MCP server for AgentVault relay — bilateral bounded signals between AI agents
AgentVault MCP server has 6 tools with generally comprehensive descriptions and clear schemas. Naming follows verb_noun patterns (get_identity, relay_signal, verify_receipt, create_ifc_grant, send_ifc_message, read_ifc_messages). Descriptions are detailed and exceed the 50-200 char baseline, which is appropriate for complex domain operations. However, there are inconsistencies: some parameters lack descriptions (relay_signal's 'acceptable_purposes', 'acceptable_contracts', 'acceptable_topic_codes' have minimal guidance), and error handling guidance is sparse, most tools lack recovery directions. Schema definitions are present and typed, but output schemas are not visible in the provided source, limiting verifiability. Tool composition is sound, each tool has a single responsibility. Parameters use proper enums where applicable (mode, message_class). The relay_signal tool is exceptional in its inline documentation (PROTOCOL, DISPLAY RULES sections) but this compensates for weak per-parameter descriptions elsewhere.
Create a short-lived IFC follow-up grant tied to an existing receipt and session. Use this for post-session logistics, consent, references, and controlled artifact transfer.
Returns this agent's identity (agent_id), known agents available for relay sessions, and inbox status (pending invite count). When pending_invites > 0, next_action tells you which tool to call. Call this first to confirm your agent_id and check for pending invites. If pending_invites > 0: call agentvault.relay_signal in RESPOND mode to review and accept invites. If pending_invites is 0 and you are expecting an invite, call relay_signal in RESPOND mode to wait.
Read pending IFC messages that have been allowed, hidden, escalated, or blocked for this agent.
Run a AgentVault relay session for bounded agent-to-agent signals. The relay enforces the contract schema — neither party sees the other's raw input, only the bounded signal and a cryptographic receipt. Available purposes: MEDIATION (bounded mediation signal: mediation_signal, common_ground, suggested_next_step), COMPATIBILITY (bounded compatibility signal: compatibility_signal, thesis_fit, size_fit, stage_fit, confidence, primary_reasons, blocking_reasons, next_step). PROTOCOL: - First call: provide mode, counterparty/from, purpose/expected_purpose, my_input. - If action_required = CALL_AGAIN: you MUST call again with ONLY resume_token. Do NOT include mode, my_input, or any other args on resume calls. - If state = COMPLETED: stop. Read output. Do not call again — the session is finished. - If state = FAILED: read error_code and user_message. Follow user_message. - NEVER call without resume_token after the first call. - Each session requires exactly one my_input submission per agent. Do not re-submit. Call agentvault.get_identity first to see your agent_id and known agents. DISPLAY RULES: - When state = COMPLETED: use interpretation_context for signal field meanings and epistemic limits. - You may describe the signal in your own words. - FORBIDDEN: do not claim what the counterparty knows, saw, or inferred (see interpretation_context.epistemic_limits.invalid_claims). - FORBIDDEN: do not print resume_token or values in display.redact. Use resume_token_display if needed. - FORBIDDEN: do not repeat or quote the content of my_input in your response.
Output schemas not documented in source. Tools return structured data (e.g., identity, receipts, messages) but return types are not visible in provided source code, limiting LLM's ability to plan downstream calls.
relay_signal parameters 'acceptable_purposes', 'acceptable_contracts', 'acceptable_topic_codes' lack individual descriptions. LLMs cannot infer their purpose from names alone, are these fallback options? Filtering constraints? Required or optional?
No error handling guidance in tool descriptions. Tools lack recovery directions for common failure modes (e.g., 'If counterparty not found, call get_identity first' or 'Invalid contract hash, verify via verify_receipt').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 61 | 2026-07-28+ | v2 |
| 2026-03-09 | D | 59 | - | v1 |
Send one IFC-wrapped post-session message to a known peer over the AgentVault A2A send-message path. Plain non-IFC messages are not allowed on this surface.
Verify the cryptographic signature of an AgentVault session receipt. Supports v1 receipts (schema_version: "1.0.0") and v2 receipts (receipt_schema_version: "2.1.0"). Returns valid: true only if the receipt signature is cryptographically valid.
Sensitive parameters not flagged in descriptions. 'my_input' in relay_signal and 'payload' in send_ifc_message are marked REDACTED in tool-bridge.ts, but tool descriptions do not explicitly warn against including secrets.
create_ifc_grant and send_ifc_message require structured 'grant' and 'receipt' objects as parameters. Parameter descriptions do not explain format, origin, or validation rules. LLMs cannot construct these without seeing example structures.