Do tool calling, MCP, streaming, images and long context work through an LLM gateway subscription?
Yes for all five on Standard Compute. Tool calls, streaming and images pass through as sent, MCP servers keep working because they run inside your agent, and every served model has a context window of 1M tokens or more. It has no embeddings endpoint.
What passes through, feature by feature
| Feature | What happens |
|---|---|
| Tool calling | Yes. tools and tool_choice are forwarded unchanged on the OpenAI APIs, and tool_use and tool_result are translated on the Anthropic Messages API |
| Parallel tool calls | Yes. parallel_tool_calls is forwarded as sent. On the Messages API, disable_parallel_tool_use is not passed on |
| JSON mode | Forwarded. response_format, including json_schema with strict, is not changed. Claude and Mistral Large 4 may wrap JSON-mode output in a markdown fence |
| MCP servers | Yes. They run inside your agent, and their tools reach the model as ordinary tool definitions |
| Provider-hosted tools | No. Tools without an input schema, such as Anthropic web search, are dropped |
| Streaming | Yes, as server-sent events on all four inference endpoints, with token usage at the end of the stream |
| Images | Yes, up to 32 images per request at 10 MiB each. Routing only picks vision models, and a pinned text-only model returns HTTP 400 |
| Context | 1M tokens or more on every model. One request takes up to about 850K input and 50K output tokens |
| Prompt caching | Yes. The gateway sets Anthropic cache breakpoints itself; OpenAI and DeepSeek models cache automatically |
| Embeddings | No. There is no /v1/embeddings endpoint, so keep a separate provider for repository indexing |
What happens to very long prompts
Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 have 1M-token windows, GPT-5.6 and GPT-6 have 1.05M, and the Gemini, DeepSeek, GLM and Kimi models about 1.05M. A single request on a paid plan is capped lower, at about 850K estimated input tokens. Above that, the gateway drops middle turns and inserts a short summary; if the request is still too large, it returns HTTP 400 with context_length_exceeded. Context compression at OpenRouter is switched off.
Requests that need more than about 500K tokens of context only go to provider endpoints proven at around 1M tokens. Responses API requests with truncation disabled get the 400 instead of the trim.
Settings for each kind of agent
- Cline, Roo Code, Kilo Code, Aider, Continue and OpenCode: base URL
https://api.stdcmpt.com/v1, your key, and the modelstandardcomputeor an exact ID. - Claude Code: set
ANTHROPIC_BASE_URLtohttps://api.stdcmpt.comwithout /v1, or runstandardcompute launch claude. - Codex CLI: run
standardcompute launch codex, or add a[model_providers]entry to~/.codex/config.toml. - Test the one feature your workflow depends on with a small task, then find the request in Logs.
The catch
A gateway translates or forwards each request, and agents notice at the edges. On the Messages API, Anthropic server tools, images inside tool results and PDF document blocks do not pass through.
Where Standard Compute fits
Standard Compute is an OpenAI- and Anthropic-compatible LLM gateway for coding agents, with monthly plans from $19 (USD) and a fixed compute budget; at the budget, requests stop with HTTP 402. Every plan serves every model, and one key works in Claude Code, Codex CLI, OpenCode, Cline, Roo Code, Aider and Continue.
FAQ
Does tool calling work properly through an OpenAI-compatible gateway subscription? Cline leans hard on function calls.
On Standard Compute, yes: tools, tool_choice and parallel_tool_calls are forwarded unchanged on the OpenAI-compatible API at https://api.stdcmpt.com/v1, so Cline's function calls reach the model with the schema it sent. Set Cline's OpenAI Compatible provider to that base URL with the model standardcompute or an exact model ID.
Do parallel tool calls and strict JSON mode survive an OpenAI-compatible gateway?
On Standard Compute both are forwarded as sent: parallel_tool_calls and response_format, including json_schema with strict: true, are not rewritten on /v1/chat/completions or /v1/responses, which Codex CLI uses. Two caveats: Claude models and Mistral Large 4 treat json_object as advisory and can wrap the JSON in a markdown fence, and how strictly a schema is enforced depends on the model's provider.
Does Claude Code with a custom base URL still work with my MCP servers through a subscription gateway?
Yes. MCP servers run on your machine inside Claude Code, and their tools reach the model as ordinary tool definitions over the Anthropic Messages API, which Standard Compute serves at ANTHROPIC_BASE_URL=https://api.stdcmpt.com. What does not pass through are Anthropic's hosted server tools without an input schema, such as Anthropic web search.
Do flat-rate API services support streaming responses for OpenCode?
Standard Compute streams on every inference endpoint: /v1/chat/completions, /v1/responses, /v1/messages and /v1/completions, as server-sent events with token usage reported at the end of the stream. OpenCode lists Standard Compute natively through models.dev, so streaming works without extra settings.
Do subscription gateways pass through vision and image inputs, for example UI screenshots in Roo Code?
Standard Compute does: image_url content parts and Anthropic image blocks pass through, up to 32 images of 10 MiB each per request. With the routed model standardcompute, a request with images only goes to a vision-capable model from your selection; if you pin a text-only model, the request returns HTTP 400 instead of being sent elsewhere.
What context window do I get per model on a flat-rate subscription, and does it silently truncate?
On Standard Compute every served model has a window of at least 1M tokens, and a single request takes up to about 850K estimated input tokens and 50K output tokens. Above that, the gateway drops middle turns and inserts a short summary before sending, and returns HTTP 400 context_length_exceeded if the request is still too large; OpenRouter's own context compression is off. For huge monorepo chunks in Aider, keep the added files under that cap.
Which models on a subscription API have 200K+ context?
On Standard Compute all of them. Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 have 1M tokens, GPT-5.6 Sol, Terra and Luna and GPT-6 Astra 1.05M, and Gemini 3.1 Pro Preview, DeepSeek V4 Pro, GLM 5.3 and Kimi K3 about 1.05M, every one on every plan from $19 a month (USD). A single request is capped at about 850K input tokens.
Is prompt caching supported through the gateway when I use Claude Code?
Yes. Standard Compute sets Anthropic cache breakpoints itself on the Messages API that Claude Code uses, on the system prompt and the latest turns, and cache reads come back as cache_read_input_tokens. The budget is drawn at the cost the provider reports, so a cached read costs a fraction of fresh input, as on Anthropic's own API.
Do coding-agent subscriptions cover embeddings for Continue's repository indexing?
Not on Standard Compute: it serves chat, completions, Responses and Messages, and has no /v1/embeddings endpoint. Keep Continue's embeddings on a separate provider or a local embeddings model, and point chat and edit at https://api.stdcmpt.com/v1.