Which model providers does Runavelo support?
Runavelo includes presets for OpenAI, Anthropic, Google Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, and configurable custom endpoints. Actual capability depends on the selected model and protocol.
BRING YOUR OWN MODEL
Choose the AI provider. Keep the workflow as the durable result.
Configure OpenAI, Claude, Gemini, DeepSeek, Kimi, Qwen, or another supported endpoint for workflow building, explanation, and repair. Approved ordinary commands then run through Runavelo's workflow engine rather than depending on repeated model inference.
MODEL CHOICE WITHOUT PRETENDING MODELS ARE IDENTICAL
Configure the protocol as well as the provider.
OpenAI Responses, OpenAI-compatible Chat Completions, Anthropic Messages, and Gemini APIs represent messages, tool calls, streaming, and provider options differently. Runavelo adapters translate those protocols into one assistant experience while allowing relevant provider-specific configuration.
1. SelectChoose a provider preset or custom configuration.
2. ConfigureSet protocol, Base URL, key, model, and supported options.
3. ValidateTest model access and tool-assisted workflow output.
4. BuildCreate or repair a workflow that remains visible after the chat.
PROVIDER PRESETS
Start quickly without losing control.
Presets fill known defaults for providers such as OpenAI, Anthropic, Google Gemini, DeepSeek, Kimi, GLM, Qwen, and MiniMax. They do not lock the fields. Users can edit the Base URL and model where the integration allows it.
For a custom service, select the protocol it actually implements. “OpenAI-compatible” does not guarantee every optional parameter, model-list endpoint, streaming event, or tool-call behavior. Run a real tool-call test before relying on the configuration.
A STABLE TOOL BOUNDARY
Let models request capabilities, not arbitrary mutations.
The assistant exposes defined tools for reading workflow context, consulting supported command documentation, validating proposed JSON, applying controlled changes, and reading execution evidence. Arguments are checked by the application before a tool runs.
OpenAI, Anthropic, and Gemini all document schema-based tool or function calling, but each uses different request and response mechanics. Runavelo handles those differences at the provider adapter rather than spreading them through workflow business logic.
PORTABLE OUTCOME
The validated visual workflow outlives the model conversation.
A model may be used to design a difficult flow and then be unavailable, changed, or replaced. The approved commands still describe the operational procedure. Ordinary browser, spreadsheet, file, and control-flow steps run without asking the model to choose the same action again.
Different models may propose different edits, so switching providers is not a promise of identical chat behavior. Portability comes from validating and preserving the workflow artifact.
LOCAL KEY STORAGE
Avoid an unnecessary credential intermediary.
The Runavelo desktop client encrypts the provider API key and stores it on the local device. Requests go to the endpoint selected in the configuration instead of sending the key to Runavelo for proxying.
Create a dedicated provider key where possible, restrict permissions and spend, rotate suspected keys, and do not place credentials in workflow descriptions, logs, screenshots, or exported files. Local encrypted storage still depends on the security of the Windows account and machine.
EXPLICIT DATA BOUNDARY
Understand what the selected provider receives.
When the assistant generates or diagnoses a workflow, the context needed for that request is sent to the configured model provider. Minimize that context, redact secrets and unnecessary personal data, and review the provider's current retention and account controls.
Local workflow commands remain local unless the flow intentionally invokes an external API, cloud OCR or CAPTCHA service, publishing service, or another network capability.
FIT TEST
Evaluate a model with the work you actually need.
Test workflow generation, a precise one-field modification, nested loops, invalid enumerations, missing required parameters, long error stacks, sequential tool calls, cancellation, and a new conversation started while an old request is ending. Validate generated workflow JSON and resulting runtime code; a fluent answer alone is not enough.
Broad model support is not universal model support. The chosen model and endpoint must implement the protocol, context capacity, and tool behavior required by the assistant.
QUESTIONS
Bring-your-own-model FAQ
Can I enter a custom API URL and model name?
Yes. You can select a supported protocol, edit the Base URL, and select or manually enter a model name. A custom endpoint must implement the required parts of that protocol.
Where is my API key stored?
The desktop client encrypts the key and stores it on the local device. It uses the key to call the provider endpoint you selected rather than uploading it to Runavelo for proxying.
Will changing models change an existing workflow?
No automatic conversion is required for ordinary workflow execution. Models may generate different proposals, but the approved visual workflow remains the durable artifact.
ENGINEERING REFERENCES
Official provider and security documentation
- OpenAI function calling documents application-defined tools and JSON Schema inputs.
- Anthropic tool use explains client-executed tools and tool-result handling.
- Gemini function calling documents function declarations and compositional calls.
- OWASP Secrets Management guidance covers storage, rotation, auditing, and lifecycle controls.
Turn the conversation into an editable workflow.
Use AI where reasoning helps, keep exact workflow logic visible, and repeat ordinary operations without a fresh inference decision.