Limits and Configuration
This page lists the limits and settings that affect how you integrate with and operate OpenProdkt. For each one it states the value and unit, who can change it, and what happens at the boundary.
How to read it:
- Enforced defaults are not service-level commitments. A value here describes what the software enforces. It is not an availability, latency, or capacity guarantee, and no throughput is promised. Latency objectives, and how far each has been validated, are on Performance Objectives; they are objectives, not commitments.
- Fixed means the value is built in and no customer or operator setting changes it.
- Plan means the value comes from the organization's entitlement, and a signed license can override it.
- Operator means a self-hosted operator sets it with an environment variable and a restart.
- Hosted-service infrastructure settings (gateway throttles, instance sizes) are not customer configuration and are not listed.
Plan quotas and rate limits
Built-in plan profiles
Entitlements define request-rate budgets and monthly quotas. The built-in profiles are below. A signed license (self-hosted Enterprise) can replace any of these values. Confirm the profile and any license override for your organization with your account team, because plan names and packaging can change.
| Profile | Requests per second (integration) | Decisions per month | Approvals per month | Notifications per month | Seats |
|---|---|---|---|---|---|
| Sandbox (hosted partner trial or self-hosted trial, 30 days) | 25 | 500,000 | 500,000 | 1,000,000 | 3 |
| Free | 1 | 5,000 | 5,000 | 10,000 | 1 |
| Business | 25 | 500,000 | 500,000 | 1,000,000 | 25 |
| Enterprise | 1,000 | No cap | No cap | No cap | 25 |
The approval quota equals the decision quota. The notification quota is twice the approval quota. A notification is an approval lifecycle event sent through email or webhook.
Dashboard budgets are separate from integration budgets. Defaults (provisional, and subject to change) are 60 requests per second for dashboard reads per organization and 20 requests per second for any single dashboard endpoint. Both are plan values and a license can override them.
Request-rate limits
- Scope. Per organization, never per key or user. Adding keys does not raise a limit.
- Window. One second. Counters are shared across all API replicas through the database.
- Budgets. Integration (
requests.per_second) covers API-key requests and JWT requests other thanGET. Dashboard (dashboard.requests_per_second) covers JWTGETrequests. A per-endpoint cap (dashboard.endpoint_requests_per_second) applies inside the dashboard budget. The budgets do not draw from each other. - At the limit. HTTP
429. Retrying later is safe; nothing was evaluated. - If the counter store is unavailable. HTTP
503. Requests are refused rather than allowed. - Zero. A license value of
0is a limit of zero and blocks all requests in that class. It does not mean unlimited. Negative values and unknown limit names are rejected when a license is loaded.
Monthly quotas
- Scope and period. Per organization, per calendar month in UTC. Usage resets at the start of the month.
- What counts. One decision unit per recorded billable decision, including a policy-evaluation failure. Not counted: unknown assets (
asset_not_registered), assets with no policies (no_policies_assigned), duplicate requests for the same pending operation, invalid input rejected before evaluation, and dashboard reads. Approvals and notifications are counted when created. - At the limit. HTTP
402withdecision_quota_exceeded,approval_quota_exceeded, ornotification_quota_exceeded. The call is not evaluated and the upstream is not called by the control plane. A retry fails until the month changes or the plan changes. - Atomic. Usage is reserved atomically with the decision, so concurrent requests cannot exceed the quota.
- Playground. Console playground evaluations of registered assets use the same decision quota.
- Read usage.
GET /v1/usagereturns consumption and limits. - Unlimited. A plan with no practical monthly limit reports the largest integer JavaScript represents exactly,
9007199254740991(2^53-1), in thelimitsofGET /v1/usageand in the quota fields ofGET /settings. Treat a value at that number as unlimited.
Licensing
Signed licenses are verified against the deployment's public key and cached for 1 minute. A license can set the limit names above, members.seats, and feature flags. Unknown names are rejected. Check Plans & Licensing for modules and delivery.
API request limits
| Item | Value | Notes |
|---|---|---|
List page size limit | 1 to 100, default 25 | Outside the range returns 400. Applies to audit, approvals, API keys, and other list endpoints that document it. |
offset | 0 or greater, default 0 | Negative returns 400. |
| Decision metadata size | 8 KiB of serialized JSON | Larger returns 400. See Policy Engine. |
| Known metadata string keys | 256 characters | provider, model, operation, environment, actor, application. |
| Metadata and policy numeric precision | 128 significant decimal digits | Fixed, not customer-configurable. Excess precision is rejected, never rounded. |
| Normalized numeric decimal exponent | -308 through 308, inclusive | Fixed, not customer-configurable. Out-of-range values are rejected. See exact numbers. |
| Numeric token length | 4,096 characters | Includes sign, decimal point, and exponent; fixed, not customer-configurable. |
| Metadata / individual policy-value depth | 64 | Root depth 0; add one per property/array element, including scalar leaves. Fixed; not a whole-request or composite-policy-tree limit. |
| AI metadata counters | Integral values in the client/runtime range | Native Go int, Python signed64, TypeScript safe integer; fractions, numeric strings, and overflow are rejected without rounding. No decision usage or upstream execution on invalid context. |
metadata.ai.attachment_count | 64 | Maximum value. |
operation_id | 1 to 128 characters of letters, digits, ., _, :, - | Evaluate requires only non-empty; status, audit filter, and proxy challenge enforce the pattern. |
traceparent request header | At most 512 characters, lowercase hex, W3C layout | Optional on evaluate and proxy evaluate. Invalid, repeated or comma-joined values are ignored and start a new trace; they are not an error. Fixed. See Trace context. |
Outcome execution_id | Lowercase UUIDv7 | Optional. A malformed or null value returns 400 invalid_request and stores nothing. Fixed. |
Outcome attempt | Integer 1 to 2147483647 | Optional. Out-of-range, fractional and non-integer values return 400 invalid_request. Fixed. |
Outcome upstream_request_id | 1 to 128 characters of letters, digits, ., _, :, - | Optional. Fixed. |
Asset ai_inventory | At most 2 KiB encoded; purpose at most 512 UTF-8 bytes; references at most 256 characters | Optional. Invalid, null, empty or unknown members return 400. Fixed. See AI inventory on assets. |
| API-key lifetime | Up to 365 days from creation or rotation | A later expiry is refused. Keys must be rotated at least yearly. |
| Total request body | No documented limit | Keep requests small. Not a guarantee of acceptance. |
Approvals and grants
| Item | Value | Changeable |
|---|---|---|
| Pending approval expiry | 24 hours | No |
| Grant window after approval | 2 minutes (120 seconds) | No |
Proxy Retry-After | 30 seconds | No |
| Operation claim lease | 5 minutes | No |
| Outcome reports per operation | First report wins | No |
Details and retry behavior are in Approvals.
Enforcement controls
| Item | Value | Changeable |
|---|---|---|
| Enforcement mode default | permissive for every organization | By a super_admin |
| Mode overrides | At most 50 per organization, shared by the environment, asset type and risk tier dimensions; value 1 to 128 characters for environment and asset type | No |
| Definition pin observation | At most 512 bytes; server_version at most 128 characters; at most 20 distinct observed digests recorded per asset and pin | No |
| Shadow report window | At most 90 days (default 30); at most 250,000 events per request; 5 second query limit | No |
| Kill switch reason | 1 to 200 characters | No |
| Active kill switches | One for the organization, one per asset, at most 100 per-asset switches | No |
| Break-glass duration | 1 to 240 minutes | No |
| Break-glass pending window | 60 minutes to decide a request | No |
| Break-glass grants | At most 5 active and 10 pending per organization | No |
| Break-glass reason and note | 1 to 200 characters (reason), at most 200 (note) | No |
Details are in Enforcement Modes, Kill Switch and Break-Glass. These are enforced limits, not a throughput guarantee.
Sign-in and sessions
| Item | Value |
|---|---|
| Sign-in code lifetime | 5 minutes |
| Minimum wait between code requests | 30 seconds |
| Code request window and maximum | 5 requests per 10 minutes |
| Code verification attempts | 5 |
| Session idle timeout | 30 minutes |
| Session maximum lifetime | 12 hours |
| Denied registration cooldown | 7 days |
These are fixed.
Integration limits
SDKs
| Item | Go | Python and TypeScript |
|---|---|---|
| Request timeout | 5 seconds default | See SDK Manual |
| Webhook signature replay window | 5 minutes | 300 seconds |
| Webhook body size read by the verifier | 256 KiB | See SDK Manual |
Approval wait and polling bounds are in the SDK Manual.
MCP stdio wrapper
| Setting | Default | Range and behavior |
|---|---|---|
pdp_timeout | 5 seconds | Go duration. Bounds each decision call and each outcome report attempt. On timeout the call is denied (pdp_unavailable). |
| Outcome report attempts | 4 attempts | Waits of 250 ms, 500 ms, then 1 s between attempts. A definitive 4xx other than 408 and 429 is not retried. Not configurable. Every attempt of one report carries the same optional execution_id, attempt and upstream_request_id. |
| Message size | 1 MiB per newline-delimited message | An oversize message gets a JSON-RPC invalid-request error. Applies to the client and the child server's output. |
| Shutdown grace | 5 seconds | Fixed. |
| Principal token file | Mode 0600, regular file, at most 64 KiB | The wrapper refuses to start otherwise. |
Envoy authorization adapter
| Setting | Default | Behavior |
|---|---|---|
OPK_AUTHZ_PDP_TIMEOUT | 2 seconds | Bounds the decision call for HTTP and gRPC routes. On timeout the adapter returns 503 pdp_unavailable and the upstream does not run. |
| MCP decision timeout | 5 seconds, fixed | MCP routes ignore OPK_AUTHZ_PDP_TIMEOUT. |
OPK_AUTHZ_LISTEN_ADDRESS | :9191 | Listen address. |
| Request body buffer | 1 MiB | Used for payload binding and MCP tools/call. Larger MCP bodies return 413 payload_too_large; larger bodies on a payload-bound HTTP route are rejected as malformed. |
MCP denials and protocol errors on the Envoy path are returned as HTTP 200 JSON-RPC bodies. GET and DELETE requests carrying a body are rejected with body_not_allowed.
Webhooks
| Item | Value | Notes |
|---|---|---|
| Connect timeout | 10 seconds | Fixed. |
| Overall request timeout | 10 seconds | Fixed. |
| Redirects and proxies | Not followed or used | |
| Success | HTTP 2xx | Anything else is a failure and is retried. |
| Default event types | approval.pending, approval.approved, approval.rejected, approval.expired, approval.revoked | Used when you create an endpoint with no eventTypes. |
| Scheme | https in saas product mode; http also allowed in other modes | Set by the deployment. |
| Destination policy | Private, loopback, link-local, and reserved ranges refused; cloud metadata addresses always refused | Self-hosted operators can allow ranges with OPK_WEBHOOK_ALLOWED_CIDRS. |
| Signature | X-OpenProdkt-Signature: v1=sha256=<hex> over timestamp + "." + body | Timestamp is generated at delivery time. |
See Approvals for retry semantics and On-Premises Webhooks for the CIDR setting.
Outbox dispatch (self-hosted)
The API server dispatches outbox events (approval emails, webhooks, approval lifecycle records). Set these on the API server and restart it. Invalid, empty, or out-of-range values fall back silently to the default; check the effective value in your configuration management.
| Variable | Default | Range |
|---|---|---|
OPK_OUTBOX_DISPATCH_ENABLED | Off (unset is off) | Boolean. Any value that is not a valid true value is off. Nothing is dispatched while off. |
OPK_OUTBOX_DISPATCH_INTERVAL | 5s | Go duration greater than zero. |
OPK_OUTBOX_DISPATCH_BATCH_SIZE | 20 | 1 to 500. A larger value is capped at 500. |
OPK_OUTBOX_DISPATCH_MAX_ATTEMPTS | 5 | 1 to 50. A larger value is capped at 50. After the last attempt the event is dead-lettered. |
OPK_OUTBOX_DISPATCH_JITTER_RATIO | 0.20 | 0 to 1. A larger value is capped at 1. |
Retry backoff starts at 1 minute, doubles, and is capped at 1 hour. It is not configurable.
Audit export and retention
Set these on the worker and restart it. See Audit for what is exported.
| Variable | Default | Range and behavior |
|---|---|---|
OPK_AUDIT_EXPORT_LAG_SECONDS | 60 seconds | Minimum age of a row before export. Unset, non-numeric, or zero or negative falls back to the default. |
OPK_AUDIT_EXPORT_INTERVAL_SECONDS | 300 seconds | Time between export cycles. Same fallback. |
OPK_AUDIT_EXPORT_BATCH_SIZE | 500 events | Maximum events per export object. Same fallback. |
OPK_ARTIFACT_OBJECT_LOCK_MODE | COMPLIANCE | COMPLIANCE or GOVERNANCE. |
OPK_ARTIFACT_RETENTION_DAYS | 2555 days | Minimum 2555. In deployment-retention builds, an explicitly empty, non-integer, zero, negative or lower value stops startup rather than silently selecting a default. |
OPK_WORKER_DEDUPE_SIZE | 10,000 message IDs | In-memory duplicate filter. It is lost on restart, so duplicates are still possible. |
OPK_WORKER_ARTIFACT_PREFIX | Deployment default | Object key prefix for approval records. |
OPK_WORKER_QUEUE_TOPIC (or OPK_OUTBOX_QUEUE_TOPIC) | Required | Queue the worker reads. |
The object-storage bucket must have Object Lock enabled when it is created. It cannot be turned on later for an existing bucket, and the worker checks at startup.
Deployment retention configuration
Availability. Applies to builds containing deployment retention configuration and native storage drivers. Confirm the build with your account team; this does not announce a new release.
An operator sets one retention policy per installation, shared by every organization, approval record and exported audit object on that installation. There is no per-organization override or Console/API setting. Changes require a restart and affect new writes only; they never shorten already retained versions. The minimum remains 2555 days and operators may extend it.
The customer deployment configuration can supply the existing environment variables alongside other operational settings:
deploy:
runtime_env:
OPK_ARTIFACT_OBJECT_LOCK_MODE: COMPLIANCE
OPK_ARTIFACT_RETENTION_DAYS: "4000"
For installations loading a runtime YAML file through OPK_INFRA_CONFIG_FILE,
the equivalent section is:
artifact_retention:
mode: COMPLIANCE
days: 4000
Environment variables take precedence over the runtime file; omitted fields use COMPLIANCE and 2555 days. Explicit invalid values fail startup. S3 startup validation requires any present bucket default to be COMPLIANCE and at least the configured deployment duration; an absent default remains allowed.
Native Azure Blob and Google Cloud Storage accept COMPLIANCE only; configuring GOVERNANCE with either native driver fails startup. Their lock-enabled startup check is not a full conformance test or a default-retention check. See storage providers and validation.
Outcome deadline monitoring
Availability. These settings exist only in worker builds that include outcome deadline monitoring. Check the release notes of the build you run.
Monitoring is off by default. When it is off the worker does nothing for it: no checks are scheduled, no coverage events are stored, and the request path is unchanged. Set these on the worker and restart it; they are read once at startup. There is no Console or API setting for them.
| Variable | Default | Range and behavior |
|---|---|---|
OPK_OUTCOME_DEADLINE_SECONDS | Unset, which is off | Whole seconds. 0, empty or unset turns monitoring off. Any value from 300 upward turns it on; it is the time after an allow by which an outcome report is expected. Values from 1 to 299, negative values and non-numeric values are rejected and the worker stops at startup instead of falling back to a default. |
OPK_OUTCOME_MONITOR_FROM | None | An RFC 3339 timestamp, for example 2026-01-01T00:00:00Z. Required when monitoring is on; the worker stops at startup if it is missing or invalid. Only allows issued at or after this time are monitored. Earlier allows are never scanned or backfilled. Ignored while monitoring is off. |
Behavior when on:
- Scope. The deadline applies to every organization on the deployment and to every allowed operation after the cutover, regardless of asset or transport. It is not configurable per organization, asset or principal.
- Result. At the deadline, if no outcome has been reported, the worker stores an
execution.outcome.missingaudit event. An allow that declaredoutcome_reporting: nonegets anexecution.outcome.not_reportableevent without waiting for the deadline. See Coverage events. - Fixed once registered. An operation's deadline is set when the worker first registers it. Changing
OPK_OUTCOME_DEADLINE_SECONDSlater affects only operations registered afterwards, and restarting does not move an existing deadline. - Dependencies. Deadline checks are delivered to the worker through the same outbox dispatch and queue as other worker events. Outbox dispatch must be enabled (see Outbox dispatch) and the worker must be running for checks to complete. Checks follow the outbox retry rules. The coverage event is stored when a check is processed, which can be later than the deadline, and a check that is dead-lettered after its last attempt produces no coverage event.
- No effect on results. Monitoring does not change the result of a decision or of an outcome report, and a late report is still accepted.
Other operator settings
| Variable | Purpose |
|---|---|
OPK_PRODUCT_MODE | Deployment mode. Default saas, which requires https webhooks. |
OPK_BREAKGLASS_EXPIRY_SWEEP_SECONDS | Seconds between the worker's checks that record breakglass.expired events (see Break-Glass). Default 60; an invalid or non-positive value falls back to 60. Set on the worker. It never changes when a grant stops working. |
OPK_WEBHOOK_ALLOWED_CIDRS | Comma-separated CIDRs allowed as webhook destinations. Set on the API and the worker. |
OPK_CORS_ALLOWED_ORIGINS | Browser origins allowed to call the API. |
OPK_CONSOLE_BASE_URL | Absolute http or https Console URL for email links. No query string or fragment. |
OPK_QUEUE_PROVIDER | Queue backend: SQS, NATS, or RabbitMQ. |
OPK_DATABASE_PROVIDER, OPK_OBJECT_STORAGE_*, OPK_SECRETS_PROVIDER | Supported self-hosted dependencies. See the customer example configuration shipped with your deployment package. |
Known documentation gaps
- Whether the license limits
users.max,assets.max,policies.max,api_keys.max, andprojects.maxare enforced by the API is not confirmed, so they are not listed. - No application-level cap on total request body size is documented. Deployment gateways and proxies may impose their own.
- Plan names and prices are not stated here.