generated: '2026-09-19'
method: searched
source: https://snhp.dev/llms.txt
corroboration:
  - 'openapi/snhp-dev-openapi.yml — operation descriptions of issue_key_v1_keys_post and negotiate_turn_endpoint_v1_negotiate_turn_post'
  - https://github.com/ryuxik/snhp/blob/main/PRICING.md   # "Abuse resistance"
  - 'live unauthenticated responses from https://snhp.dev on 2026-09-19'
checked: '2026-09-19'
summary: >-
  Limits are published precisely and in three places (llms.txt "Cost model", the key-issuance
  operation's own description, and the negotiate/turn description): an in-memory token bucket per
  single instance giving 60 requests/min per IP to keyless callers, 600/min per key to callers who
  send a gt_* key in a HEADER (a key in the request body "does NOT raise your limit"), 10/hour per
  IP on key issuance, and — for the opt-in LLM dispute routes — 40/hour per IP plus a persistent
  $5/day spend cap. The runtime signal is a 429 that "always carries Retry-After (whole seconds
  until a token frees up)". No RateLimit-* or X-RateLimit-* headers were observed on a live 200
  and the OpenAPI declares no 429 on any operation. The issued key's response carries
  rate_limit_per_minute: 600 as a field.
limit_count: 5
scopes:
  - scope: per-IP
    tier: keyless (the free floor)
    surface: 'all /v1/* calls'
    window: 1 minute
    limit: 60
    unit: requests
    burst: 'token bucket (burst size not published)'
    status_on_exhaustion: 429
    verbatim: 'keyless /v1/* calls:   60/min per IP   (the free floor)'
    source: https://snhp.dev/llms.txt
  - scope: per-key
    tier: 'any gt_* key sent as a header'
    surface: 'all /v1/* calls'
    window: 1 minute
    limit: 600
    unit: requests
    burst: 'token bucket (burst size not published)'
    status_on_exhaustion: 429
    verbatim: 'keyed /v1/* calls:    600/min per key  (send the key as a HEADER — `Authorization: Bearer gt_*` or `X-API-Key: gt_*`; a key in the request body does NOT raise your limit, it stays on the 60/min floor)'
    source: https://snhp.dev/llms.txt
    note: 'The key-issuance response echoes this as rate_limit_per_minute: 600 (operation description).'
  - scope: per-IP
    tier: keyless
    surface: 'POST /v1/keys (issuance)'
    window: 1 hour
    limit: 10
    unit: requests
    status_on_exhaustion: 429
    verbatim: 'POST /v1/keys:         10/hour per IP   (issuance is unauthenticated)'
    source: https://snhp.dev/llms.txt
  - scope: per-IP
    tier: 'LLM extras (off by default)'
    surface: '/v1/dispute/* drafting and coaching routes'
    window: 1 hour
    limit: 40
    unit: requests
    status_on_exhaustion: 429
    verbatim: 'a per-IP hourly limit (SNHP_LLM_PER_IP_HOURLY, default 40)'
    source: https://github.com/ryuxik/snhp/blob/main/PRICING.md
  - scope: per-deployment
    tier: 'LLM extras (off by default)'
    surface: '/v1/dispute/* drafting and coaching routes'
    window: 1 day
    limit: '$5 USD of upstream LLM spend'
    unit: dollars
    status_on_exhaustion: 429
    verbatim: 'a daily USD cap (SNHP_DAILY_LLM_USD, default $5) that persists across restarts ... Past the cap, calls 429.'
    source: https://github.com/ryuxik/snhp/blob/main/PRICING.md
headers:
  documented: true
  documented_headers: ['Retry-After  (on 429 only; whole seconds)']
  observed: false
  observed_note: >-
    Live 200 responses on 2026-09-19 (POST /v1/negotiate/turn keyless, GET /health, GET
    /v1/catalog) carried no rate-limit header of any family; the response header set is Fly.io's
    (server, via, fly-request-id) plus the provider's security headers (strict-transport-security,
    x-content-type-options, x-frame-options, referrer-policy, content-security-policy). No 429 was
    provoked, so Retry-After was not observed directly.
  absent_on_200:
    - X-RateLimit-Limit
    - X-RateLimit-Remaining
    - X-RateLimit-Reset
    - RateLimit
    - RateLimit-Policy
    - RateLimit-Limit
statuses:
  rate_limited: 429
  rate_limited_note: 'Documented, not declared: no operation in the OpenAPI lists a 429 response (overlays/ proposes one).'
implementation_note_verbatim: 'Rate limits (in-memory token bucket, per single instance)'
implementation_reading: 'Per-instance in-memory buckets on a single deployment: limits are not shared across instances and reset on restart. The provider states both facts.'
mcp_note: 'No separate limit is published for the MCP doors; the paid MCP tools take the same gt_* key as an argument, and the HTTP limiter "only reads headers", so whether an MCP call rides the 60/min or 600/min lane is not stated.'
agent_guidance: >-
  Issue a key once (POST /v1/keys, idempotent on agent_id for 24h) and send it in Authorization:
  Bearer — that alone multiplies the ceiling by 10. On 429 sleep exactly Retry-After seconds. Do
  not put the key in the body expecting a higher lane.