CloudSigma AI Mission Control
One plugin. Every leading AI model. Connect OpenClaw to CloudSigma Mission Control for GPT, Claude, Gemini, Kimi, Qwen and more—with streaming, tool calling, live model discovery, session affinity, optimised autorouter and prompt-cache-aware routing built in.
Install
openclaw plugins install clawhub:@cloudsigma/openclaw-taas-providerCloudSigma AI Mission Control
One plugin. Every leading AI model. Connect OpenClaw to CloudSigma Mission Control for GPT, Claude, Gemini, Kimi, Qwen and more—with streaming, tool calling, live model discovery, session affinity, optimised autorouter and prompt-cache-aware routing built in.
CloudSigma AI Mission Control is the first-class OpenClaw provider for CloudSigma TaaS. It includes API-key onboarding, a conservative static model catalog, OpenAI-compatible request handling, native OpenClaw session affinity, and browser WebRTC Talk through gpt-realtime-2.1.
What this plugin provides
- Provider id:
cloudsigma - Provider alias:
cloudsigma-taas - Default model:
cloudsigma/gpt-5.5 - Default endpoint:
https://taas.cloudsigma.com/v1 - Auth via
CLOUDSIGMA_API_KEY - OpenAI-compatible chat/completions transport
- Native session affinity and bounded request correlation for agent, tool, and simple-completion streams
- Streaming usage support
- Tool-call compatibility through OpenAI-style tool payloads
- Replay/history compatibility for OpenAI-compatible provider families
- Live model discovery from CloudSigma TaaS
/models, with safe static fallback - Conservative image-capability metadata: image support is advertised only for models verified or explicitly marked as vision-capable
- SSRF-safe model discovery restricted to the CloudSigma TaaS host
Install
openclaw plugins install clawhub:@cloudsigma/openclaw-taas-provider
[!IMPORTANT] This is the one-plugin CloudSigma setup. New installations must not also install
openclaw-taas-affinity; this provider owns thecloudsigmaruntime hooks and includes its supported affinity behavior. Legacy affinity-only installations can remain temporarily unchanged until migrated as a controlled configuration update.
OpenClaw install source:
clawhub:@cloudsigma/openclaw-taas-provider
ClawHub package page:
https://clawhub.ai/plugins/@cloudsigma/openclaw-taas-provider
Source repository:
https://github.com/cloudsigma/openclaw-taas-provider
Requirements
- OpenClaw
>=2026.7.1-2 - A CloudSigma Mission Control API key
- Network access to
https://taas.cloudsigma.com/v1
Configure
Set the CloudSigma Mission Control API key in the OpenClaw runtime environment:
export CLOUDSIGMA_API_KEY=...
Then use CloudSigma models with model refs like:
cloudsigma/gpt-5.5
cloudsigma/claude-sonnet-4.6
cloudsigma/gemini-3.1-flash-lite
The provider alias cloudsigma-taas resolves to the same provider auth as cloudsigma.
Session affinity and request correlation
CloudSigma TaaS uses a stable logical session key to preserve eligible upstream
continuations and prompt-cache locality. The provider forwards this only when
OpenClaw supplies its native per-conversation sessionId; it never creates an
identity from a workspace path, hostname, process state, time, prompt content,
or agent name.
Some generic/background and direct-simple OpenClaw paths carry the same native
identity through the public model_call_started lifecycle event rather than
options.sessionId. Where that hook is available, the provider correlates only
an exact valid W3C traceId + spanId from that event with the invocation's
traceparent header. Invocation options.sessionId wins, then a runtime
wrapper ctx.sessionId, then this exact trace bridge. The bridge is bounded to
1,024 entries, has a 30-minute sliding TTL, supports retry-safe reuse, and
fails closed for malformed or conflicting traces. It never uses timing,
workspace, process, agent, or environment state to correlate calls.
For each eligible HTTP/WebSocket transport turn, the plugin sends:
X-Session-IdandX-OpenClaw-Session-IdX-OpenClaw-Plugin-VersionX-OpenClaw-Agent-Idwhen it is encoded in the native session keyX-OpenClaw-Turn-IdandX-OpenClaw-Attempt
For both normal agent/tool streams and direct/simple completions, it adds the following OpenAI-compatible payload metadata when the caller has not already provided that field:
metadata.session_idmetadata.sticky_keymetadata.openclaw_correlation(schema/source/plugin version/session, agent when available, provider, and model)
Existing metadata fields are preserved, including caller-supplied session,
sticky, correlation, and requester_runtime values. A missing native session
means no affinity metadata is added. This makes unattributable traffic explicit
rather than silently conflating independent conversations.
AutoRouter response details and an optional AutoRouter algorithm preference are kept in a bounded in-memory map keyed by that same native session id. They are not shared across sessions and are not a replacement identity mechanism.
The plugin requests startup activation because its optional AutoRouter gateway RPC methods and trace lifecycle subscription must be registered in the live Gateway dispatch table. Provider-only lazy activation is not sufficient for those Gateway surfaces.
Privacy and security
Session identifiers are sent to CloudSigma TaaS because they are necessary for session affinity. The provider validates them as bounded, header-safe strings before forwarding. The correlation envelope contains no prompt text, workspace paths, hostnames, environment values, git metadata, secrets, or requester-bridge data. The plugin does not implement the retired requester bridge, does not poll external services, and does not persist affinity state locally.
Realtime Talk (WebRTC)
CloudSigma Talk is registered as realtime voice provider cloudsigma. It currently supports browser WebRTC only and uses the fixed model gpt-realtime-2.1.
[!IMPORTANT] In OpenClaw
2026.7.1-2,browserOriginis static plugin configuration; it is not inferred from each browser request. It must exactly equal the active OpenClaw UI origin and an origin on the CloudSigma TaaS allowlist. Scheme, hostname, and port must all match. If the UI is served from another origin (including another port), update this setting and the TaaS allowlist before using Talk. Do not use a page URL: paths, queries, fragments, credentials, and a trailing slash are rejected.
{
plugins: {
entries: {
cloudsigma: {
enabled: true,
config: {
// Optional. SecretInput literals and secret references are supported.
apiKey: { source: "env", provider: "default", id: "CLOUDSIGMA_API_KEY" },
browserOrigin: "https://openclaw.example.com",
realtimeRequestTimeoutMs: 25000
}
}
}
},
talk: {
realtime: {
provider: "cloudsigma",
transport: "webrtc"
}
}
}
Talk credential precedence is deliberately isolated from OpenAI authentication:
- Explicit
plugins.entries.cloudsigma.config.apiKey - A CloudSigma
api_keyauth profile CLOUDSIGMA_API_KEYTAAS_API_KEY
OpenAI API keys, OpenAI/Codex OAuth, and external OpenAI CLI credentials are never considered. The long-lived key remains server-side; the browser receives only the short-lived client secret and the fixed offer URL https://taas.cloudsigma.com/v1/realtime/calls.
apiKey, browserOrigin, and realtimeRequestTimeoutMs are authoritative only in plugins.entries.cloudsigma.config. Values with those names under talk.realtime.providers.cloudsigma are ignored; there are currently no supported provider-local Talk fields. OpenClaw materializes plugin SecretInput references before provider resolution, and the resolved configuration is then passed unchanged through readiness checks and browser-session minting. The realtime mint timeout defaults to 25 seconds and is clamped to 5–60 seconds; it remains a single bounded attempt and does not retry or reuse ephemeral credentials.
Browser sessions are minted with OpenAI-compatible audio.input.turn_detection server VAD, automatic response creation, and response interruption enabled. OpenClaw's vadThreshold, prefixPaddingMs, and silenceDurationMs Talk options are forwarded when set. This provider therefore advertises server-VAD barge-in support for its browser WebRTC transport.
Onboarding behavior
During OpenClaw provider setup, the plugin exposes an API-key auth choice:
- Label:
CloudSigma Mission Control API key - CLI flag:
--cloudsigma-api-key <key> - Environment variable:
CLOUDSIGMA_API_KEY - Default model:
cloudsigma/gpt-5.5
The plugin writes the CloudSigma provider catalog into OpenClaw config and retains the existing CloudSigma TaaS default-model alias for compatibility.
Static catalog included in the package
The package ships a conservative static catalog so OpenClaw can start even before live model discovery succeeds.
Current seed models include:
| Model ref | Capabilities | Context | Max output | Notes |
|---|---|---|---|---|
cloudsigma/gpt-5.5 | text | 400k | 128k | Default route |
cloudsigma/gpt-5.4 | text | 1.05M | 128k | Large-context GPT route |
cloudsigma/claude-fable-5 | text | 1M | 128k | Reasoning-capable Claude route |
cloudsigma/claude-opus-4.8 | text | 200k | 32k | High-end Claude route |
cloudsigma/claude-sonnet-4.6 | text | 1M | 64k | Balanced Claude route |
cloudsigma/gemini-3.1-flash-lite | text, image | 1M | 65k | Vision-capable Gemini route |
cloudsigma/kimi-k2.6 | text | 262k | 262k | Long-output route |
cloudsigma/qwen3-vl | text, image | 262k | 8k | Vision-language route |
cloudsigma/pixtral-large | text, image | 128k | 4k | Vision-language route |
cloudsigma/qwen2.5-vl-72b | text, image | 32k | 8k | Vision-language route |
cloudsigma/llama-3.3-70b | text | 128k | 8k | Open model route |
cloudsigma/qwen3-235b | text | 256k | 8k | Large Qwen route |
cloudsigma/glm-4.5-air | text | 128k | 8k | Lightweight GLM route |
Live discovery can supplement this list with additional CloudSigma TaaS chat-capable models when the API key is available.
Model discovery
At runtime, the plugin attempts to fetch CloudSigma TaaS model metadata from:
https://taas.cloudsigma.com/v1/models
Discovery behavior:
- If no usable API key is present, OpenClaw uses the static catalog.
- If the endpoint returns an error, malformed JSON, or no usable chat models, OpenClaw uses the static catalog.
- If discovery succeeds, live models supplement the static catalog.
- Static seed rows remain authoritative for known model ids, so verified defaults do not drift unexpectedly.
Security notes
- The plugin does not store API keys in the package.
- Authentication is delegated to OpenClaw provider auth via
CLOUDSIGMA_API_KEY/ setup flow. - Model discovery uses OpenClaw's SSRF-safe fetch helpers and restricts network access to the configured CloudSigma TaaS hostname.
- Talk client-secret minting is pinned to the exact
https://taas.cloudsigma.comorigin, denies redirects, bounds JSON and error bodies, and rejects conflicting GA/legacy credentials. - Realtime server bridges fail closed; the initial Talk implementation is browser WebRTC only.
- The package manifest declares no tools, services, HTTP routes, bundled skills, or background daemons.
- The plugin executes provider registration code only; it does not install native dependencies.
Compatibility metadata
This package declares the OpenClaw compatibility and build contract required for external ClawHub plugins:
{
"openclaw": {
"extensions": ["./dist/index.js"],
"providers": ["cloudsigma"],
"compat": {
"pluginApi": ">=2026.7.1-2",
"minGatewayVersion": "2026.7.1-2"
},
"build": {
"openclawVersion": "2026.7.1-2",
"pluginSdkVersion": "2026.7.1-2"
}
}
}
Validation and release gates
Before publishing, releases are expected to pass:
npm ci
npm run typecheck
npm run build
npm test
npx -y clawhub@0.21.0 package validate . --json
npx -y clawhub@0.21.0 package pack . --pack-destination artifacts --json
The test suite covers:
- Plugin registration
- Provider id, alias, auth metadata, and catalog registration
- Native transport headers (session, plugin version, agent, turn, and attempt)
- Session/sticky/correlation metadata for GPT, Claude, AutoRouter, and simple-completion paths
- Preservation of caller-provided metadata and no identity fallback without a native session
- Onboarding/default-model config writes
- Static catalog shape
- Live discovery success path
- Live discovery fallback on HTTP errors, network failures, malformed JSON, and empty model lists
- Conservative image-capability filtering
Troubleshooting
CLOUDSIGMA_API_KEY is missing
Set the key in the OpenClaw runtime environment and restart/reload the gateway process that loads plugins.
export CLOUDSIGMA_API_KEY=...
A model is missing from the picker
The plugin always ships a static seed catalog. Additional models require live discovery, which needs a valid key and a successful response from https://taas.cloudsigma.com/v1/models.
A vision model is not marked as image-capable
That is intentional unless the route is verified or the CloudSigma TaaS metadata explicitly declares vision/image support. The plugin avoids advertising image input for routes that are only text-safe.
Requests fail with auth errors
Confirm the same environment that runs OpenClaw has CLOUDSIGMA_API_KEY set. Shell-local exports do not help if OpenClaw is running under a service manager with a different environment.
Continuations do not stay on one TaaS route
Confirm that this provider plugin is enabled and that the OpenClaw request has a
native session id. The provider deliberately does not derive identity from a
workspace or process. On a new setup, remove any legacy
openclaw-taas-affinity plugin configuration so the single provider owns the
CloudSigma hook surface, then reload/restart OpenClaw through your normal
controlled operations process.
Package identity
- Package:
@cloudsigma/openclaw-taas-provider - Runtime/provider id:
cloudsigma - Alias:
cloudsigma-taas - License: MIT
- Publisher: CloudSigma
