Providers
SwarmClaw supports 25 built-in providers plus compatible custom endpoints. Each agent can use a different runtime.
CLI Providers
CLI providers spawn a local binary and let that runtime manage its own tool loop.
| Provider | Binary | Notes |
|---|---|---|
| Claude Code CLI | claude | Native tool/runtime loop through Claude Code. |
| OpenAI Codex CLI | codex | Native Codex runtime with session continuity and streamed events. |
| OpenCode CLI | opencode | CLI-backed coding/runtime provider. |
| Gemini CLI | gemini | CLI-backed Gemini runtime. |
| GitHub Copilot CLI | copilot | CLI-backed Copilot runtime with multi-model support through Copilot. |
| Factory Droid CLI | droid | CLI-backed Factory Droid runtime with droid exec stream-json output, session resume, and autonomy controls. |
| Cursor Agent CLI | cursor-agent | CLI-backed Cursor runtime with structured JSON output and resume support. |
| Qwen Code CLI | qwen | CLI-backed Qwen coding runtime with headless structured output. |
API Providers
API providers make direct HTTP calls.
Provider connection checks return a concise diagnostic timeline covering endpoint resolution, model discovery, fallback model selection, and final chat verification. Token-like values are redacted before those diagnostics are shown in setup, provider settings, or agent editing.
| Provider | Endpoint Family | Notes |
|---|---|---|
| Anthropic | api.anthropic.com | Claude models |
| OpenAI | api.openai.com | GPT and reasoning models |
| OpenRouter | openrouter.ai/api/v1 | Routed model catalog through an OpenAI-compatible API |
| Google Gemini | generativelanguage.googleapis.com | Gemini API models |
| DeepSeek | api.deepseek.com | DeepSeek models |
| Groq | api.groq.com | Groq-hosted open models |
| Together | api.together.xyz | Together-hosted open models |
| Mistral | api.mistral.ai | Mistral models |
| xAI | api.x.ai | Grok models |
| Fireworks | api.fireworks.ai | Fireworks-hosted models |
| Nebius | api.tokenfactory.nebius.com | Nebius-hosted models |
| DeepInfra | api.deepinfra.com | DeepInfra-hosted models |
Local, Gateway, and Runtime-Managed Providers
| Provider | Type | Notes |
|---|---|---|
| Ollama | Local or cloud | Explicit Local/Cloud mode. Local typically targets your own Ollama endpoint; Cloud uses Ollama's hosted path with credentials. |
| LM Studio | Local OpenAI-compatible API | Connect to a local or remote LM Studio server. Bare hosts are normalized to /v1 before discovery and chat. |
| OpenClaw | Gateway | Route execution through saved OpenClaw gateway profiles or direct gateway configuration. |
| Goose | Runtime-managed CLI | Run Goose through its local CLI/runtime stack while SwarmClaw preserves worker sessions and provider selection. |
| Hermes Agent | Runtime-managed API server | Connect to a local or remote Hermes /v1 endpoint while Hermes keeps control of its own tool/runtime loop. |
OpenClaw
OpenClaw is an autonomous agent runtime that SwarmClaw can deploy, connect to, and manage as part of a larger swarm.
In the agent editor, OpenClaw is enabled through the OpenClaw Gateway toggle instead of the normal provider/model dropdown.
Gateway Profiles
OpenClaw gateways can be managed centrally:
- create named gateway profiles with endpoint, credential, notes, tags, and deployment metadata
- use Smart Deploy for local bring-up, VPS bundles, or SSH-managed remote install
- mark one profile as the default
- activate, drain, cordon, or request restart for saved profiles before maintenance
- run discovery, health checks, and dashboard open flows from the control plane
- import, export, and clone saved gateway profiles
Agents can inherit the default profile, select another saved profile, or use direct gateway details when needed. Automatic OpenClaw routing skips profiles that are draining or cordoned so new work lands only on gateways marked active.
Fleet Visibility
SwarmClaw also tracks external OpenClaw runtimes:
- gateway health and verification state
- runtime counts and last-seen data
- external agent runtime lifecycle state
- remote lifecycle controls for SSH-managed installs
See OpenClaw Setup for the full deployment flow.
Hermes Agent
Hermes Agent integrates through its OpenAI-compatible API server instead of the OpenClaw websocket gateway path.
- default local endpoint:
http://127.0.0.1:8642/v1 - remote endpoints work the same way if they are reachable from the SwarmClaw host
- auth can be optional or enabled on the Hermes side
- Hermes stays in charge of its own tool/runtime loop, so SwarmClaw treats it as a runtime-managed worker provider
Goose
Goose integrates as a runtime-managed worker provider through the local goose CLI.
- SwarmClaw uses Goose's existing local configuration and auth state
- the default model entry is
default, but you can still type a model manually - Goose remains in charge of its own tool/runtime loop and extensions
- session continuity is preserved across turns through named Goose sessions
LM Studio
LM Studio integrates through its OpenAI-compatible API server.
- default local endpoint:
http://127.0.0.1:1234/v1 - auth is optional and follows the API key configured in LM Studio
- live model discovery reads the loaded model catalog from
/v1/models - agents can type or save local model IDs that are not part of the default list
Custom Providers
Add any OpenAI-compatible endpoint as a custom provider. Common examples include:
- local vLLM or LM Studio style endpoints
- internal OpenAI-compatible gateways
Custom providers appear in agent and chat dropdowns immediately after saving.
Model Selection
SwarmClaw can populate model lists from provider discovery, but you are not limited to discovered models. You can still type a newer or custom model name manually.
Credentials
Provider credentials are encrypted before storage. Add them from provider setup flows or from the secrets/credentials surfaces.
Ollama Mode Selection
Ollama transport is explicit:
- choose Local to talk to a local or self-hosted Ollama endpoint
- choose Cloud to use Ollama's hosted endpoint with a credential
Model names do not silently switch transport for you.
Model Failover
Agents can be configured with fallback credentials to retry through another credential when a provider fails with a rate-limit, auth, or server-side issue.
See Failover for the full model.
Provider Limitations
- CLI providers, OpenClaw, Goose, and Hermes Agent manage their own runtime/tool loop, so SwarmClaw does not expose the same local tool toggles for those agents.
- OpenClaw runs tools remotely on the gateway, not inside the local SwarmClaw process.
- Goose runs tools inside the Goose runtime, not inside the local SwarmClaw process.
- Hermes Agent runs tools inside the Hermes runtime behind its API server, not inside the local SwarmClaw process.
- Custom providers must be OpenAI-compatible.