A FastAPI-based gateway for MCP delegation that validates delegation proofs and authorizes actions using cryptographic credentials and privilege attenuation
Server has 9 tools with complete input schemas and descriptions, but quality is uneven. Tool names follow verb_noun convention (authorize_action, issue_credential, build_proof, inspect_chain, verify_receipt, list_credentials, read_s3_file, write_s3_file, list_s3_files). Descriptions are present (avg ~150 chars) but lack actionable context for LLM selection. Parameters have types and descriptions, but lack constraints (enums, ranges, patterns). No output schemas documented. Error handling is absent, no recovery guidance, no categorization. Security concerns: proof_path and credential_store expose file system paths; no validation against path traversal. Composition is reasonable (single responsibility per tool), but S3 tools lack pagination support despite being list/read operations. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk labels present but not integrated into schema.
Validate a delegation proof and authorize an action using AVCAD gateway. Checks cryptographic signatures, credential chain validity, time bounds, privilege attenuation, and action authorization. Returns ALLOW/DENY decision with a signed tamper-evident receipt.
Build a DelegationProof from a leaf credential. Traverses the credential chain from leaf to root following delegated_from links, then signs the proof with the holder's private key.
Inspect a delegation proof's credential chain. Shows each credential in the chain with its issuer, holder, permissions, expiration, and delegation links. Useful for debugging authorization failures.
Issue a new DelegationCredential. Creates a cryptographically signed credential granting specific permissions from an issuer to a holder. For child credentials, specify delegated_from to link to the parent. Permissions must be attenuated (subset) when delegating.
List available credentials in the credential store. Shows credential ID, issuer, holder, service, and expiration for each.
No output schemas documented. LLMs cannot infer what fields to expect from tool responses, forcing unstructured parsing and increasing hallucination risk.
Parameters lack constraints (enums, ranges, patterns). 'service' accepts free-form strings (should be enum: s3, postgres, slack). 'actions' and 'resources' accept unbounded arrays. LLMs will hallucinate invalid values.
File path parameters (proof_path, leaf_credential_path, credential_store) expose file system paths without validation. No protection against path traversal attacks (../../../etc/passwd). LLMs can be tricked via prompt injection.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
List files in an S3/MinIO bucket, optionally filtered by prefix.
Read a file from S3/MinIO. Returns the file content as text.
Verify a signed decision receipt against the gateway's public key. Confirms the receipt has not been tampered with since it was issued. Returns validity status and receipt details.
Write content to an S3/MinIO file.
No error handling guidance. Tools return success/failure but do not tell LLM what to do next (retry? ask user? fatal?). No recovery paths for common failures (credential expired, proof invalid, file not found).
list_s3_files lacks pagination (no limit, offset, or cursor parameters). Returning unbounded file lists will exhaust context window and degrade LLM reasoning.