features.md

Feature Matrix

Exhaustive list of grob capabilities, extracted from the codebase. The current version is in Cargo.toml.

Core Proxy

Feature	Description	Config
Multi-provider failover	Priority-based provider chains with exponential backoff	[[models.mappings]] priority
Circuit breakers	Auto-disable unhealthy providers (Closed/Open/HalfOpen)	[security] circuit_breaker = true
Adaptive provider scoring	EWMA latency + rolling success rate ranks providers dynamically	[security] adaptive_scoring = true
Task classification routing	Intent-based: think, websearch, background, regex, default	[router] think/background/websearch
Prompt-based routing	Regex rules with capture groups route specific prompts	[[router.prompt_rules]]
Auto-mapping	Regex model name transformation (e.g., ^claude- → default)	[router] auto_map_regex
Fan-out racing	Parallel dispatch: fastest, best_quality, or weighted selection	strategy = "fan_out"
Response caching	Dedup temperature=0 requests (moka LRU, configurable TTL)	[cache] enabled = true
Streaming SSE	Full SSE streaming for all endpoints and providers	Built-in
Tool calling	Function calling support across all providers	Built-in
GDPR routing	Region-based provider filtering (EU-only)	[router] gdpr = true

API Compatibility

Endpoint	Format	Streaming	Tool Calling
/v1/messages	Anthropic native	Yes	Yes
/v1/chat/completions	OpenAI compatible	Yes	Yes
/v1/responses	OpenAI Responses API (Codex CLI)	Yes	Yes
/v1/models	OpenAI model listing	N/A	N/A
/v1/messages/count_tokens	Token counting	N/A	N/A

Providers (14+)

Provider	Type	Auth Methods
Anthropic	anthropic	API key, OAuth PKCE (Max/Pro)
OpenAI	openai	API key
Gemini	gemini	API key, OAuth PKCE (Pro)
Vertex AI	vertex-ai	Application Default Credentials
OpenRouter	openrouter	API key (200+ models)
Mistral	openai	API key (custom base_url)
Ollama	openai	None (local)
Groq	openai	API key
DeepSeek	openai	API key
Together	openai	API key
z.ai	z.ai	API key
MiniMax	minimax	API key
Kimi Coding	kimi-coding	API key
Zenmux	zenmux	API key
Any OpenAI-compatible	openai	API key + custom base_url

DLP (Data Loss Prevention)

Scan Type	Description	Actions	Direction
Secret scanning	API keys, tokens, PEM blocks, credentials (25 built-in rules + custom)	redact, block, warn	Request + Response
PII detection	Email, phone, credit card (Luhn), IBAN/BIC	redact, block, warn	Request + Response
Name pseudonymization	Reversible mapping (real names → consistent pseudonyms)	pseudonymize	Request (anonymize) / Response (de-anonymize)
Prompt injection	Pattern-based detection (custom + built-in)	block, warn	Request
URL exfiltration	Anti-EchoLeak: domain whitelist/blacklist filtering	block, warn	Response
Canary tokens	Watermark redacted secrets for leak traceability	inject	Request
Streaming DLP	Per-chunk SSE scanning with SPRT cross-boundary detection	redact, block	Streaming responses
Entropy detection	SPRT-based high-entropy content identification	flag	Streaming

Additional DLP features:

• Per-session context (multi-turn conversation tracking)
• Hot-reloadable domain lists and rules
• Signed config verification (Merkle tree integrity)
• Homoglyph attack prevention (Unicode normalization)
• External rules files (TOML, hot-reloadable)

Policy Engine

Feature	Description	Config
Policy engine	Unified per-tenant/zone/compliance policy evaluation with glob matching	[[policies]]
HIT Gateway	Per-action human authorization for tool_use blocks (auto-approve, require-approval, deny)	[policies.hit]
Encrypted audit export	Age envelope encryption with per-policy auditor matrix	[policies.log_export] content = "encrypted"

Security

Feature	Description	Config
Rate limiting	Per-tenant token bucket (RPS + burst)	[security] rate_limit_rps
Circuit breakers	Per-provider failure tracking (Closed/Open/HalfOpen)	[security] circuit_breaker
OWASP security headers	HSTS, CSP, X-Frame-Options, X-Content-Type-Options, Referrer-Policy, Permissions-Policy	[security] security_headers
Constant-time auth	Timing-attack-safe API key comparison (subtle crate)	Built-in
Credentials encryption	AES-256-GCM at rest for OAuth tokens and virtual keys	Automatic
Cross-platform permissions	Owner-only file permissions (Unix 0o600 + Windows DACL)	Automatic
SecretString wrapping	API keys wrapped in secrecy::SecretString — impossible to log accidentally	Built-in
CodeQL analysis	Automated security scanning (CI + local)	.github/workflows/codeql.yml

Regulatory Compliance

Grob maps its features to specific regulatory requirements. The table below shows which grob capability satisfies which article/control.

EU AI Act

Article	Requirement	Grob Feature	Config
Art. 12	Record-keeping (logging of AI system usage)	Signed audit log with model name, token counts, timestamps, hash chain	[compliance] audit_model_name = true, audit_token_counts = true
Art. 14	Human oversight (risk classification)	Per-request risk scoring with escalation webhook	[compliance] risk_classification = true, escalation_webhook
Art. 15	Accuracy, robustness, cybersecurity	Prompt injection detection, DLP, circuit breakers	[dlp] injection = "block"
Art. 52	Transparency obligations	X-AI-Provider, X-AI-Model, X-AI-Generated response headers	[compliance] transparency_headers = true

Preset: grob preset apply eu-ai-act enables everything in one command.

GDPR / RGPD

Requirement	Grob Feature	Config
Data minimization (Art. 5)	DLP redacts PII before it reaches the LLM provider	[dlp] pii = "redact"
Data residency (Art. 44-49)	GDPR routing mode restricts to EU-only providers	[router] gdpr = true, region = "eu"
Pseudonymization (Art. 4)	Name pseudonymization with reversible mapping	[dlp] names = "pseudonymize"
Right to erasure (Art. 17)	No PII stored — redacted before leaving the proxy	[dlp] pii = "redact"
Data breach notification (Art. 33)	Canary tokens detect data leaks	[dlp] canary = true
Audit trail (Art. 30)	Signed audit log with hash chain	[security] audit_dir

Preset: grob preset apply gdpr enables EU-only routing + DLP.

HDS / PCI DSS / SecNumCloud

Control	Grob Feature	Code Reference
Audit trail integrity	Hash-chained entries with ECDSA-P256 / Ed25519 / HMAC-SHA256 signatures	src/security/audit_log.rs
Merkle tree batch signing	Inclusion proofs for batch verification	src/security/merkle.rs
Access control	Per-tenant virtual keys with budget + rate limit + model allowlist	src/auth/virtual_keys.rs
Circuit breaker (PCA/PRA)	Provider availability with failure thresholds and auto-recovery	src/security/circuit_breaker.rs
Rate limiting	Per-tenant token bucket (NIS2 DoS protection)	src/security/rate_limit.rs
Data classification	Audit entry classification levels: NC, C1, C2, C3	src/security/audit_log.rs
Credentials at rest	AES-256-GCM encryption for all stored secrets	src/storage/encrypt.rs
Key management	File-based key with owner-only permissions (cross-platform)	src/auth/token_store.rs
Security headers	OWASP-compliant HTTP response headers	src/security/headers.rs

SOC 2 / ISO 27001 / HIPAA

Control area	Grob Feature	Status
Access control	Virtual keys, JWT auth, rate limiting	Implemented
Audit logging	Signed, hash-chained, tamper-evident audit log	Implemented
Data protection	DLP (secrets, PII, injection), encryption at rest	Implemented
Availability	Circuit breakers, multi-provider failover, zero-downtime upgrades	Implemented
Monitoring	Prometheus, OpenTelemetry, live TUI, log export	Implemented
Incident response	Canary tokens, escalation webhooks, risk classification	Implemented
Certification	SOC 2 / ISO 27001 / HIPAA formal audits	Not certified (features are present, certification requires third-party audit)

NIS2 / DORA

Requirement	Grob Feature
ICT risk management	Circuit breakers, adaptive provider scoring, spend budgets
Incident reporting	Signed audit log, escalation webhooks, canary tokens
Digital operational resilience	Multi-provider failover, zero-downtime upgrades, connection warmup
Third-party risk	Per-provider spend tracking, rate limiting, model allowlists

Compliance presets

Preset	What it enables
grob preset apply gdpr	EU-only routing + DLP (PII redaction, pseudonymization)
grob preset apply eu-ai-act	GDPR + signed audit log + transparency headers + risk classification

What grob does NOT provide

Item	Why
SOC 2 / ISO 27001 / HIPAA certification	Requires third-party audit ($30-100k). Features are present, certification is a business process.
ANSSI qualification (SecNumCloud)	Requires dedicated infrastructure and audit. Grob can be deployed in a qualified environment.
Data residency guarantees	Grob filters providers by region tag, but does not control where providers process data.
End-to-end encryption	TLS to providers is standard HTTPS. Grob does not encrypt the LLM payload itself (the provider must see it to respond).

Implementation verification (audited 2026-03-18)

Every compliance claim was verified against the actual codebase:

#	Feature	Status	Code Evidence
1	EU AI Act Art. 12 — record-keeping	Implemented	audit_log.rs:122-130 model_name + tokens in signed entries
2	EU AI Act Art. 14 — risk scoring	Implemented	risk.rs:20-30 scoring: injection=critical, blocked+PII=high
3	EU AI Act Art. 14 — escalation webhook	Implemented	risk.rs:49-87 async POST to configured URL
4	EU AI Act Art. 15 — injection detection	Implemented	prompt_injection.rs 28 languages + anti-obfuscation
5	EU AI Act Art. 52 — transparency headers	Implemented	middleware.rs:36-52 X-AI-Provider/Model/Generated
6	GDPR — region routing	Implemented	helpers.rs filters providers by region when gdpr=true
7	GDPR — PII redaction	Implemented	pii.rs credit cards (Luhn), IBANs (mod97), BICs
8	GDPR — name pseudonymization	Implemented	names.rs reversible HMAC-SHA256 mapping
9	GDPR — canary tokens	Implemented	canary.rs + dfa.rs canary injected on redaction
10	HDS/PCI — audit hash chain	Implemented	audit_log.rs:344-374 SHA-256 chaining
11	HDS/PCI — classification NC/C1/C2/C3	Implemented	audit.rs dynamic: injection=C3, PII=C2, DLP=C1, none=Nc
12	HDS/PCI — Merkle batch signing	Implemented	merkle.rs + audit_log.rs:419-453 inclusion proofs
13	HDS/PCI — signing algorithms	Implemented	audit_signer.rs ECDSA-P256, Ed25519, HMAC-SHA256
14	NIS2/DORA — escalation webhook	Implemented	Same as #3

Authentication

Method	Description	Config
None	No authentication (default)	[auth] mode = "none"
API key	Static key (Bearer or x-api-key header)	[auth] mode = "api_key"
JWT	RS256/HS256 with tenant extraction and JWKS refresh	[auth] mode = "jwt"
Virtual keys	Per-tenant keys with budget, rate limit, and model allowlist	grob key create
OAuth PKCE	Browser-based login for Anthropic Max, OpenAI, Gemini Pro	auth_type = "oauth"

Multi-Tenant Virtual Keys

Feature	Description
Key generation	grob_ prefix + 32 hex chars, SHA-256 hashed at rest
Per-key budget	Monthly USD cap per virtual key
Per-key rate limit	RPS override per key
Model allowlist	Restrict accessible models per key
Key expiration	TTL in days
Revocation	Instant via grob key revoke
Tenant isolation	Spend tracking and rate limiting per tenant_id

Observability

Feature	Description	Config
Prometheus metrics	/metrics endpoint (request count, latency, spend, cache stats)	Built-in
OpenTelemetry	OTLP trace export (gRPC)	[otel] enabled = true (feature otel)
Log export	Structured request logs to stdout, JSONL file, or HTTP webhook	[log_export] enabled = true
Live TUI	grob watch — real-time traffic inspector with DLP/fallback events	grob watch (feature watch)
SSE event stream	GET /api/events for programmatic monitoring	Built-in
Message tracing	Per-request trace IDs with structured logging	[server.tracing] enabled = true
Spend tracking	Persistent monthly spend per provider/model/tenant (JSONL journals)	Built-in
Budget alerts	Warning at configurable threshold (default 80%)	[budget] warn_at_percent

Operations

Feature	Description
Single binary	6 MB container (FROM scratch), TLS bundled via rustls
Zero-downtime upgrades	SO_REUSEPORT + SIGUSR1 graceful drain
Native TLS + ACME	Built-in HTTPS with Let's Encrypt auto-certificates
Presets	One-command configuration (perf, medium, cheap, local, gdpr, eu-ai-act)
Setup wizard	Interactive first-run: tool selection, auth, compliance, budget
Config hot-reload	POST /api/config/reload — atomic swap without restart
Connection warmup	Pre-TLS handshake on startup for all providers
Record & replay	Capture live traffic → replay through mock backend
Health endpoints	/health, /live, /ready (Kubernetes-compatible)
Diagnostics	grob doctor — config, providers, storage, service checks

CLI Commands

Command	Description
grob setup	Interactive first-run wizard
grob start [-d]	Start server (foreground or detached)
grob stop	Stop server
grob restart [-d]	Restart server
grob exec -- <cmd>	Run command behind proxy (auto start/stop)
grob watch	Live TUI traffic inspector
grob status	Service status + spend summary
grob spend	Monthly spend breakdown by provider/model
grob key create/list/revoke	Virtual API key management
grob validate	Test all providers with real API calls
grob doctor	Diagnostic health checks
grob preset list/apply/export	Preset management
grob connect [provider]	Interactive credential setup
grob env	Environment variable status
grob model	Model and routing info
grob init	Create .grob.toml project config
grob config-diff [target]	Compare config vs preset
grob upgrade	Zero-downtime binary upgrade
grob run	Container mode (0.0.0.0, JSON logs, no PID)
grob completions <shell>	Generate shell completions

Feature Flags

Flag	Default	Purpose
dlp	Yes	Data Loss Prevention
oauth	Yes	OAuth PKCE flows
tap	Yes	Webhook event emission
compliance	Yes	Signed audit logging
mcp	Yes	MCP tool matrix
watch	Yes	Live TUI dashboard
tls	No	TLS with rustls
acme	No	Let's Encrypt auto-certs
otel	No	OpenTelemetry export
harness	No	Record & replay testing

Architecture

• Language: Rust (tokio async runtime, axum HTTP framework)
• Storage: Atomic files + JSONL journals (no PostgreSQL, no Redis, no embedded DB)
• Allocator: jemalloc (non-MSVC) for ~20% throughput improvement
• Container: 6 MB FROM scratch, rustls TLS bundled
• Codebase: ~66K lines of Rust, 1193 tests
• Traits: 7 core abstractions (DlpPipeline, RequestRouter, Tracer, SpendTracking, AuditWriter, EventTap, ProviderAvailability)