Anthropic Messages API¶
SMG serves the Anthropic Messages API at /v1/messages and counts input tokens at /v1/messages/count_tokens. What happens to a request depends on the backend: with gRPC workers SMG implements the API itself, with HTTP workers it forwards the request to an engine that implements the Messages API, and with the Anthropic backend it forwards the request to the Anthropic API.
Endpoints¶
| Method | Path | Description |
|---|---|---|
POST |
/v1/messages |
Create a message. Set "stream": true to receive Server-Sent Events. |
POST |
/v1/messages/count_tokens |
Count the input tokens of a Messages request without generating. |
Both endpoints go through the same authentication and admission control as the other inference endpoints.
Supported Backends¶
| Backend | Configure with | /v1/messages |
/v1/messages/count_tokens |
|---|---|---|---|
| gRPC workers | grpc:// worker URLs |
SMG renders, tokenizes, and parses (details) | Counted by SMG itself (details) |
| ZMQ workers | ipc:// or tcp:// worker URLs |
Same as gRPC workers | Same as gRPC workers |
| HTTP workers | http:// or https:// worker URLs |
Forwarded to the worker | Forwarded to one worker |
| HTTP PD | --pd-disaggregation with HTTP workers |
Prefill/decode dual dispatch | Forwarded to one prefill worker |
| Anthropic API | --backend anthropic |
Forwarded to the Anthropic API, with an optional SMG-run MCP tool loop | 501 Not Implemented |
HTTP workers must implement the Messages API themselves; SMG forwards the request without translating it.
In IGW mode (--enable-igw), SMG picks the backend per request from the workers that serve the requested model. A model served by an external Anthropic worker goes to the Anthropic API backend (see External Providers). The other external providers answer /v1/messages with 501.
Authentication¶
SMG's own authentication is separate from the credentials it passes to backends:
| Hop | Headers | Behavior |
|---|---|---|
| Client to SMG | Authorization: Bearer <key> |
Checked when --api-key or --tenant-api-key is set. x-api-key does not authenticate to SMG. A missing or unknown bearer token gets 401 Unauthorized. |
| SMG to HTTP workers | Authorization |
The client's header is forwarded with the other allowlisted headers; x-api-key is not. A matched gateway key is removed at authentication, and the worker's own configured key, when it has one, is sent instead. See Authentication for worker API keys. |
| SMG to the Anthropic API | x-api-key, Authorization |
The client's own key headers are forwarded as sent. When the client sends neither, SMG sends the worker's stored key as x-api-key, with anthropic-version: 2023-06-01 added when the client sent no version header. |
Anthropic SDKs send their key in x-api-key, which the Anthropic API backend passes upstream. When gateway authentication is on, clients must also send the gateway key as Authorization: Bearer <key>. The gateway key never reaches Anthropic: it is removed from Authorization at authentication, and from x-api-key too when sent there, while the client's own Anthropic key in x-api-key passes through.
Create a Message¶
POST /v1/messagesRequest Body¶
| Field | Type | Required | Description |
|---|---|---|---|
model |
string | Yes | Model ID. Must not be empty. |
messages |
array | Yes | At least one message, each with a role (user, assistant, or system) and content (a string or an array of content blocks). |
max_tokens |
integer | Yes | Maximum number of tokens to generate. 0 is accepted; the response depends on the backend — see max_tokens of 0. |
system |
string or array | No | System prompt, as a string or an array of text blocks. |
stream |
boolean | No | Stream the response as Server-Sent Events. |
temperature |
number | No | Sampling temperature, 0 to 1. |
top_p |
number | No | Nucleus sampling, greater than 0 and at most 1. |
top_k |
integer | No | Sample only from the top K tokens, 1 to 2147483647. Unlike Chat Completions and Responses, where top_k is signed and -1 disables it, this field is unsigned and has no disable value: a negative value is refused with 400. |
stop_sequences |
array of strings | No | Sequences that stop generation. |
tools |
array | No | Tool definitions; see Tool Use. |
tool_choice |
object | No | {"type": "auto"}, {"type": "any"}, {"type": "tool", "name": "..."}, or {"type": "none"}. Every value except none requires tools, and tool must name one of them. auto, any, and tool also take disable_parallel_tool_use: when true, the response carries at most one tool call — see Tool Use. |
thinking |
object | No | {"type": "enabled", "budget_tokens": N}, {"type": "adaptive"}, {"type": "disabled"}, or {"type": "between_tools"} (no thinking up front; the model may think between its tool calls). enabled needs budget_tokens of at least 1024, and the budget must stay under max_tokens unless the interleaved-thinking beta is on — see Request Headers. enabled and adaptive also take display (summarized, omitted, or updates). between_tools takes type alone: a request that adds any other field, display or budget_tokens included, is refused with 400, and the error message names each extra field. |
output_config |
object | No | Output configuration. format asks for structured output, a JSON answer held to a schema; see Structured Output. effort, a string or a number taken as its text, is the reasoning effort the prompt is rendered with on gRPC and ZMQ workers. Other keys are kept and forwarded like unmodeled fields. |
output_format |
object | No | The structured output format under its earlier top-level name, with the same shape as output_config.format; see Structured Output. When both are sent, output_config.format wins. |
metadata |
object | No | Request metadata (user_id). |
service_tier |
string | No | auto or standard_only. |
container |
object | No | Container for code execution (beta). |
mcp_servers |
array | No | MCP servers for the request (beta). Required when tools has an mcp_toolset entry; see MCP Tools. |
rid |
string | No | Extension: request ID passed to the backend for log correlation. |
SMG validates the body before routing and answers 400 when a required field is missing or a rule above is broken. The one exception is a body that the HTTP router streams through unparsed; see HTTP and PD Workers. Fields SMG does not model, such as context_management, are kept and forwarded to HTTP workers and the Anthropic API.
Content blocks in messages may be text, image, document, tool_use, tool_result, thinking, redacted_thinking, server_tool_use, search_result, web_search_tool_result, tool_search_tool_result, or tool_reference. Any other block type is rejected with 400. A system-role message inside messages is accepted, because some clients send one mid-conversation.
Example Request¶
curl http://localhost:30000/v1/messages \
-H "Content-Type: application/json" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "meta-llama/Llama-3.1-8B-Instruct",
"max_tokens": 1024,
"system": "You are a concise assistant.",
"messages": [
{"role": "user", "content": "What is the meaning of life?"}
]
}'curl http://localhost:30000/v1/messages \
-H "Content-Type: application/json" \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "claude-sonnet-4-20250514",
"max_tokens": 1024,
"messages": [
{"role": "user", "content": "What is the meaning of life?"}
]
}'Response¶
{
"id": "msg_abc123",
"type": "message",
"role": "assistant",
"content": [
{"type": "text", "text": "There is no single answer..."}
],
"model": "meta-llama/Llama-3.1-8B-Instruct",
"stop_reason": "end_turn",
"stop_sequence": null,
"stop_details": null,
"usage": {
"input_tokens": 27,
"output_tokens": 112,
"cache_creation_input_tokens": 0,
"cache_read_input_tokens": 0,
"cache_creation": {"ephemeral_5m_input_tokens": 0, "ephemeral_1h_input_tokens": 0},
"service_tier": "standard",
"inference_geo": "not_available"
},
"container": null
}With gRPC workers SMG builds this response (details): input_tokens is the prompt outside the engine's prefix cache and cache_read_input_tokens the part the cache served, their sum the whole prompt. With HTTP workers and the Anthropic API the response comes from the backend.
max_tokens of 0¶
A max_tokens of 0 passes validation. On gRPC and ZMQ workers, SMG answers the request itself without generating: an empty message with stop_reason of max_tokens, whose usage reports the prompt's token count, counted exactly as /v1/messages/count_tokens counts it. No worker takes part, and no tenant rate-limit budget is reserved, though the request still counts in the gateway's request metrics.
{
"id": "msg_abc123",
"type": "message",
"role": "assistant",
"content": [],
"model": "meta-llama/Llama-3.1-8B-Instruct",
"stop_reason": "max_tokens",
"stop_sequence": null,
"stop_details": null,
"usage": {
"input_tokens": 27,
"output_tokens": 0,
"cache_creation_input_tokens": 0,
"cache_read_input_tokens": 0,
"cache_creation": {"ephemeral_5m_input_tokens": 0, "ephemeral_1h_input_tokens": 0},
"service_tier": "standard",
"inference_geo": "not_available"
},
"container": null
}With "stream": true, the answer is a three-event stream: message_start with the empty message — its usage already carries the prompt's token count — then message_delta with the stop reason and the usage counts, and message_stop.
A request whose prompt cannot be counted is refused as /v1/messages/count_tokens refuses it — an unknown model gets 404, a request that does not render gets 400, and the other refusals under Backend Support apply unchanged. When the model's workers advertise a context window, a prompt with more tokens than the smallest advertised window is refused with 400 context_length_exceeded, as a generating request would be.
On HTTP and PD workers, SMG refuses max_tokens: 0 with 400 (code max_tokens_zero_not_relayed) before any worker sees the request: the relay forwards a Messages request as sent and renders no prompt, so it cannot build the empty message. As with the other request rules, a body the HTTP router streams through unparsed reaches the worker unvalidated.
The Anthropic API backend forwards the request like any other and relays the upstream response.
In IGW mode, a max_tokens: 0 request for a model served over both gRPC and HTTP always goes to a gRPC router, so it gets the empty message; a model only HTTP workers serve gets the relay's 400.
Tool Use¶
Define each tool with a JSON Schema input_schema:
{
"tools": [
{
"name": "get_weather",
"description": "Get the current weather for a city",
"input_schema": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"]
}
}
],
"tool_choice": {"type": "auto"}
}A tool call comes back as a tool_use block, with stop_reason set to tool_use. The block's caller reports who issued the call — {"type": "direct"} is the model itself. Every tool_use block SMG builds or re-serializes carries the key; a caller the Anthropic API reported, such as a server tool, is kept as it came.
{
"content": [
{
"type": "tool_use",
"id": "toolu_0b7e5f1c9a2d4e6f8a1b3c5d",
"name": "get_weather",
"input": {"city": "Paris"},
"caller": {"type": "direct"}
}
],
"stop_reason": "tool_use"
}Return the result in the next user message as a tool_result block whose tool_use_id is the id of the call.
To get at most one tool call per turn, set disable_parallel_tool_use: true on a tool_choice of auto, any, or tool. On gRPC and ZMQ workers, SMG enforces it: when the model writes several calls, only the first comes back as a tool_use block — in a stream, only that block is emitted — and the later calls are dropped, returned neither as calls nor as text. HTTP workers and the Anthropic API receive the field with the rest of tool_choice; enforcing it is left to them.
Structured Output¶
Ask for a JSON answer with output_config.format, which names a JSON schema the model's text answer must fit:
{
"output_config": {
"format": {
"type": "json_schema",
"schema": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"]
}
}
}
}json_schema is the only type: a format with another type or without schema is refused with 400, as is a schema that is not a JSON object. The same format is also accepted under its earlier top-level name, output_format; when both fields are sent, output_config.format wins.
On gRPC and ZMQ workers, SMG enforces the schema with constrained decoding: the engine is held to the schema as it generates, so the returned text block is JSON that fits it. The engine needs a grammar backend (TokenSpeed and TensorRT-LLM must have one switched on — see Enforcement). Three exceptions:
- A
tool_choicethat forces a tool call —anywhen at least one custom tool is declared, ortoolnaming one of the custom tools — takes precedence: the turn ends in a tool call, not a text answer, so the format constrains nothing. - On a model the
hfSymphony table reads, a format the table has no tag for is refused with400(invalid_parser_configuration), unless the worker takes the whole-turn grammar described there. - MLX workers do not support constrained decoding and refuse a request that asks for a format with
400(invalid_request_parameters).
HTTP workers and the Anthropic API receive output_config and output_format as sent; enforcing the format is left to them.
Streaming¶
Set "stream": true to receive the response as Server-Sent Events. Each event has an event: line naming its type and a data: line with the JSON payload. The stream ends with message_stop; there is no data: [DONE] line.
| Event | Payload |
|---|---|
message_start |
The message, with empty content; on gRPC and ZMQ workers its usage already carries the rendered prompt's token count |
content_block_start |
A new content block at index: thinking, text, or tool_use |
content_block_delta |
A text_delta, thinking_delta, or input_json_delta (a fragment of the tool input JSON) for the block at index |
content_block_stop |
The block at index is complete |
message_delta |
stop_reason, stop_sequence, and the final usage |
message_stop |
End of the message |
ping |
Keep-alive. Relayed from upstream when a backend sends one; SMG itself writes one on any backend's stream after --sse-keepalive-secs (default 15) seconds of silence |
error |
A failure after the stream started; see Errors |
HTTP workers and the Anthropic API produce their own events, which SMG relays; the Anthropic API can send block and delta types beyond this table. A gRPC worker's stream for a short text answer looks like this:
event: message_start
data: {"type":"message_start","message":{"id":"msg_abc123","type":"message","role":"assistant","content":[],"model":"meta-llama/Llama-3.1-8B-Instruct","stop_reason":null,"stop_sequence":null,"stop_details":null,"usage":{"input_tokens":27,"output_tokens":0,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"service_tier":"standard","inference_geo":"not_available"},"container":null}}
event: content_block_start
data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}
event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"There is no single answer"}}
event: content_block_stop
data: {"type":"content_block_stop","index":0}
event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null,"stop_details":null,"container":null},"usage":{"output_tokens":112,"input_tokens":27,"cache_creation_input_tokens":0,"cache_read_input_tokens":0}}
event: message_stop
data: {"type":"message_stop"}Count Tokens¶
POST /v1/messages/count_tokensReturns the number of input tokens a Messages request would use, without generating. Who counts depends on the backend: on gRPC and ZMQ workers, SMG counts the tokens itself from the model's tokenizer; on HTTP workers, it forwards the request to a worker that implements the endpoint and relays the worker's response. See Backend Support.
Request Body¶
| Field | Type | Required | Description |
|---|---|---|---|
model |
string | Yes | Model ID; selects the worker (HTTP) or the tokenizer (gRPC/ZMQ). |
messages |
array | Yes | Conversation, in the same format as /v1/messages. |
system |
string or array | No | System prompt. |
tools |
array | No | Tool definitions. |
tool_choice |
object | No | Tool selection. |
thinking |
object | No | Thinking configuration. |
output_config |
object | No | effort shapes the rendered prompt as it does on /v1/messages; a format is accepted but takes no part in the count. |
There is no max_tokens. Any other field, such as context_management or mcp_servers, is forwarded unchanged to HTTP workers; on gRPC and ZMQ backends such fields are accepted but take no part in the count, which uses only the fields above.
Example¶
curl http://localhost:30000/v1/messages/count_tokens \
-H "Content-Type: application/json" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "meta-llama/Llama-3.1-8B-Instruct",
"system": "You are a concise assistant.",
"messages": [
{"role": "user", "content": "What is the meaning of life?"}
]
}'{
"input_tokens": 27
}Backend Support¶
| Deployment | Behavior |
|---|---|
| HTTP workers | Routed like any other request (routing policy, retries, circuit breaker, load tracking) to one worker's /v1/messages/count_tokens. 404 (model_not_found) when no worker serves the model. |
| HTTP PD | Sent to one prefill worker. No decode worker is involved and no bootstrap fields are added. 503 (no_prefill_servers) when no prefill worker is available. |
| gRPC or ZMQ workers | Answered by SMG itself from the model's tokenizer; no worker takes part, so regular, PD (--pd-disaggregation), and EPD (--epd-disaggregation) deployments all serve it. See below. |
| Anthropic API backend | 501 Not Implemented |
| IGW mode | A model with an external worker goes to that provider's backend (501). Otherwise any of the model's self-hosted pools answers — gRPC, ZMQ, HTTP, and PD alike, a disaggregated pool qualifying by its prefill workers alone. A model no pool serves, including an unknown model, gets 404 (model_not_found). |
On gRPC and ZMQ backends, SMG renders the request through the model's chat template exactly as /v1/messages renders it — the same tool filtering by tool_choice and the same media placeholders — and returns the rendered prompt's token count. Media is not fetched: an image counts as the placeholder the template writes for it, and a document counts as its text (a document source with no text is refused, as /v1/messages refuses it).
That path refuses a request the way /v1/messages would. An unknown model gets 404 (model_not_found), even when a removed worker's tokenizer is still registered.
A served model whose tokenizer the gateway is still loading gets 503 (tokenizer_not_ready) with a Retry-After: 1 header. A served model whose tokenizer will not load on its own — no path to load it from, the load failed, or the tokenizer was removed — gets 503 (tokenizer_load_failed) with no Retry-After and a message naming the reason. Retrying changes nothing until a tokenizer registers for the model, either by a worker re-registering with a usable path or by POST /v1/tokenizers. Under --disable-tokenizer-autoload, a served model with no tokenizer and no such load failure gets 500 (tokenizer_not_found).
A request that does not render gets 400 (process_messages_failed), and a rendered prompt that fails to encode gets 500 (tokenization_failed).
Request Headers¶
Headers specific to the Messages API:
| Header | Effect |
|---|---|
X-SMG-MCP |
enabled makes SMG run the MCP tool loop itself for a request whose tools include an mcp_toolset entry, on the Anthropic API backend. See MCP Tools and MCP. Other backends ignore it, and SMG does not forward it. |
anthropic-version |
Checked on /v1/messages and /v1/messages/count_tokens before routing: the value must be a real calendar date of 2023 or later, written YYYY-MM-DD. Anything else, including an impossible date such as 2023-02-30, is refused with 400 (invalid_anthropic_version). A request without the header is served, and a valid value does not change how SMG handles the request. Forwarded to HTTP and PD workers and to the Anthropic API. |
anthropic-beta |
Read on /v1/messages for one rule: an entry starting with interleaved-thinking lets thinking.budget_tokens be at or above max_tokens; without it, such a budget is refused with 400 (thinking_budget_exceeds_max_tokens). Forwarded to HTTP and PD workers and to the Anthropic API. |
x-api-key |
Not used to authenticate to SMG. Forwarded only to the Anthropic API; a gateway key sent there is removed at authentication. See Authentication. |
Like the request-body rules, the anthropic-version and anthropic-beta checks are skipped for a /v1/messages request whose body the HTTP router streams through unparsed (see HTTP and PD Workers). A /v1/messages/count_tokens body is always buffered, so its anthropic-version check always runs.
Forwarded Headers¶
For HTTP and PD workers, SMG forwards only these request headers (it sets Content-Type itself):
| Header | Purpose |
|---|---|
Authorization |
Credentials for the worker; see Authentication |
anthropic-version, anthropic-beta |
API version and beta features for engines that serve /v1/messages |
x-request-id, x-correlation-id, and any header starting with x-request-id- |
Request correlation |
traceparent, tracestate |
W3C trace context |
x-smg-routing-key |
SMG routing hint |
x-override-priority |
SGLang's request priority, read by an engine started with SGLANG_ENABLE_REQUEST_HEADER_OVERRIDES=true and not by SMG |
Every other request header, including x-api-key, is dropped. The Anthropic API backend forwards a different set — Authorization, x-api-key, anthropic-version, and anthropic-beta — and nothing else. When the request carries neither Authorization nor x-api-key (a gateway key is removed at authentication), SMG adds the worker's stored key as x-api-key, with anthropic-version: 2023-06-01 when that header is absent. gRPC and ZMQ workers receive a tokenized request, not HTTP headers.
gRPC and ZMQ Workers¶
With gRPC workers (SGLang, vLLM, TensorRT-LLM, TokenSpeed, or MLX), and ZMQ workers that use the same pipeline, SMG implements the Messages API itself. It renders the conversation with the model's chat template, tokenizes it, sends token IDs to the engine, and builds the Anthropic response from the generated tokens. The pipeline serves regular, PD (--pd-disaggregation), and EPD (--epd-disaggregation) deployments. SMG needs the model's tokenizer (--model-path); see gRPC Workers.
| Request field | Handling |
|---|---|
system, messages |
Rendered with the chat template. Text and image blocks in user turns, tool_use and thinking blocks in assistant turns, and tool_result blocks are converted. A system-role message inside messages keeps its position; a template that only renders a leading system prompt gets such a message as a user turn at its position. Images go through SMG's multimodal pipeline. |
max_tokens, temperature, top_p, top_k |
Sent to the engine as sampling parameters. A max_tokens of 0 never reaches the engine; see max_tokens of 0. |
stop_sequences |
Sent as stop strings to vLLM, TensorRT-LLM, and TokenSpeed gRPC workers. For SGLang gRPC workers and ZMQ workers, SMG matches them itself and sends only single-token stops to the engine, as stop token IDs. |
tools |
Custom tools (those with an input_schema) are passed to the chat template and the tool parser; with a tool_choice of tool, only that tool is passed to the chat template. Other tool types, such as mcp_toolset, bash, text editor, web search, and tool search, are dropped. |
tool_choice |
Mapped to Chat Completions semantics: auto to auto, any to required, tool to that function, none to none. SMG enforces any and tool with constrained decoding; with disable_parallel_tool_use: true, the any constraint is held to one call only on models a Symphony table reads. On the legacy parsers the constraint still allows several calls, which the engine may generate and bill in usage.output_tokens. Whatever the choice, only the first tool call is returned (see Tool Use). |
thinking |
enabled and adaptive turn on the chat template's thinking mode and run the model's reasoning parser. disabled and between_tools turn thinking mode off. Without thinking, the template default applies. A display of omitted withholds the reasoning text: the thinking blocks and their signature stay in place, with empty text; in a stream, the thinking block opens and closes without thinking_delta events. With summarized, updates, or no display, the reasoning text is returned as the model wrote it; SMG does not summarize. budget_tokens is not sent to the engine. |
output_config.format, output_format |
The answer's JSON schema is sent to the engine as a constraint and enforced with constrained decoding, unless tool_choice forces a tool call; see Structured Output. |
output_config.effort |
Rendered into the prompt as the chat template's reasoning_effort; not itself sent to the engine. |
metadata, service_tier, container, mcp_servers, output_config's other keys, extra fields such as context_management |
Not sent to the engine. |
MLX workers answer 400 to a request with images, with stop_sequences, with a structured output format (output_config.format or output_format), or with a tool_choice of any or tool.
How the response is built:
- Tool calls and reasoning are extracted by SMG's tool and reasoning parsers, which SMG picks for the model or you set with
--tool-call-parserand--reasoning-parser. Athinkingblock comes first, then text, thentool_useblocks. Thinking blocks carry an emptysignature, and everytool_useblock carries"caller": {"type": "direct"}— the call is the model's own. tool_useids use Anthropic'stoolu_prefix: a parser id of the formcall_<suffix>is returned astoolu_<suffix>, the same in streaming and non-streaming responses. Model-specific id formats are returned unchanged.stop_reasonis the first of these that applies:max_tokenswhen the token limit was reached (tool calls in the truncated output do not override it),tool_usewhen the output has tool calls,stop_sequencewhen one ofstop_sequencesended generation (the matched string is instop_sequence), andend_turnotherwise.usagesplits the prompt over the engine's prefix cache:input_tokensis the prompt outside the cache andcache_read_input_tokensthe part the cache served, their sum the whole prompt.cache_creation_input_tokensstays0— SMG writes no cache entries of its own — and every cache counter is an integer, nevernull. Thecache_creationbreakdown,service_tier(standard), andinference_geo(not_available) are always written, andoutput_tokens_detailswith itsthinking_tokenswhen the turn ran with thinking. In a stream,message_startcarries the rendered prompt's whole token count asinput_tokenswith zero cache counters; the finalmessage_deltacarriesoutput_tokensand the engine's measured split ofinput_tokensandcache_read_input_tokens(input_tokensisnullwhen the engine never reported its count).- Context window: when the selected worker advertises its context length, SMG rejects a prompt with more tokens than the window with
400context_length_exceededbefore dispatch. In PD mode the smaller of the prefill and decode windows applies. Whether the prompt plusmax_tokensmust fit is left to the engine. - Tenant rate limiting, when enabled, covers this endpoint; see Tenant Rate Limiting.
HTTP and PD Workers¶
With HTTP workers, SMG forwards /v1/messages to the selected worker. It normally buffers the request, validates it, and re-serializes it: fields SMG does not model are kept, a --model-alias name is replaced by the canonical model ID, and numbers keep the values the client sent. When the HTTP router instead streams the body through (see Request Streaming), it forwards the raw body and leaves validation to the worker. Either way, only the allowlisted headers are forwarded. The worker's successful responses and SSE streams are relayed unchanged; a worker's error keeps its status and headers, but its body leaves in the Messages error envelope.
With --pd-disaggregation and HTTP workers, /v1/messages uses the same prefill/decode dual dispatch as chat completions: SMG selects a prefill and decode pair and dispatches the request to both. See PD Disaggregation.
On both deployments, a parsed body with max_tokens: 0 is refused with 400 before a worker is selected; see max_tokens of 0.
Anthropic API Backend¶
Point SMG at the Anthropic API:
smg launch --backend anthropic --worker-urls https://api.anthropic.comClients then call SMG as in the Anthropic API example.
- SMG sends each request to
<worker URL>/v1/messages, so the worker URL is the API origin without/v1. The request carries the client'sx-api-key,Authorization,anthropic-version, andanthropic-betaheaders; when it arrives with neither key header, SMG adds the worker's stored key asx-api-key(see Authentication). - Streaming responses are relayed as they arrive. Non-streaming responses are parsed into SMG's Messages schema and re-serialized; a body that does not parse returns
502with codeparse_error. - An upstream error is returned with its status, its error type and message kept in the error envelope, minus the organization and project identifiers in the message (see Provider Error Responses). An error Anthropic answers as an SSE stream is relayed as it arrives.
/v1/messages/count_tokensreturns501.
MCP Tools¶
SMG can run MCP tools for Messages requests on the Anthropic API backend. gRPC workers drop mcp_toolset tools, and HTTP workers receive them in the request body.
SMG runs the tool loop itself when the request has the X-SMG-MCP: enabled header and tools contains at least one mcp_toolset entry. Without the header, mcp_servers and the mcp_toolset entries are forwarded to Anthropic, together with the client's anthropic-beta header, and Anthropic's MCP connector runs the tools.
curl http://localhost:30000/v1/messages \
-H "Content-Type: application/json" \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "X-SMG-MCP: enabled" \
-d '{
"model": "claude-sonnet-4-20250514",
"max_tokens": 1024,
"messages": [
{"role": "user", "content": "Search the web for SMG and summarize the top result."}
],
"mcp_servers": [
{"type": "url", "name": "search", "url": "https://mcp.example.com/mcp"}
],
"tools": [
{"type": "mcp_toolset", "mcp_server_name": "search"}
]
}'| Field | Description |
|---|---|
mcp_servers[].name |
Server name that mcp_toolset entries refer to. |
mcp_servers[].url |
http:// or https:// URL. SMG connects over Streamable HTTP. A URL containing /sse selects the legacy SSE transport, which SMG can no longer connect to. |
mcp_servers[].authorization_token |
Sent to the MCP server as Authorization: Bearer <token>. |
mcp_toolset.mcp_server_name |
The server whose tools the model may call. |
mcp_toolset.default_config.enabled |
Whether the server's tools are enabled (default true). |
mcp_toolset.configs |
Per-tool settings keyed by tool name. When present, only the tools listed here are exposed, each if its enabled is true (or unset, falling back to default_config.enabled). |
defer_loading in default_config or configs |
Copied onto the tools SMG passes to the model. |
What SMG does:
- Connects to the servers in
mcp_servers. If none connects, it returns502with codemcp_connection_failed. - Replaces
mcp_serversand themcp_toolsetentries with the enabled MCP tools as regular tools, and setstool_choicetoautoif the request had none. - Runs each tool the model calls, sends the results back to the model, and repeats until the model stops calling tools, for at most 10 rounds of tool calls. If the model still calls tools after the 10th round, a non-streaming request fails with
502(mcp_max_iterations). A stream ends with anerrorevent right after the 10th round, without calling the model again. - Returns each call as an
mcp_tool_useblock (id prefixedmcptoolu_) followed by itsmcp_tool_resultblock, then the model's final content:
{
"content": [
{
"type": "mcp_tool_use",
"id": "mcptoolu_01AbCdEf",
"name": "web_search",
"server_name": "search",
"input": {"query": "SMG"}
},
{
"type": "mcp_tool_result",
"tool_use_id": "mcptoolu_01AbCdEf",
"content": [{"type": "text", "text": "..."}],
"is_error": null
},
{"type": "text", "text": "The top result describes..."}
],
"stop_reason": "end_turn"
}In a stream, SMG forwards the upstream events, emits the model's tool calls as mcp_tool_use blocks and each result as an mcp_tool_result block, and ends with a single message_delta and message_stop. SMG does not accept mcp_tool_use or mcp_tool_result blocks in messages (see the accepted block types under Request Body).
Tool calls use policy-only approval: SMG never pauses the request to ask the client. Each call is allowed or denied by the policy section of the MCP configuration file; with no policy configured, every tool is allowed. A denied call does not run: the model receives the denial as the tool's result, and the response reports it in the call's mcp_tool_result block with is_error set. See MCP for server configuration and the policy file.
Errors¶
Every error on /v1/messages and /v1/messages/count_tokens leaves in Anthropic's error envelope, whatever produced it inside — a request-body refusal, gateway authentication or admission, routing, the gRPC pipeline, the MCP tool loop, or a relayed backend body. For example, an over-long prompt on a gRPC worker:
{
"type": "error",
"error": {
"type": "invalid_request_error",
"message": "This model's maximum context length is 32768 tokens. However, your request has 40210 input tokens. Please reduce the length of the input."
},
"request_id": "msg_abc123"
}error.typefollows the status:authentication_errorfor a401,permission_errorfor a403,not_found_errorfor a404,request_too_largefor a413,rate_limit_errorfor a429,overloaded_errorfor a503(or an upstream529),timeout_errorfor a504,api_errorfor any other 5xx, andinvalid_request_errorfor any other 4xx.request_idrepeats thex-request-idresponse header.- An error the gateway produces itself carries its machine-readable code — the codes this page gives in parentheses, such as
context_length_exceededormodel_not_found— in theX-SMG-Error-Coderesponse header. A request-body refusal carries no code header; its message names the field and the rule it broke. - An error relayed from an HTTP or PD worker keeps its status and headers. An error from the Anthropic API keeps its status but not its upstream headers —
retry-afterand theanthropic-ratelimit-*headers are dropped — unless Anthropic answers the error as an SSE stream, which is relayed with its headers. A body already in the envelope keeps itserror.typeanderror.messageand gains therequest_id; any other body is wrapped, with its error message carried over. An Anthropic error message first loses its organization and project identifiers (see Anthropic API Backend).
A /v1/messages/count_tokens body that does not parse or fails validation is refused with the same 400 envelope as a /v1/messages body.
Errors that SMG detects before the stream starts are returned as ordinary HTTP error responses, not as a stream, even when stream is true. After the stream has started, the gRPC path and SMG's MCP tool loop report a failure as an Anthropic error event and end the stream. The MCP tool loop starts its stream before it first calls the model, so even an upstream error on that first call arrives this way:
event: error
data: {"type":"error","error":{"type":"api_error","message":"..."}}HTTP workers and the Anthropic API stream their own error events, which SMG relays.