MCP Server for AI Network - provides tools for interacting with AI Network blockchain database and DAG operations
The server defines 5 tools with basic schemas and descriptions, but critical issues limit production readiness. Tool descriptions are present but generic; most lack depth on WHEN to use them vs. alternatives. Input schemas use Zod validation and are partially visible, but output schemas are undocumented. Parameter descriptions exist but are minimal (many under 50 chars). Error handling returns structured responses but lacks recovery guidance. No tool annotations (readOnlyHint/destructiveHint/idempotentHint). The 'publish-instruction' tool description contains sample JSON that LLMs may reuse literally. Overall, the definitions meet baseline requirements but lack the specificity and completeness expected of production tools.
Add content to the AI Network DAG via gRPC. - You can provide a simple string in 'message' to store as content, - When you want to upload a file, you should use 'data' as the file content (string(base64)) and 'message' as a file name. - The resulting 'cid' (content identifier) will be generated and returned in the result.
Get content from the AI Network DAG by CID.
Get the value of a path from the database on AI Network
Publish an instruction to the AI Network DAG via gRPC. Sample instruction to run "mnist-train" container would be { topic: 'mnist-train' instruction: { command: "train", container_id: "mnist-train", container_args: { worker_pk: "workerpk1", data: "cid1" }}.
Set a value to a specific path in the AI Network database
No output schemas documented. Callers cannot predict what fields to expect in responses, forcing LLMs to guess at data extraction and downstream tool composition. All tools return {content, isError} but field structure, types, and nested object schemas are undocumented.
publish-instruction description includes sample JSON literal (topic: 'mnist-train', command: 'train'). LLMs frequently copy example values verbatim into real calls, bypassing actual intent. Replace example payloads with formal constraints (enum, pattern, schema reference).
Parameter descriptions are minimal (10 - 30 chars). E.g., 'Database path to retrieve value from' for get-value lacks expected format, constraints, or error cases. Descriptions should clarify: what format? Max length? What happens on path not found? This forces LLMs to infer behavior.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 52 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 44 | - | v1 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Agents cannot distinguish safe reads (get-value, get-content) from destructive writes (set-value, publish-instruction, add-content) without LLM reasoning. Annotations enable safe retries and dry-run logic.
Error responses return only a text message (error.message wrapped in {content, isError: true}). No recovery guidance, categorization (retryable vs. fatal), or actionable next steps. LLMs receive raw errors and must guess what to do.
add-content accepts optional parameters (cid, message, data, children) with no documentation of valid combinations. Are all optional, or is at least one required? Must exactly one of (message, data) be provided? Undocumented dependencies invite misuse.
publish-instruction expects 'instruction' as {type: 'any'}. No schema, no valid fields, no constraints. LLMs have no guidance on what instruction objects look like or which fields are required. Publish will fail unpredictably if malformed.
add-content description states file content should be 'base64-encoded string', but {type: 'any'} does not enforce this. LLMs may pass raw strings, binary, or other types. No validation message guides correction if encoding is wrong.