Skip to content

critique.sh/api

Critique Developer APIs

One API key, three ways to build. Whether you need raw model speed, sandboxed execution, or automated pipeline control, Critique routes you straight to the right surface.

One control plane · One key family · Three surfaces

Start at the gateway.
Pick the layer that matches the job.

Inference is token-in token-out. Coding Agent is remote execution with patch and PR delivery. Platform API is the orchestration layer for passports, MCP, and automation around your merge boundary. The point of this page is to stop forcing developers to reverse-engineer that split.

Route map

POSTRaw model tokens
/api/v1/chat/completions
POSTSandboxed coding runs
/api/v1/coding-agent/runs
POSTPlatform control + MCP
/api/mcp

Auth is still Authorization: Bearer crt_.... The split is product shape, not account sprawl.

What do you need?

Do not make developers hunt through docs just to discover whether they need completions, a live sandbox, or platform control. Put the choice at the top.

Raw model tokens

Inference API

Devs building their own agent loops, sidecars, or eval harnesses who just want raw completions.

The hook

Fully OpenAI-compatible. Change your baseURL to critique.sh/api/v1, drop in your crt_ key, and keep your existing SDK.

  • DeepSeek V4 Flash, Tencent Hy3 Preview, and frontier model routes on the same credit pool.
  • Training opt-in pricing can cut DeepSeek V4 Flash and Hy3 token rates by 75%.
  • Best fit when you want speed, model choice, and no managed sandbox in the middle.

Managed sandbox runtime

Coding Agent API

Teams who want Critique to spin up a cloud coding agent, write a patch, and open a draft PR over HTTP.

The hook

Remote coding runs execute inside ephemeral E2B sandboxes, and warm sessions stay alive for fast follow-up prompts.

  • Prompt in, patch and PR out, with status polling, streaming, and follow-up messages on the same run id.
  • Reuses our OpenCode and Builder runtime under the hood, without making you operate sandbox infra.
  • Best fit when the output needs to be code, validation, and optionally a GitHub branch or draft PR.

Pipelines and orchestration

Platform API

Devs automating GitHub actions, exporting merge passports, or spinning up Model Context Protocol integrations.

The hook

Programmatically control the Critique platform with the same crt_ key family.

  • Trigger review-adjacent automation, export signed passport bundles, and list or fetch governed change records.
  • Manage MCP access over HTTP for external agents that need Critique context without living in our UI.
  • Best fit when you need control-plane plumbing, policy evidence, and integration hooks around the coding loop.

One crt_ key across the stack

New keys already carry the scopes needed for inference and coding-agent automation, while platform-facing scopes control passport and MCP access. The auth story should feel unified because it is.

ScopeSurfacePrimary routes
Inference APIGET /models, POST /chat/completions
Coding Agent APIruns, follow-ups, status, stream, models
Platform APIpassports, exports, MCP-readable control data

The unified developer experience

This is where Critique stops looking like three disconnected endpoints and starts looking like a real platform: the same auth model, the same billing logic, and a clear control-plane story around code generation and safe merge.

One crt_ key

Scopes handle access across inference, coding-agent runs, passports, and MCP. Rotate one auth primitive instead of juggling separate tokens for every surface.

One credit economy

Inference tokens and managed sandbox runs draw from the same Critique balance. On coding-agent runs, you can switch model billing to your OpenRouter account when you want cheaper BYOK economics.

One client shape

The stack already shares auth, billing language, and route conventions, which is what makes a unified SDK path realistic for completions, sandbox launches, and passport status checks.

One merge boundary

Coding agents create code, inference powers sidecars, and the platform layer records the governed result through passports, review runs, and MCP-readable state.

The destination hierarchy is simple: use /api as the gateway, keep /inference-api and /coding-agent-api as focused product pages, and let the platform docs own the orchestration details.