Agent integrations
Start with the client you already use, point it at a governed OpenAI- or Anthropic-compatible endpoint, and keep the path to budgets, policy and traceability open as usage grows. These public recipes explain the workflow; actual key creation and configuration happen after sign-in.
POST /v1/chat/completionsmodel "selected-id"region "allowed"Capabilities and boundaries
Each capability below describes its operating scope so technical and commercial teams can evaluate the same proposition.
Treat each recipe as a versioned compatibility record. A named client is ready for a stronger claim only after a clean configuration, one successful request and a recorded completion receipt.
The Anthropic-compatible path and OpenClaw have documented connection paths today. Each recipe is designed to get from configuration to a successful request in minutes; named client claims remain versioned evidence work.
Cline, Roo Code, Aider and similar tools can use the documented OpenAI-compatible path where their current version exposes a custom base URL. Codex-native clients are tracked separately because their response contract differs.
Codex CLI and Desktop already have a verified current-version TaaS reference using the native Responses endpoint and file-backed provider auth. Hermes is a named custom-provider validation track; Grok Bot (stable Cursor app 0.75.5) is a separate xAI app product with no documented customer-selectable model endpoint. Its Enterprise beta Inference Hooks are a possible policy integration, not an inference provider route.
A personal harness connection can become a governed team workload without replacing the application. Move policy to the control plane as adoption grows.
Five-minute integration path
Once signed in, follow the same short path whether you are testing Claude Code, OpenClaw, Codex or another compatible harness. The service keeps the key and request evidence in your account; the recipe runs from your machine.
Start with a free account, then verify your email and open the Get Started guide.
Choose a live model, region and optional budget at the key boundary. The secret is shown once and stays on your side.
Copy the harness-specific command, replace the local placeholder and make one small request. Successful terminal execution is the completion signal.
Move from a personal test to projects, central billing, BYOK, policy routing and traceable usage when the workload becomes a team service.
Copy, configure, verify
Use a named recipe when it is verified. For validation tracks, keep the client version visible and treat the first successful request as the gate for public compatibility claims.
Transport: Anthropic messages · Evidence: endpoint smoke-tested; named Claude Code CLI version pending.
Claude Code and other Anthropic-compatible clients can use the same governed path; verify the exact client version before publishing a compatibility claim.
export ANTHROPIC_BASE_URL="https://taas.cloudsigma.com"
export ANTHROPIC_AUTH_TOKEN="sk-taas-..."
export ANTHROPIC_MODEL="auto:orchestrator"After sign-in, the Get Started wizard provides the same local setup flow. Review the first-party script before running it: Claude Code setup script.
Transport: OpenAI chat completions · Evidence: isolated local turn completed.
Use the OpenAI-compatible provider and route through the governed catalogue.
baseUrl: https://taas.cloudsigma.com/v1
api: openai-completions
model: taas/autoTransport: OpenAI-compatible /v1 · Evidence: client-specific validation required.
Where the current client exposes a custom OpenAI-compatible base URL, use /v1 and verify with a small request before a longer run. Codex uses the separate native recipe below.
base_url = "https://taas.cloudsigma.com/v1"
model = "auto"
api_key = "sk-taas-..."Transport: native Responses at /v1/codex · Evidence: current CLI completion receipt; Desktop shares the provider surface.
This is the redacted native provider pattern used by the current CloudSigma TaaS Codex installation. The current Codex CLI path has passed a disposable-key native completion test, and Desktop uses the same provider surface. The CLI and Desktop app-server share $CODEX_HOME/config.toml; load the key through a secret-manager command.
model_provider = "cloudsigma"
model = "gpt-6-luna"
[model_providers.cloudsigma]
name = "CloudSigma TaaS"
base_url = "https://taas.cloudsigma.com/v1/codex"
wire_api = "responses"
supports_websockets = false
[model_providers.cloudsigma.auth]
command = "/bin/cat"
args = ["/secure/path/to/taas-api-key"]After sign-in, the Get Started wizard provides the same local setup flow. Review the first-party script before running it: Codex CLI/Desktop setup script.
Hermes: OpenAI-compatible /v1 · named custom-provider text-completion smoke receipt for NousResearch/hermes-agent v0.22.5; tool/function and explicit streaming validation pending.
Grok Bot: stable Cursor app 0.75.5 has no documented customer-selectable model endpoint. Enterprise Inference Hooks are a separate policy-interception opportunity, not a TaaS inference route.
# Validation fixture — NousResearch/hermes-agent v0.22.5
# ~/.hermes/config.yaml
providers:
taas:
provider: "custom: taas"
api: "https://taas.cloudsigma.com/v1"
key_env: "TAAS_API_KEY"
transport: "chat_completions"
model:
default: "auto"
provider: "custom:taas"Build resources
Use the maintained API references and practical guide for current request formats, authentication and supported model identifiers.
Make the next decision concrete
Review workload, control, spending, location and integration requirements with our team.