CoralBricks

CoralBricks is a hosted inference platform built for coding and research agents that work through long contexts, call tools, and take several steps to finish a task. It serves open models through familiar API formats, so a developer can point an existing client or agent at the CoralBricks gateway instead of building around a new protocol. The focus is less on a one-off chatbot response and more on the sustained workload of an agent loop, where a growing transcript is sent repeatedly and output speed, context capacity, cache behavior, and reliable streaming all affect whether the job feels usable.

The platform exposes three ways to run those workloads under the same key and model identifiers. Chat Completions is the conventional stateless option, with the application sending the full conversation on every request. The Responses API can carry thread state between turns through a previous response identifier, which saves the client from replaying the entire transcript. Stored responses can be retrieved for auditing or restoring an interface after a reload, and there is an endpoint to delete one when it's no longer needed. Background Responses is aimed at autonomous work that can run for minutes or longer. A client can start a job, poll it later, and resume its event stream without holding one HTTP connection open for the whole run. Each surface supports streaming, and the official documentation also covers tool calling for agent applications.

CoralBricks currently lists models from families including GLM, DeepSeek, and GPT-OSS. Its integration guides show GLM and DeepSeek variants with context windows up to one million tokens, while GPT-OSS 120B is listed with a 128,000 token context. Model capabilities come from the live models endpoint, including whether a model accepts image input, so applications don't have to assume that every model supports the same content. Unsupported content receives a clear error instead of being ignored. The service reports decode speed, first-token latency, and cache hit measurements over recent 30 minute and daily windows. A quiet model has no value for an unmeasured window rather than a misleading zero. That gives developers a practical way to inspect serving behavior rather than choosing solely from a model name or a theoretical benchmark.

Compatibility is a central part of the product. CoralBricks implements OpenAI-compatible Chat Completions, the OpenAI Responses API, and an Anthropic Messages surface. Official setup guides cover Codex CLI, Claude Code, OpenCode, Cursor, GitHub Copilot, Kilo Code, Pi, Cline, and other clients that accept a custom provider. In many cases setup consists of adding the CoralBricks base URL and an API key, then selecting a supported model identifier. The native Responses implementation matters for agent tools that can't use a chat-only provider without an additional translation layer. The same gateway also preserves streaming tool calls and reasoning controls expected by supported clients.

The pricing model is usage based and prepaid. Published integration tables quote separate input and output rates per million tokens, with different prices for each model. Cached input reads are free, which is particularly relevant to coding agents that resend a large, mostly unchanged prefix on every turn. Novel input can be retained as a cache write, and the service reports prompt, cached, cache-write, completion, and cost details in each response. Streamed calls include the cost information on the final usage-bearing chunk, so an application can record spend without estimating it separately. An authenticated quota endpoint returns the remaining credit balance. Requests that would run with a non-positive balance are rejected before inference, so this isn't presented as an unlimited free service or a subscription with hidden overages.

CoralBricks is a good fit for teams that already have an agent harness and want faster access to open reasoning models without operating GPU infrastructure. It also suits developers who need very long context, resumable background jobs, or direct compatibility with several popular coding interfaces. The narrower model catalog and account-level limits mean it won't replace a broad marketplace for every project, and the official API reference still describes access as a design-partner program. Limits for rate, concurrency, and context are assigned per key, with larger parallel workloads handled by contacting the team. For a buyer evaluating it now, the grounded proposition is clear. It's paid inference optimized around long-running agent traffic, free cache reads, and common agent API formats.