	`1MB`	all	Size of the JSON input
`max_output_bytes`	`1MB`	capabilities that publish data or generate CSV	Size of the published JSON, and of the data a generated CSV is built from
`max_json_depth`	`32`	all	Nesting depth of input and output JSON
`max_json_items`	`100000`	all	Total number of values in input and output JSON
`rate_limit`	none	all	`{requests, per, burst}`. See below
`expose_in_openapi`	`true`	all	Whether the capability appears in OpenAPI and the MCP tool list

Sizes accept plain byte counts or units such as 256KB, 8MiB, or 1GB; KB and KiB both mean 1024 bytes. Every limit must be a positive, finite value; 0 does not mean unlimited.

File limits

These apply only to capabilities that accept or return files:

Key	Default	Meaning
`max_input_files`	`16`	Number of uploaded files per call
`max_input_file_bytes`	`32MiB`	Size of each uploaded file. A file input’s `max_bytes` can lower it
`max_total_input_file_bytes`	`64MiB`	Combined size of uploaded files per call
`max_output_files`	`16`	Number of returned files per call
`max_output_file_bytes`	`32MiB`	Size of each returned file
`max_total_output_file_bytes`	`64MiB`	Combined size of returned files per call
`file_ttl`	`15m`	How long download links and retained files stay valid
`mcp_inline_total_bytes`	`1MiB`	Size of an MCP tool result with `delivery.mcp: inline`, including Base64 encoding

The gateway has its own request size limits for uploads; see Gateway configuration. Uploads and returned files are staged in the agent’s spool, which has a total quota set by runtime.spool_total_bytes; see Agent configuration.

Rate limits

rate_limit gives each capability a token bucket shared by all its callers. The bucket starts with burst tokens and refills at requests per per:


policy:
  rate_limit: {requests: 10, per: 1s, burst: 20}

A call that finds the bucket empty fails with RATE_LIMITED before it runs. A capability takes the whole rate_limit object from defaults unless it sets its own. Buckets survive gateway reconnects and reset when the agent restarts.

This limit is enforced by the agent, so it holds even if the gateway’s own per-IP limit is bypassed. The agent also caps concurrent executions with runtime.max_concurrent_requests; calls beyond that and its waiting queue fail with BUSY.

Timeouts

timeout covers the run itself and checking its output. A caller that disconnects or sets a shorter deadline ends the call earlier. Receiving uploaded files and sending large downloads are bounded separately by runtime.transfer_timeout, so a client or proxy in front of the gateway needs a timeout that covers both; see Deployment.

If a mutation, HTTP call, or code capability times out after it started, the result may be unknown. Read Uncertain outcomes before retrying.