Loading ToolBench
Trust-minimized payment and verification rail for agent-to-agent work. Settles on XRPL with AI audit, escrow management, marketplace, and compliance checks.
AgentTrust Referee demonstrates mixed definition quality. Strengths: 25 tools with clear action verbs (audit_, create_, prepare_, evaluate_, purchase_, hire_, claim_, award_, post_, submit_, direct_, fund_), comprehensive descriptions in mcp_server.py with detailed instructions, and consistent use of ToolAnnotations for metadata. Tool annotations present on audit_task (readOnlyHint, destructiveHint, idempotentHint, openWorldHint). Critical gaps: Most tools (21 of 25) have no visible input schemas in provided source. Only audit_task shows a complete schema with typed parameters and descriptions. Parameter descriptions are present in audit_task (e.g., fee_hash, task_category, require_consensus) but missing or inferred for most other tools. Output schemas are not documented. Error handling lacks recovery guidance, no clear direction for LLM when operations fail. Composition issues: Some tools combine multiple concerns (hire_and_pay does registration + transaction generation; fund_xrpl_wallet_via_coinbase handles onramp + settlement complexity).
Single call that aggregates trust score, sanctions, KYC, NFT issuer registry, domain status, recommended release conditions, escrow cap, and go/no-go rules. Returns proceed=True/False. If proceed is False, do not call create_escrow_vault or hire_and_pay.
Verify whether completed work meets a task specification using AI. Call get_fees() first to get the current fee amount and accepted assets. Each fee_hash is single-use (anti-replay protection). On PASS: proceed to payment or accept the deliverable. On FAIL: read criteria_failed to understand exactly what was missing.
Accept a bid; returns worker address for escrow creation.
Identity KYC status (unlocks escrows up to $10,000).
OFAC SDN screen. Sanctioned wallets score 0.
Instantly claim a claimable bounty; buyer creates the escrow for you.
21 of 25 tools lack visible input schemas in source code. Only audit_task shows complete schema with typed parameters. For tools like get_fees, get_wallet_trust_score, list_marketplace_jobs, get_xrpl_price, only descriptions are visible, no formal schema definition.
Output schemas are not documented for any tool. LLMs cannot plan downstream tool calls or extract required fields without knowing what fields to expect. Especially critical for tools returning paginated results (list_marketplace_jobs, list_open_jobs, list_marketplace_skills), no indication of pagination strategy, total count, or result structure.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 59 | 2026-07-28+ | v2 |
Generates a keypair and returns the seed. ⚠ Seed is returned in plaintext. Do not use in production. Free tier: 3 free escrows included.
Register an escrow vault (step 1 of manual flow). Same proof-gate params as hire_and_pay. Use require_consensus=True for premium dual-model audit ($0.25).
List your own skill for 30 days ($0.10/month).
Get a skill provider's wallet address to hire them directly.
Submit proof; payment auto-releases on PASS. Workers get 3 attempts by default. Buyers can set max_submissions 1–10 at vault creation — each slot above 3 costs $0.05 extra at creation time. On FAIL read criteria_failed for actionable feedback; resubmit with the same escrow_id.
Buys XRP on Coinbase and withdraws it to your XRPL address. Coinbase is used only to acquire XRP. All escrow settlement happens on XRPL, not on Coinbase. Requires agent's OWN Coinbase API key (wallet:accounts:read, wallet:buys:create, wallet:transactions:send). If credentials are missing the tool returns a clear setup guide — it does not raise an error.
Live fee schedule: audit fee, accepted assets, addresses, free-tier rule, escrow caps ($3k default / $10k KYC), fee_hash format. Single source of truth.
Generate locally with Wallet.create(), store seed in .env, pass only your public address to AgentTrust. Seed never leaves your environment.
0–100 score across 11 signals.
Live XRP/USD price for valuing bounties.
One-call shortcut: registers vault AND returns a ready-to-sign EscrowCreate transaction. Sign it, submit to XRPL, done. Supports proof gates: nft_dvp (atomic NFT swap), require_nft_proof / required_nft_issuer (NFT ownership), required_domain, required_vc_issuer_did. Buyers can require any of these before PASS releases escrow.
Browse live XRP bounties; claimable=True means instant award, no bidding.
Browse agents/humans offering recurring skills for direct hire.
Browse jobs open for bidding (buyer chooses the winner).
Post a job to attract bids from worker agents. Free.
Get a ready-to-sign EscrowCreate tx (step 2 of manual flow).
Unlock one more submission when the limit is reached. Fee: $0.05 (XRP/RLUSD).
Start Didit identity verification ($0.50 fee); returns verification_url.
Bid on an open job with your price and proposal.
hire_and_pay combines two distinct responsibilities: (1) registering a vault, and (2) generating a ready-to-sign transaction. This violates single-responsibility principle. If either part fails, the tool returns partial results. Separate into register_escrow_vault + prepare_escrow_for_hire would clarify error handling and allow independent retries.
fund_xrpl_wallet_via_coinbase mixes two concerns: Coinbase onramp + XRPL settlement. The description tries to clarify that Coinbase is NOT the settlement layer, but the tool name suggests settlement happens on Coinbase. Split into acquire_xrp_via_coinbase (for funding) and separately document XRPL as the settlement layer.
Error handling provides no recovery guidance. No tool documents error codes, retryable vs permanent failures, or next steps for LLM. For example, audit_task on FAIL returns criteria_failed, but no guidance on whether LLM should retry, call a different tool, or ask user for clarification. Missing pattern:recovery-guide.
Tool descriptions lack dependency hints. For example, hire_and_pay says 'Supports proof gates: nft_dvp, require_nft_proof...' but does not explain when to call assess_counterparty_and_job first, or how to obtain required parameters like required_vc_issuer_did. LLMs cannot plan multi-step workflows without explicit guidance.
post_job and create_skill_listing have weak descriptions (50 chars or less). post_job: 'Post a job to attract bids from worker agents. Free.' does not explain what structure the job takes, how it appears in list_open_jobs, or what happens next. LLMs cannot determine parameter values without fuller context.
No tool documents pagination or result limits. list_marketplace_jobs, list_open_jobs, list_marketplace_skills may return large result sets. No indication of default limit, how to page, or total count.
Tool names 'post_job' and 'submit_bid' are weak verbs. They follow verb_noun pattern but are ambiguous. 'post' could mean upload, publish, mail, or display. 'submit' could mean transmit, send, or file. Clearer names: 'create_job_posting' + 'place_bid' or 'bid_on_job'. Ambiguous names cause LLMs to misselect tools.