Loading ToolBench
Sovereign Ollama Bridge — MCP server for local and cloud Ollama models. One file. Two deps. Every model.
Ollama-Omega demonstrates strong tool definitions with comprehensive descriptions, proper JSON schemas, and clear parameter documentation. All 6 tools are explicitly registered with names, descriptions, and input schemas. Descriptions are well-written (average ~280 chars, within baseline 194 range upper quartile) and include context about WHEN to use each tool. Parameter types and descriptions are present throughout. Output schemas are documented in the code. However, there are gaps: no output schemas are exposed in the tool registration (only in code comments), error handling lacks actionable recovery guidance, and some advanced patterns (e.g., pagination for list_models, batching) are absent. Security is handled well (no credentials in params, proper timeouts), but the lack of visible confirmation steps for destructive operations (ollama_pull_model is WRITE but has no dry-run mechanism) is a notable gap.
Send a message sequence to an Ollama model and receive a streaming or buffered chat response. Use this tool for multi-turn conversations with an LLM. Streaming is enabled by default; set stream=false in arguments to buffer the entire response before returning. For single, non-conversational requests, use ollama_generate instead.
Request a single completion from an Ollama model given a text prompt. Use this tool for non-conversational, single-request text generation (e.g., summarization, code completion, creative writing). For multi-turn conversations, use ollama_chat instead. Streaming is enabled by default; set stream=false in arguments to buffer the entire response before returning.
Check Ollama daemon connectivity and list currently running models. Use this tool as the first call to verify the Ollama service is reachable before calling any other tool in this server. Do not use this to list all installed models — use ollama_list_models instead. Behavior: Read-only, idempotent, safe to retry. No authentication required. No rate limits. Makes a single HTTP GET to the Ollama daemon. On connection failure returns an error object without throwing.
List all Ollama models installed on the local machine with their memory load status. Use this tool to discover available model names before calling ollama_chat, ollama_generate, or ollama_show_model. Do not use this to check if a specific model is installed — use ollama_show_model with the model name instead.
Output schemas are documented in code comments (HEALTH_OUTPUT, LIST_MODELS_OUTPUT, etc.) but not exposed in the MCP tool definitions returned to the client. LLMs cannot see what fields to expect from tool responses, breaking downstream planning and field extraction.
Error responses use generic JSON structure without actionable recovery guidance. Examples: 'Invalid JSON from Ollama', 'Missing required argument'. Per pattern:recovery-guide, errors should tell the LLM what to do next (e.g., 'Model not found. Call ollama_list_models to see available models.').
ollama_list_models and ollama_chat with streaming responses lack pagination parameters (limit, offset, page). For models endpoints returning many results, unbounded responses waste tokens and risk context window exhaustion.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2025-06-18+ | v2 |
Download and install a model from the Ollama library or a custom registry. Use this tool to make a model available for inference. The model name must be a valid Ollama model identifier (e.g., 'llama3:8b', 'mistral:7b'). Download progress is streamed in real-time; the tool returns success only when the download is complete. Large models may take several minutes to download depending on network speed.
Retrieve metadata and configuration for an installed Ollama model, including its Modelfile, parameters, prompt template, and architecture details. Use this tool to understand a model's capabilities and settings before using it with ollama_chat or ollama_generate. If the model does not exist, an error is returned.
ollama_pull_model is a destructive operation (WRITE, downloads/modifies local state) with no dry-run, confirmation step, or undo mechanism. Per pattern:confirmation-request, irreversible operations should support confirmation before execution.
Tool descriptions use correct guidance (e.g., 'Use this tool as the first call', 'Do not use this to list all installed models, use ollama_list_models instead'), but do not include dependency hints or error recovery paths. Example: if ollama_chat fails due to model not being installed, the description should suggest calling ollama_list_models or ollama_pull_model.