Use this page to find a capability and its setup guide. Features can require
configuration or an optional build flag; their presence does not mean they are
active in your deployment. The current source version is in
Cargo.toml.
| Capability | What it does | Setup and limits |
|---|---|---|
| Routing | Select a logical model using configured rules and complexity tiers | Routing order |
| Provider fallback | Try the model's configured providers in priority order | Add a fallback |
| Circuit breakers | Temporarily skip a provider after repeated failures | Security reference |
| Adaptive scoring | Rank providers using observed latency and success | Scoring configuration; opt-in, scores are not persisted |
| Context-window guard | Estimate request size before dispatch and return compact hints | Error reference |
| Response cache | Reuse eligible non-streaming responses | Caching conditions |
| Fan-out | Send to several providers and select a response | Selection and accounting limits |
| Region filtering | Filter provider mappings by configured region labels | GDPR mode; global is an exception, labels do not prove residency |
| Client format | Endpoint | Details |
|---|---|---|
| Anthropic Messages | /v1/messages |
Compatibility |
| OpenAI Chat Completions | /v1/chat/completions |
Compatibility |
| OpenAI Responses | /v1/responses |
Compatibility |
Streaming and tool support depend on the provider and translated fields. Read the protocol fidelity matrix for preserved, translated and unsupported behavior, and conformance for tested combinations. Other HTTP endpoints are described in the OpenAPI spec.
Anthropic, OpenAI, Gemini, Vertex AI, OpenRouter and other backends have
setup recipes. An OpenAI-compatible service can use a
custom base_url. Provider credentials, supported models and account access
remain provider-specific; Grob does not grant a subscription or model entitlement.
| Protection | Scope | Guide |
|---|---|---|
| Secret rules | Built-in and custom credential patterns | Secret scanning |
| Financial identifiers | Credit cards, IBAN and BIC detection | PII configuration |
| Configured names | Replace listed names with reversible pseudonyms | Name rules |
| Prompt injection | Pattern detection on configured input/output paths | Injection controls |
| URL filtering | Detect response URLs outside the configured domain policy | URL exfiltration |
| Canary tokens | Replace selected secrets with traceable fake values | Custom rules |
| Streaming scanning | Inspect response chunks with cross-chunk state | DLP reference |
DLP must be enabled. It detects supported patterns; it is not a guarantee that all personal data, secrets or injection attempts are removed. Generic email and phone detection is not provided by the financial-identifier scanner.
Policies apply configured rules to tenants, tools and providers. The human-in-the-loop (HIT) gateway can require approval or deny matched tool actions. Audit exports can use recipient-based encryption. These controls need explicit policies and working dependencies; see the configuration reference.
| Control | Configuration and boundary |
|---|---|
| Client authentication | Authentication reference; separate from provider OAuth |
| Virtual access keys | Per-identity permissions and limits; key commands |
| Rate limits | Disabled by default; configure rates and burst together |
| Stored credentials | Encryption and key custody; not all stored data is encrypted |
| Memory protection | Locking and deployment limits |
| Service credential replacement | HTTP gateway with local or Vault authority |
| Security headers and request limits | Security reference |
| CI analysis | CodeQL, Semgrep, dependency and secret checks; results apply to the checked revision |
Secret wrappers reduce accidental exposure. Explicit access to a secret, process compromise, traces and operator configuration still matter. See the threat model.
The following are technical capabilities that may contribute evidence to your organization's review. They do not establish compliance, certification, data residency or a right-to-erasure process.
Grob can add provider/model headers, enrich audit records and emit risk signals. See compliance configuration. Request risk scores are not a legal classification of your AI system.
DLP, configured-name pseudonyms and region filtering can support data controls. Traces or log exports may still contain personal data; operators must define retention, access and deletion procedures. Region labels alone do not establish where a provider processes data.
Signed audit records, access control and encrypted credential storage can support a security review. Their presence does not qualify Grob or the hosting environment.
Grob supplies technical controls, not an organizational audit or certification. Review deployment, policies and evidence with the responsible specialists.
Fallback, monitoring, audit records and escalation webhooks can support incident handling. A webhook is not an incident notification to an authority. Reporting, governance and supplier review remain organizational responsibilities.
gdpr and eu-ai-act are starting configurations. Inspect the selected providers,
DLP settings and audit destinations before applying a preset. See
preset operations and the routing limits.
Grob does not certify your organization, verify a provider's residency, encrypt model input from the provider that must process it, or remove all personal data from every storage destination.
The older dated implementation table was a source-review snapshot, not continuous assurance. Use the linked behavior references, tests and CI results for the revision you deploy. A successful check does not establish regulatory compliance.
Choose incoming-client authentication separately from provider credentials. API keys issued by Grob do not replace provider credentials by themselves; the configured proxy supplies those when dispatching requests.
Virtual keys attach an identity, permitted models/providers and optional limits to client requests. See authentication for the lifecycle and replica consistency for which limits and state are shared across processes.
Observability covers metrics, live traffic, tracing and log
export. OpenTelemetry requires the otel build feature. Protect monitoring
access and decide which request content, if any, may be retained.
Operations covers presets, reloads, upgrades and timeouts. Deployment covers containers, authentication and persistent state. A standalone binary or scratch container does not require a separate database; image size and native TLS support depend on the build.
Use grob --help for your installed binary and the CLI reference for
command details. grob doctor checks local setup; grob validate makes real
provider calls and can consume quota.
The authoritative defaults are in Cargo.toml, under [features].
Optional tls, acme, otel and harness features require a build that includes
them. Compile-time availability and runtime activation are separate.
See the architecture overview for request flow and component responsibilities. Proposed designs, including the optional supervisor, are not current configuration options.