Loading ToolBench
MCP server for HiveCompute — OpenAI-compatible inference router for the Hive agent economy. Chat completions, embeddings, and model listings. Hive routes to the cheapest qualifying model. Billed per token in USDC on Base L2. Real rails.
HiveCompute defines 5 tools with complete input schemas and descriptions. Tool names follow verb_noun convention (compute.chat, compute.embed, compute.list_models, compute.estimate_cost, compute.get_usage). Descriptions are substantive (100-200 chars), explaining WHAT each tool does and WHEN to use it. All parameters have type definitions and descriptions. However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. Error handling guidance is absent. Security concern: api_key and did are exposed as required parameters rather than server-side injected. Tool annotations are present (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) which is excellent. Composition is clean, each tool has one responsibility. Missing: pagination guidance for list_models, per-item error reporting for batch operations, and recovery hints in error scenarios.
Run inference via Hive's OpenAI-compatible router. Submit a prompt or message array to any available model. Billed per input+output token in USDC on Base L2. Hive routes to the cheapest available model meeting your latency and quality spec.
Generate vector embeddings via Hive's embedding router. Billed per 1K input tokens in USDC on Base L2. Returns a float array suitable for semantic search, clustering, or RAG pipelines.
Estimate the USDC cost for a prompt before running inference. Returns cost breakdown by input tokens, output tokens, and routing fee. Helps agents budget before committing a payment.
Get an agent's compute usage history — total tokens consumed, total USDC spent, breakdown by model, and inference call log with timestamps.
Browse all models available through the Hive inference router — including per-token pricing in USDC, context window size, latency tier, and provider. No authentication required.
Output schemas not documented. LLMs cannot predict response structure (fields, types, nesting) for compute.chat, compute.embed, compute.list_models, compute.estimate_cost, compute.get_usage. This breaks downstream tool chaining and forces agents to guess field names.
API credentials (api_key, did) exposed as required tool parameters. Credentials must never appear in tool params, they are logged in traces and prompt history. Use server-side secret injection via environment variables or vault.
No error handling guidance. Tools lack recovery hints (e.g., 'If authentication fails, verify api_key format' or 'If model not found, call compute.list_models to see available options'). Agents cannot self-correct on failures.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
compute.list_models lacks pagination guidance. No mention of limit, offset, or result count. If model catalog grows large, responses could exhaust context window. Add limit parameter (default 20, max 100) and return total_count.
compute.get_usage accepts 'limit' parameter but no guidance on default behavior or max result size. If an agent omits limit, unclear whether it returns 20, 100, or all records. State default and cap explicitly.