Features

Every request routed through sphyr_net must carry a valid HMAC-SHA256 signature computed over the canonical request string. The signature is verified at the Cloudflare edge before any downstream processing begins — a tampered or replayed request is rejected immediately. Timestamps are enforced within a ±300 second window to prevent replay attacks without requiring clock synchronization infrastructure.

# HMAC-SHA256 over the canonical 8-part payload (colon-delimited):
# nonce:ts:url:mthd:sha256(body):sha256(canonical_headers):cat:trace_id
#
# Headers: sorted alphabetically, each prefixed with its UTF-8 byte length
# cat (1–64 chars) is required; trace_id defaults to ""

sphyr_net({
  url:        "https://api.example.com/data",
  mthd:       "GET",
  sig:        "<HMAC-SHA256 over canonical payload>",
  ts:         1743800400,
  sess_token: "<sess_token from audit_sess>",
  cat:        "api-call",
})
# nonce is returned by audit_sess and embedded in `sig` — never passed
# as a separate sphyr_net argument.

Exfiltration attacks often embed secrets — API keys, credentials, private keys — inside seemingly normal outbound requests. Sphyr computes Shannon entropy on request payloads and rejects payloads that exceed the configured threshold. Known high-entropy but benign patterns (UUIDs, base64-encoded binaries) are excluded via gateway-managed allow-patterns. This blocks high-entropy payloads before they reach their intended destination — catching exfiltration attempts in transit through the Sphyr gateway.

# Entropy configuration (per-key)
{
  "entropy_threshold": 4.5   # bits/char — triggers block above this
                              # Allowlist patterns (UUIDs, base64 binaries, etc.)
                              # are managed by the gateway, not per-key config.
}

Rate limiting is applied at three levels: per-agent (isolating a single misbehaving agent within a fleet), per-key (capping combined throughput across all agents sharing one credential), and per-IP (a DoS backstop protecting the gateway from unauthenticated flood attacks). Each level uses a fixed-window counter backed by Cloudflare Durable Objects for serialized, race-free consistency at the edge. When a limit is exceeded, the response includes precise retry-after timing so well-behaved agents can back off gracefully without polling. Default limits: 10 req/s per agent · 100 req/s per key · 300 req/s per IP · 500 req/s pre-auth backstop.

# Rate limit response headers
HTTP/1.1 429 Too Many Requests
X-Credits-Remaining: 4823
X-Quota-Remaining: 0
X-Request-ID: req_a8b2c3d4

# sphyr_net rate limit response body
{
  "code": "RATE_LIMITED",
  "message": "Per-key rate limit exceeded",
  "retryable": true,
  "request_id": "req_a8b2c3d4",
  "docs_url": "https://sphyr.io/docs/errors/RATE_LIMITED",
  "remediation": {
    "action": "BACKOFF",
    "parameters": { "retry_after_ms": 8400 }
  }
}

Sessions are established via the audit_sess handshake: the agent presents its Sphyr credential, Sphyr verifies the HMAC-SHA256 signature against the stored signing secret, and issues a 30-minute session token (UUID v4). The token is IP-bound — requests from a different IP are rejected, preventing session token theft. Note: agents running in serverless environments (AWS Lambda, GitHub Actions, Vercel Functions) that share a NAT gateway may share an egress IP with other agents — IP binding provides no isolation between co-tenants on the same NAT. All requests within a session are grouped in the forensics view, giving operators a coherent view of what each agent task did across its lifetime.

# audit_sess handshake
POST /mcp  (tool: audit_sess)
{ "key_id": "key_2Y5k...", "agent_id": "...", "ts": "...", "client_nonce": "...", "sig": "..." }

# Response
{
  "sess_token": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
  "nonce": "5f4dcc3b-5aa7-4c1d-9f6f-8b1e2c3d4e5f",  # included in every sphyr_net HMAC
  "expires_in_sec": 1800,                            # session TTL — 30 minutes
  "key_id": "key_2Y5k...",
  "issued_at": "2026-04-30T17:23:42.000Z",
  "expires_at": "2026-04-30T17:53:42.000Z",
  "balance": 48750
}

Every session has a security story. Trace surfaces blocked request counts and per-session peak severity at a glance — and for every blocked request, shows the exact inline rationale that triggered the block. Not a tooltip. Not a log you have to grep. Inline, in the console, on every blocked call.

# session · a3f8b2c1… · 12 requests · 3 blocked · risk 71
POST https://api.example.com/upload  BLOCKED  risk 71
  ↳ Entropy threshold exceeded — Shannon entropy 6.42 bits/char on URL parameter value; pattern matches API key format.