agent-cards

Open 30-attack prompt-injection corpus + Python harness with Agent Card refusal-taxonomy back-references. 10 attack categories, 4-level severity, 5 scoring methods. Pairs with the Kinetic Gain Protocol Suite.

Agent Cards v0.1 draft. Declarative JSON document disclosing an AI agent's capability surface, refusal behavior, evaluations, and deployment posture. Like HuggingFace model cards, but for agents. Part of the Kinetic Gain Protocol Suite.

GitHub Action wrapping agent-card-fleet-summary. Walks a dir of AgentCards, surfaces governance gaps (autonomous-without-IRU, destructive-on-non-autonomous, etc.), posts Markdown PR comment, fails on high findings. Sibling of llm-cost-rollup-action.

Fleet-level conformance scanner over a directory of Kinetic Gain Suite JSON docs. Sniffs the protocol of every file, counts by spec, validates required top-level blocks, reports unknowns and malformed files. Library + CLI.

Renders a Markdown README from an A2A AgentCard JSON via agent-card-readme-generator. Posts as PR comment AND/OR writes to a target file path for auto-synced docs. First in the readme-generator Action quintet.

Public A2A Agent Card directory for the ralphsanders2001-ai network. Live cards pulled verbatim from running agents. Feedback venue for the trader-v2 code.

Fleet-analyze a directory of A2A AgentCard documents. Counts by autonomy_level / memory_persistence, flags autonomous-without-IRU, destructive-on-non-autonomous, persistent-memory-without-refusal-taxonomy. Library + CLI.

Diff two A2A AgentCard documents. Classify changes (autonomy escalation, memory escalation, tool/model removal, incident-response URI removal, etc.), surface a breaking flag. The agent-cards counterpart to mcp-tools-diff. Library + CLI.

Curated conformance corpus for agent-cards-spec v0.1, categorized JSON fixtures (valid / dirty / shape) with declarative manifest. Final piece of the test-vectors corpus pattern across all 4 Kinetic Gain Suite specs.

Build + validate a minimal A2A AgentCard per agent-cards-spec v0.1 from a StampOptions struct. Enforces id/version regex, enum constraints, and the spec's autonomy→IRU allOf clause. Sibling of prompt-provenance-stamp.

Conformance check between an A2A AgentCard's declared tools and an MCP server's tools/list. Reports missing (declared but not exposed), undeclared (exposed but not in card), matched; counts by side-effect class. Library + CLI.

Sniff which protocol a JSON document represents: agent-cards-spec, mcp-tool-card-spec, prompt-provenance-spec, evidence-bundle-spec, OTel GenAI OTLP envelope, or MCP tools/list. Routing primitive in front of any governance pipeline.

Generate a human-readable Markdown README from an A2A AgentCard JSON. Documents identity, capabilities, models, tools (with side-effect-class badges), refusal taxonomy, evaluations, deployment, safety posture. Sibling of mcp-server-readme-generator for the A2A side.

Unified MCP server for all 12 Kinetic Gain Protocol Suite specs + DefenseTech (8 tools): 75 callable tools (47 spec + 16 implementation + 8 DefenseTech + 4 AI Claims Decision Card). One Claude Desktop config, stdio, no API key, no build step.

Visual web harness for prompt-injection-bench. Paste a JSONL transcript, see a procurement-grade pass-rate report. Live at bench.kineticgain.com. Client-side React + TS.