OpenAI-Compatible API & SDK Docs

Quickstart

Base URL: https://api.sentiogate.dev/v1
Auth header: Authorization: Bearer sg_live_... (create a gateway key in the Vault; provider keys stay in your Vault, BYOK)
Client: any OpenAI SDK, or our open-source SDKs (Apache-2.0): sentiogate-python on PyPI (pip install sentiogate-python) and @sentiogate/sdk on npm (npm i @sentiogate/sdk). The SentioGate service itself is proprietary.
model: "auto" picks the cheapest qualifying model; any provider/model id pins one.
Per-request controls: route_preference (cost, quality or latency) and task_class in the body; X-Session-ID keeps a conversation on the same route; X-SentioGate-Cache, X-SentioGate-Tool-Trim, X-SentioGate-Loop-Stop and X-SentioGate-Tool-Guard headers switch those features for one request. Account-wide switches are under Routing policy in the dashboard.

curl https://api.sentiogate.dev/v1/chat/completions \
  -H "Authorization: Bearer sg_live_your_gateway_key" \
  -H "Content-Type: application/json" \
  -d '{"model": "auto", "messages": [{"role": "user", "content": "Hello"}]}'

How `model: "auto"` works

Send model: "auto" and SentioGate ranks the models you enabled in your Vault for each request, by cost, latency and current health. It automatically falls back to another model if one fails: on a rate limit (HTTP 429) or a server error it tries your next key for the same model, then the next ranked model, which can be on another provider. A 429 puts that model on that key into a cooldown shared across requests, honoring the provider's Retry-After, so later requests skip it.

Send a provider/model id to pin one model. A pinned model only gets key failover; it never switches to a different model. OpenRouter model ids keep their own slash, for example openrouter/anthropic/claude-sonnet-4. GET /v1/models lists the ids you can route to.

Authentication and your provider keys (BYOK)

Every request carries a SentioGate gateway key (sg_live_...) as Authorization: Bearer sg_live_.... Create one in the dashboard.
The Anthropic-compatible /v1/messages endpoint also accepts the key as x-api-key, which is what Anthropic SDKs and Claude Code send.
Your provider keys (for example OpenAI, Anthropic, Gemini, Groq, Mistral, DeepSeek, xAI or OpenRouter) go in the Vault, encrypted at rest. Providers bill your own accounts directly, with 0% token markup.
You can store up to 5 keys per provider, in priority order, with optional per-key spend and balance limits that pause a key.
model: "auto" only routes to models you enabled in the Vault.

Endpoints

`POST /v1/chat/completions`	OpenAI-compatible Chat Completions. With `stream: true` the Server-Sent Events stream is passed through and router metadata arrives in the `X-SentioGate-Router` response header.
`GET /v1/models`	Lists the model ids you can route to.
`POST /v1/messages`	Anthropic Messages API, for Claude Code and Anthropic SDKs. Claude models need your own Anthropic or OpenRouter key.
`POST /v1/messages/count_tokens`	Counts input tokens for a Messages request.

Free community pool (no provider keys yet)

With a gateway key and no provider keys in your Vault, model: "auto" is answered by the free community pool: 600 requests per 5 hours per user during the beta, with no credit card. It is meant for testing your integration. Other model ids need your own provider key. Once you add keys, the same model: "auto" call routes across them with no code change.

Errors and retries

Errors use the OpenAI error shape: {"error": {"message", "type", "code"}}. Check error.code:

`401 invalid_api_key`	Missing or invalid gateway key.
`403 no_enabled_models`	`auto` found no enabled models in your Vault. Enable at least one model.
`403 community_model_only`	No provider keys yet, and the request asked for a specific model. Send `auto`, or add your own key.
`403 community_pool_disabled_with_byok`	The request asked for the community pool while provider keys are saved. Send `auto` or a `provider/model` id.
`429 upstream_rate_limited`	Your providers rate-limited every model SentioGate tried. Wait for `Retry-After`, or add another key or provider.
`429 community_pool_exhausted`	You used your community pool requests for this window. `Retry-After` is the time until it frees up.
`503 community_busy`	The shared community pool is briefly busy. Retry after the short `Retry-After`.

Wait at least Retry-After before retrying, and retry in one layer only: if you use the OpenAI SDK's built-in retries, leave SentioGate SDK retries off, or the other way round.

Which model answered: router metadata

Non-streaming responses include a router object next to the usual OpenAI fields, with the policy that ran, the reason a model was picked and any fallback attempts. Log it to see which provider and model actually answered. The full request and response schemas are in the endpoint reference below.

Endpoint reference

Inference

POST/chat/completionsCreate a chat completion (OpenAI-compatible)

With `stream: true` the upstream Server-Sent Events stream is passed through unchanged and router metadata is returned in the `X-SentioGate-Router` response header (JSON). Non-streaming responses carry the same metadata in the `router` field.

Parameters

`X-Session-ID` header	`string` Conversation or agent-session id (max 256 chars). Optional: keeps the requests of one session together, so `model: "auto"` stays consistent across tool turns and provider prompt caching works better.
`X-SentioGate-Prompt-Cache` header	`"off"` Send `off` to stop SentioGate adding a provider prompt-cache breakpoint. Client cache_control fields are always forwarded.
`X-SentioGate-Cascade` header	`"on" \| "off"` Cascade (opt-in, non-streaming, `model: "auto"`): `on` first tries a cheaper model from your vault and, if its answer does not pass SentioGate's quality checks, retries once with the normal pick. Both attempts are billed by your provider. `off` turns it off for this request. The outcome is reported in `router.cascade`.
`X-SentioGate-Length-Cap` header	`"on" \| "off"` Send `on` to enable the answer-length cap for this request (off by default), or `off` to disable it.
`X-SentioGate-Cache` header	`"off" \| "bypass" \| "force"` Send `off` (or `bypass`) to skip the response cache, or `force` to cache this request at any temperature (still never for streaming, tools or n>1). Responses carry `X-SentioGate-Cache` (`exact`, `semantic`, `miss` or `skip-<reason>`) and, on a hit, `X-SentioGate-Cache-Saved-USD`.
`X-SentioGate-Tool-Trim` header	`"on" \| "off"` Tool-output trimming (on by default for `model: "auto"` and your own keys): shortens large tool results from earlier agent steps. System, user and assistant messages and the newest tool results are never changed. Send `off` to opt out for this request, or `on` to apply it when your account setting is off. Savings are reported in `router.tool_trim`.
`X-SentioGate-Loop-Stop` header	`"on" \| "off"` Runaway-loop stop (on by default). When an agent looks stuck repeating the same steps, the request is flagged in the `X-SentioGate-Loop` response header and answered with 400 `loop_detected` instead of calling a model. Send `off` to only flag it (overrides the account setting); `on` forces the stop.
`X-SentioGate-Tool-Guard` header	`"flag" \| "off"` Dangerous tool-call flag (on by default). When the model asks to run a tool call that looks dangerous, such as deleting the root or home directory, piping a download into a shell, sending credential files or secrets out, or force-pushing to main, the response is flagged in the `X-SentioGate-Guard` response header and in `router.guard`. On streams the flag arrives as an SSE comment line starting with `: x-sentiogate-guard` before the end of the stream. Flag only: the answer is never changed or blocked. Send `off` to skip the check for this request (overrides the account setting).
`X-SentioGate-Allow-Queue` header	`"on" \| "off"` Advanced, for background jobs only. On a non-streaming request served by the free community pool, `on` lets a busy pool answer 202 Accepted with a poll link (`Location: /v1/batch/poll/{jobId}`) instead of a busy error. Requests on your own provider keys are always answered directly.

Request body `ChatCompletionRequest`

model

string

`auto`, `provider/model` (e.g. `openai/gpt-4o-mini`) or a bare model id.

messagesrequired

Message[]

fields

`role`required	`"system" \| "user" \| "assistant" \| "tool"`
`content`required	`string \| object[]`
`name`	`string`

stream boolean

temperature

number

Sent upstream only when set; unset uses the provider default.

max_tokens integer

top_p number

frequency_penalty number

presence_penalty number

stop string | string[]

task_class

Optional hint for `model: "auto"` that overrides the automatic task detection.

route_preference

"cost" | "quality" | "latency"

`cost` (default): lowest price among suitable models. `quality`: prefer more capable models. `latency`: prefer faster models.

session_id

string

Same as header X-Session-ID.

cache_control

object

Forwarded untouched to the provider. Block-level cache_control inside messages is also forwarded untouched.

prompt_cache_key

string

Forwarded (OpenAI-style cache routing key).

prompt_cache

boolean

false stops SentioGate auto-inserting a prompt-cache breakpoint.

cascade

boolean

Same as header X-SentioGate-Cascade.

length_cap

boolean

true enables the answer-length cap for this request (off by default; same as header `X-SentioGate-Length-Cap: on`). It applies only when you set no max_tokens / max_completion_tokens and the request is not streaming; an answer cut by the cap is retried once without it.

cache

boolean

false skips the response cache (same as header `X-SentioGate-Cache: off`); true opts in at any temperature (same as `force`). By default only deterministic requests are cached (a low temperature, or a single-turn factual question with temperature unset).

tool_trim

boolean

false turns tool-output trimming off for this request (same as header `X-SentioGate-Tool-Trim: off`).

compress

"auto" | "lossless" | "on" | "off"

Context compression. `auto` (default): lossless clean-up only, plus summarizing older turns when a long session nears the model's context limit. System prompts, code blocks and tool-call pairs are never changed, and provider prompt caching keeps working. `lossless`: never summarize. `on`: summarize older turns more readily. `off`: send messages unchanged. Header `X-SentioGate-Compress` is equivalent. Savings are reported in `router.compression`.

metadata

object

fields

`task_class`	`"classification" \| "simple_qa" \| "summarization" \| "general" \| "long_context" \| "coding" \| "reasoning"` Optional hint for `model: "auto"` that overrides the automatic task detection.
`route_preference`	`"cost" \| "quality" \| "latency"` `cost` (default): lowest price among suitable models. `quality`: prefer more capable models. `latency`: prefer faster models.

tools

object[]

OpenAI-style tools; forwarded upstream.

tool_choice

object

Forwarded upstream with tools.

parallel_tool_calls

boolean

Forwarded upstream when set.

functions

object[]

Legacy OpenAI functions; forwarded when tools is absent.

provider_keys

object

Per-request provider key overrides (not stored).

Responses

200

Completion. `text/event-stream` when `stream: true`.

ChatCompletionResponse

id string

object string

created integer

model string

choices

object[]

fields

index integer

message

Message

fields

`role`required	`"system" \| "user" \| "assistant" \| "tool"`
`content`required	`string \| object[]`
`name`	`string`

finish_reason stringnull

usage

object

fields

`prompt_tokens`	`integer`
`completion_tokens`	`integer`
`total_tokens`	`integer`

cached boolean

router

RouterInfo

SentioGate routing metadata. Informational; other fields may appear.

fields

latency_ms

integer

End-to-end time at the edge (non-streaming).

provider string

model string

policy "sentiogate_auto" | "explicit" | "cache"

task_class

Optional hint for `model: "auto"` that overrides the automatic task detection.

min_tier integer

model_tier integernull

estimated_prompt_tokens integer

estimated_output_tokens integer

estimated_cost_usd numbernull

pricing_source string

reason string

candidates integer

attempts

string[]

`provider/model:status` per upstream attempt.

route_preference

"cost" | "quality" | "latency"

`cost` (default): lowest price among suitable models. `quality`: prefer more capable models. `latency`: prefer faster models.

measured_latency_ms

numbernull

Recently measured latency of the chosen model, if known.

prompt_cache

object

Provider prompt-cache usage for this response and the estimated saving.

fields

`cached_tokens`	`integer`
`cache_write_tokens`	`integer`
`saved_usd`	`numbernull`
`saved_source`	`stringnull`
`auto`	`objectnull`

agent

object

Present on `model: "auto"` requests that carry tools.

fields

`tools`	`boolean`
`tool_rounds`	`integer`

cost_source

string

How estimated_cost_usd was computed (for example from the provider's published prices, or `unknown_price`).

reasoning_retry

object

Present when a reasoning model returned an empty answer and the request was retried automatically. Both attempts are billed.

cascade

object

Present when cascade was requested.

fields

`attempted`	`boolean`
`escalated`	`boolean`
`reason`	`string` `passed`, or why the cheaper answer was not used.
`first_model`	`string`
`first_cost_usd`	`number`
`first_latency_ms`	`integer`

length_cap

object

Answer-length cap: whether it applied, the token limit, and whether a truncated answer was retried uncapped.

fields

`cap_applied`	`boolean`
`cap_tokens`	`integer`
`cap_retry`	`boolean`
`cap_skipped`	`"not_enabled" \| "client_limit" \| "streaming"`

cache

object

fields

`hit`	`boolean`
`layer`	`"exact" \| "semantic"`
`cost_usd`	`number`
`saved_usd`	`number`
`skipped`	`string`

tool_trim

object

Present when tool-output trimming changed this request.

fields

`trimmed`	`integer`
`chars_saved`	`integer`
`tokens_saved_est`	`integer`

loop

object

Present when the agent looks stuck in a loop.

fields

`suspected`	`boolean`
`count`	`integer`

guard

object

Present when the answer contains tool calls and the dangerous tool-call flag is on. Flag only: the answer is not changed.

fields

`mode`	`"flag"`
`inspected`	`integer` Tool calls checked.
`flagged`	`integer` Tool calls that look dangerous.
`rules`	`string[]` Ids of the matched rules, for example `rm-root-or-home`, `pipe-to-shell`, `secret-file-exfil`, `force-push-main`, `secret-in-outbound-args`.

compression

object

fields

`applied`	`boolean`
`original_tokens`	`integer`
`compressed_tokens`	`integer`
`saved_tokens`	`integer`
`saved_pct`	`number`
`mode`	`string`
`lossy`	`boolean`
`summary_cost_usd`	`number`

400

Invalid request body. Also when the agent looks stuck repeating the same steps and runaway-loop stop is on, the default (`loop_detected`). Change the approach or input and send again; retrying the same request gives the same answer.

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

401

Missing or invalid SentioGate API key (`invalid_api_key`).

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

402

No provider key in your vault for the requested provider, or no keys at all (`no_provider_key`).

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

403

Model disabled by vault policy (`model_disabled_by_policy`), or auto found no capable/enabled model (`no_capable_model`, `no_enabled_models`).

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

429

Vault spend/request cap reached (`policy_limit_reached`).

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

501

Selected provider is not OpenAI-compatible and cannot be forwarded by the edge (`provider_not_supported_on_edge`).

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

502

Upstream provider returned an error or was unreachable.

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

503

Temporarily unavailable, for example the free community pool is busy (`community_busy`, with `Retry-After`).

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

Anthropic

POST/messagesCreate a message (Anthropic Messages API)

Anthropic-compatible endpoint for Claude Code and Anthropic SDKs: set `ANTHROPIC_BASE_URL=https://api.sentiogate.dev` and `ANTHROPIC_API_KEY` to your `sg_live_...` key (sent as `x-api-key` or `Authorization: Bearer`). `?beta=true` is accepted. With your own Anthropic key in the Vault, `claude-*` models are sent natively to Anthropic, so the request (thinking, tools, cache_control, documents) and the response or SSE stream are passed through as Anthropic sent them. Any other model (`auto`, `provider/model`, a bare id) is translated to an OpenAI-compatible chat completion on your vault keys and the answer is translated back into Anthropic's message shape and SSE event order (`message_start`, `content_block_*`, `message_delta`, `message_stop`). Without vault keys, only `auto` (or `sentiogate-community`) is served, on the free community models; other ids return 403 `permission_error`. For Claude Code set `ANTHROPIC_MODEL=auto` and `ANTHROPIC_SMALL_FAST_MODEL=auto` (and `ANTHROPIC_DEFAULT_HAIKU_MODEL=auto`). Claude models need your own Anthropic or OpenRouter key. A 429 carries the upstream `Retry-After` when the provider sent one. Router metadata is returned in the `X-SentioGate-Router` header.

Parameters

`anthropic-version` header	`string` Anthropic API version (default `2023-06-01`). Forwarded on native Anthropic calls.
`anthropic-beta` header	`string` Anthropic beta flags. Forwarded on native Anthropic calls.
`X-SentioGate-Tool-Trim` header	`"on" \| "off"` Tool-output trimming (on by default for `model: "auto"` and your own keys): shortens large tool results from earlier agent steps. System, user and assistant messages and the newest tool results are never changed. Send `off` to opt out for this request, or `on` to apply it when your account setting is off. Savings are reported in `router.tool_trim`.
`X-SentioGate-Loop-Stop` header	`"on" \| "off"` Runaway-loop stop (on by default). When an agent looks stuck repeating the same steps, the request is flagged in the `X-SentioGate-Loop` response header and answered with 400 `loop_detected` instead of calling a model. Send `off` to only flag it (overrides the account setting); `on` forces the stop.
`X-SentioGate-Tool-Guard` header	`"flag" \| "off"` Dangerous tool-call flag (on by default). When the model asks to run a tool call that looks dangerous, such as deleting the root or home directory, piping a download into a shell, sending credential files or secrets out, or force-pushing to main, the response is flagged in the `X-SentioGate-Guard` response header and in `router.guard`. On streams the flag arrives as an SSE comment line starting with `: x-sentiogate-guard` before the end of the stream. Flag only: the answer is never changed or blocked. Send `off` to skip the check for this request (overrides the account setting).

Request body `MessagesRequest`

modelrequired

string

`claude-*` (native with your Anthropic key; via OpenRouter as `anthropic/<id>` otherwise), `auto`, `sentiogate-community`, or `provider/model`.

max_tokensrequired integer

messagesrequired

object[]

fields

`role`required	`"user" \| "assistant"`
`content`required	`string \| object[]`

system string | object[]

stream boolean

temperature number

top_p number

top_k integer

stop_sequences string[]

tools

object[]

Custom tools are supported on every route; Anthropic server tools only on the native route.

tool_choice object

thinking object

metadata object

Responses

200

Message. `text/event-stream` when `stream: true`.

MessagesResponse

id string

type string

role string

model string

content object[]

stop_sequence stringnull

usage

object

fields

`input_tokens`	`integer`
`output_tokens`	`integer`
`cache_creation_input_tokens`	`integer`
`cache_read_input_tokens`	`integer`

400

Invalid request (`invalid_request_error`). Also when the agent looks stuck repeating the same steps and runaway-loop stop is on, the default (`invalid_request_error`). Change the approach or input and send again; retrying the same request gives the same answer.

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

401

Missing or invalid SentioGate API key (`authentication_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

402

No provider key in your vault for the requested model (`billing_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

403

Model not available without your own provider key, or disabled by vault policy (`permission_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

405

Only POST (and HEAD) are supported.

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

413

Request too large (`request_too_large`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

429

Rate limited (`rate_limit_error`); `Retry-After` is set when known.

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

501

The requested provider cannot be served by the edge (`api_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

502

Upstream provider error (`api_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

504

Upstream provider timed out (`timeout_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

529

Upstream provider overloaded (`overloaded_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

POST/messages/count_tokensCount input tokens (Anthropic Messages API)

For `claude-*` models with your own Anthropic key in the Vault, Anthropic counts the tokens. Otherwise SentioGate returns an estimate. The `X-SentioGate-Token-Count` response header says which (`anthropic` or `estimate`).

Parameters

`anthropic-version` header	`string` Anthropic API version (default `2023-06-01`). Forwarded on native Anthropic calls.
`anthropic-beta` header	`string` Anthropic beta flags. Forwarded on native Anthropic calls.

Request body `MessagesRequest`

modelrequired

string

`claude-*` (native with your Anthropic key; via OpenRouter as `anthropic/<id>` otherwise), `auto`, `sentiogate-community`, or `provider/model`.

max_tokensrequired integer

messagesrequired

object[]

fields

`role`required	`"user" \| "assistant"`
`content`required	`string \| object[]`

system string | object[]

stream boolean

temperature number

top_p number

top_k integer

stop_sequences string[]

tools

object[]

Custom tools are supported on every route; Anthropic server tools only on the native route.

tool_choice object

thinking object

metadata object

Responses

200

Token count

object

input_tokensrequired integer

400

Invalid request (`invalid_request_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

401

Missing or invalid SentioGate API key (`authentication_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

429

Rate limited (`rate_limit_error`).

AnthropicError

typerequired string

errorrequired

object

fields

`type`required	`"invalid_request_error" \| "authentication_error" \| "billing_error" \| "permission_error" \| "not_found_error" \| "request_too_large" \| "rate_limit_error" \| "timeout_error" \| "api_error" \| "overloaded_error"`
`message`required	`string`

Catalog

GET/modelsList routable model idspublic

Responses

200

Model list

ModelList

object string

data

object[]

fields

`id`	`string`
`object`	`string`
`owned_by`	`string`

GET/providersList supported providerspublic

Responses

200

Providers

object

object

GET/pricingModel list prices used for routing and cost estimatespublic

Responses

200

Pricing snapshot

object

object

GET/pricing/modelsPer-model list prices (USD per token), context length, tool support and capability tierpublic

Responses

200

Model price catalog

object

`source`	`string`
`fetched_at`	`stringnull`
`count`	`integer`
`models`	`object`

POST/waitlistGet an email when a coming-soon paid plan launches ("Notify me" on /pricing)public

No account needed; when signed in, the account is linked to the sign-up. Stores only the email, the plan and the time. Signing up twice for the same plan is fine and returns the same answer. Nothing is emailed until the plan launches. Rate limited per IP.

Request body

`email`required	`string`
`plan`required	`"pro" \| "team"`

Responses

200

On the list

object

`status`	`"subscribed"`
`plan`	`"pro" \| "team"`
`message`	`string`

400

Invalid email or plan

Detail

detail string

429

Too many sign-ups from this address; see Retry-After

Error

errorrequired

object

fields

`message`required	`string`
`type`	`string`
`code`	`"invalid_api_key" \| "no_provider_key" \| "model_disabled_by_policy" \| "no_capable_model" \| "no_enabled_models" \| "policy_limit_reached" \| "provider_not_supported_on_edge"`
`provider`	`stringnull`
`model`	`stringnull`

503

Temporarily unavailable

Detail

detail string

GET/benchmarksPer-model latency and success rates measured from recent routed trafficpublic

Responses

200

Measured model stats

BenchmarkList

window string

measured boolean

benchmarks

object[]

fields

`provider`	`string`
`model`	`string`
`samples`	`integer`
`p50_latency_ms`	`numbernull`
`avg_ttft_ms`	`numbernull`
`error_rate_pct`	`numbernull`

Auth

GET/auth/meCurrent user

Responses

200

User and workspaces

object

object

401

Not authenticated

Detail

detail string

GET/auth/keysList SentioGate API keys (prefixes only)

Responses

200

Keys

object

object

401

Not authenticated

Detail

detail string

POST/auth/keysCreate a SentioGate API key (`sg_live_...`, shown once)

Request body

name string

Responses

201

Created key; `raw_key` is only returned here

object

object

401

Not authenticated

Detail

detail string

DELETE/auth/keys/{key_id}Revoke a SentioGate API key

Parameters

key_id pathrequired string

Responses

200

Revoked

object

object

401

Not authenticated

Detail

detail string

Vault

GET/vault/keysList stored provider keys (masked) and model policy

Responses

200

Vault

object

object

401

Not authenticated

Detail

detail string

POST/vault/keysStore or update a provider key (encrypted at rest) and model policy

Request body

`provider`required	`string`
`api_key`	`string`
`model_overrides`	`object`
`key_policy_defaults`	`object`

Responses

200

Stored

object

object

400

Invalid key or provider

Detail

detail string

401

Not authenticated

Detail

detail string

DELETE/vault/keys/{provider}Delete a stored provider key

Parameters

provider pathrequired string

Responses

200

Deleted

object

object

401

Not authenticated

Detail

detail string

Cache

GET/cache/statsYour response-cache entries, hits and recorded savings

Responses

200

Cache stats

object

object

401

Not authenticated

Detail

detail string

DELETE/cacheDelete your cached responses (other users are never affected)

Responses

200

Cleared

object

object

401

Not authenticated

Detail

detail string

Integrations

GET/integrations/langfuseYour Langfuse export settings (the secret key is never returned)

Responses

200

Settings

LangfuseSettings

`configured`	`boolean`
`enabled`	`boolean`
`include_content`	`boolean`
`host`	`string` Langfuse host, for example `https://cloud.langfuse.com` or `https://us.cloud.langfuse.com`.
`public_key`	`stringnull` Masked public key.
`has_secret_key`	`boolean`
`updated_at`	`integernull` Unix time in milliseconds.

401

Not authenticated

Detail

detail string

503

Temporarily unavailable

Detail

detail string

PUT/integrations/langfuseSave Langfuse export settings

Opt-in export of a trace per chat request to your own Langfuse project. Off by default. Fields you leave out keep their saved value; `secret_key` is write-only and stored encrypted. Enabling needs both keys. Prompts and answers are only exported when `include_content` is true; otherwise traces carry metadata only (model, provider, token usage, cost, latency, status).

Request body `LangfuseSettingsUpdate`

`enabled`	`boolean`
`include_content`	`boolean` Also export prompts and answers. Default false (metadata only).
`host`	`string` HTTPS Langfuse host (self-hosted allowed; private addresses are rejected).
`public_key`	`string` Starts with `pk-lf-`.
`secret_key`	`string` Starts with `sk-lf-`. Write-only: stored encrypted and never returned.

Responses

200

Saved settings

LangfuseSettings

`configured`	`boolean`
`enabled`	`boolean`
`include_content`	`boolean`
`host`	`string` Langfuse host, for example `https://cloud.langfuse.com` or `https://us.cloud.langfuse.com`.
`public_key`	`stringnull` Masked public key.
`has_secret_key`	`boolean`
`updated_at`	`integernull` Unix time in milliseconds.

400

Invalid JSON, host or key, or enabling without both keys

Detail

detail string

401

Not authenticated

Detail

detail string

503

Temporarily unavailable

Detail

detail string

DELETE/integrations/langfuseTurn off the Langfuse export and delete the saved keys

Responses

200

Settings after deletion (`configured: false`)

LangfuseSettings

`configured`	`boolean`
`enabled`	`boolean`
`include_content`	`boolean`
`host`	`string` Langfuse host, for example `https://cloud.langfuse.com` or `https://us.cloud.langfuse.com`.
`public_key`	`stringnull` Masked public key.
`has_secret_key`	`boolean`
`updated_at`	`integernull` Unix time in milliseconds.

401

Not authenticated

Detail

detail string

503

Temporarily unavailable

Detail

detail string

POST/integrations/langfuse/testCheck the saved Langfuse keys against your Langfuse host

Responses

200

Result. `ok: false` with a `detail` when the keys are missing, rejected or the host is unreachable; `status` is the HTTP status Langfuse answered (0 when unreachable).

object

`ok`required	`boolean`
`status`	`integer`
`detail`	`string`

401

Not authenticated

Detail

detail string

429

Tested too often; `Retry-After` says when to try again.

object

`ok`	`boolean`
`detail`	`string`

503

Temporarily unavailable

Detail

detail string

Discover

POST/workload/analyzeEstimate savings from an uploaded log export (JSON, JSONL or CSV, max 5 MB). Nothing is stored.public

Request body

`raw`	`string` File contents
`records`	`object[]`

Responses

200

Estimated report

object

object

400

Empty or unparseable input

Detail

detail string

413

Upload too large

Detail

detail string

Batch

POST/batch/enqueueSubmit an asynchronous task to the free community pool

Request body

item_text string

messages

Message[]

fields

`role`required	`"system" \| "user" \| "assistant" \| "tool"`
`content`required	`string \| object[]`
`name`	`string`

model string

Responses

202

Queued task with poll_url

object

`job_id`	`string`
`status`	`string`
`poll_url`	`string`

400

Invalid payload

Detail

detail string

401

Sign-in or API key required

Detail

detail string

403

The free community pool is not available for this account (for example when you have provider keys in the Vault)

Detail

detail string

413

Item too large

Detail

detail string

429

Personal allowance (600 requests per 5 hours) used up, too many pending items, or rate limited

Detail

detail string

GET/batch/poll/{jobId}Check status and retrieve results of an enqueued batch task

Parameters

jobId pathrequired string

Responses

200

Job status and completed model output (owner only)

object

`job_id`	`string`
`status`	`string`
`response`	`object`

401

Sign-in or API key required

Detail

detail string

404

Job not found (or not yours)

Detail

detail string

Quickstart

How model: "auto" works

Authentication and your provider keys (BYOK)

Endpoints

Free community pool (no provider keys yet)

Errors and retries

Which model answered: router metadata

Endpoint reference

Inference

Parameters

Request body ChatCompletionRequest

Responses

Anthropic

Parameters

Request body MessagesRequest

Responses

Parameters

Request body MessagesRequest

Responses

Catalog

Responses

Responses

Responses

Responses

Request body

Responses

Responses

Auth

Request body

Responses

Responses

Responses

Request body

Responses

Parameters

Responses

Vault

Responses

Request body

Responses

Parameters

Responses

Cache

Responses

Responses

Integrations

Responses

Request body LangfuseSettingsUpdate

Responses

Responses

Responses

Discover

Request body

Responses

Batch

Request body

Responses

Parameters

Responses

How `model: "auto"` works

Request body `ChatCompletionRequest`

Request body `MessagesRequest`

Request body `MessagesRequest`

Request body `LangfuseSettingsUpdate`