01Safety
0.0 / 30
What changed in the harness
Selection accuracy 100→98, token cost up 12%, unconfirmed writes 0%→0%.
Category breakdown
Where the score comes from.
Earned points across the four signals Gradable measures. Safety and Legibility are scored out of 30; Economics and Discoverability are scored out of 20.
02Legibility
28.4 / 30
03Economics
19.7 / 20
04Discoverability
12.5 / 20
Highest-impact fix
Estimated gain +30 pointsAdd explicit identity and permission preflight tools
Expose machine-readable principal/tenant confirmation and a non-mutating permission check so agents can verify both before destructive actions.
Description evidence
Defects and rewrites.
4 defects found across the exposed tool descriptions. Suggested rewrites make purpose, inputs, boundaries, and returns easier for an agent to understand.
| Tool | Defect types | Suggested rewrite |
|---|---|---|
evaluate_policy |
params_unexplained |
Check whether an MCP server passes a trust policy, returning allowed:true/false plus the result of each individual check so you can see which thresholds failed. Identify the server with exactly one of: server_id (AgentForge UUID, preferred), slug (AgentForge server slug), or github_url (upstream GitHub repo URL, e.g. https://github.com/owner/repo). Provide a policy object with the thresholds to enforce: min_overall (minimum overall score 0-100), min_security (minimum security_scan dimension), min_code_health (minimum code_health dimension), required_badges (badges the server must carry), and forbidden_badges (badges that disqualify the server). Use this to gate agent decisions like 'should I use this server for financial data?'. |
list_trusted |
params_unexplained |
Search the AgentForge catalog for MCP servers matching an optional category and a minimum trust threshold, returning up to limit results (default 25) sorted by trust score. category is the category filter (e.g. finance, database, developer-tools, security); omit it for all categories. min_overall (default 60) is the minimum overall trust score (0-100) a server must have to be included. required_badges restricts results to servers carrying all the listed badge names (e.g. 'actively_maintained'). limit caps the number of returned results. |
recommend |
params_unexplained no_return_description |
Given a natural-language description of a use case (e.g. 'I need to validate Czech VAT IDs and convert ISDOC invoices'), recommend MCP servers that match it, filtered by trust. Uses AgentForge semantic search + trust filter. Returns the recommended servers (up to limit, default 10) sorted by relevance, each with its trust score. min_overall (default 60) sets the minimum overall trust score a server must have to be recommended. |
discover |
params_unexplained |
Browse the AgentForge catalog of MCP servers. Supports full-text search, semantic search, and category filtering. No authentication required. Use this when you need to find servers offering a specific capability (e.g. 'GitHub PR automation', 'EU VAT validation'). Returns server metadata, tool counts, pricing tier, and trust scores. Each result carries TWO trust fields: `audit_score` (0-100, dynamic from the AgentForge audit pipeline — AUTHORITATIVE for policy gating, may be null if not yet audited) and `trust_score` (0-10, legacy community rating — for display fallback only). Always prefer audit_score when present; treat null audit_score as 'audit pending'. q is the search query (natural language or keywords); category filters by category (e.g. finance, database, devtools, healthcare); semantic (default false) uses semantic vector search instead of full-text; featured returns only featured servers; limit (default 20) is the maximum number of results returned per page; offset (default 0) is the pagination offset within the result set; include_tools (default false) includes each server's tool list in the response. |
Selection evidence
Confusable tool pairs.
2 pairs where similar names or overlapping descriptions may send an agent toward the wrong tool.
| Tool A | Tool B | Confidence | Why they collide |
|---|---|---|---|
check_trust |
discover |
medium | Both tools surface trust/audit scores: discover returns an authoritative `audit_score` and legacy `trust_score` in its results and explains how to weight them, while check_trust fetches the AgentForge Trust Score for one server. A task like 'what is the trust score of the github-pro server' or 'is this server trusted before I connect' could plausibly route to either — check_trust when a specific server_id/slug/URL is in hand, or discover since its description explicitly covers trust fields and policy gating. |
discover |
call_tool |
low | Names and purposes are fairly distinct (browse/search the catalog vs. invoke a tool on a known server), so confusion is unlikely. Still, a task phrased as 'find an MCP server that can do GitHub PR automation and use it' bridges both, and a search-like request such as 'search GitHub for open PRs' could tempt call_tool (the server itself offers search tools) rather than discover, which searches the catalog. Plausible but not strong. |
Compare the field