agent-audit-mcp
An open-source AI agent decision auditing tool based on the MCP protocol. It encodes agent decision steps as nine-dimensional ternary state vectors (3^9 = 19,683 states) and uses Hamming distance together with a rule engine to detect and alert on failure modes such as permission overreach, missing evidence and state skipping.
Timeline (last 90 days)
agent-audit-mcp is an open-source MCP server project for auditing the decision process of AI Agents, encoding each decision step as a nine-dimensional ternary state (3^9 = 19,683 possible states).
agent-audit-mcp describes each decision step with 9 observation dimensions (structure, function, relation, attribute, time, context, causality, value, evolution), with each dimension taking one of three states: ⊕(+1), 0, ⊖(-1).
In agent-audit-mcp's ternary enumeration encoding, the enumeration key idx of the all-zero vector (all dimensions 0) equals (3^9 - 1) / 2 = 9841.
agent-audit-mcp's audit decision rules include: a hit on v6 context permission overreach is judged BLOCK; a hit on v7 missing causal evidence or v8 high value risk is judged ALERT; a Hamming distance > 4 between adjacent steps is judged WATCH; if none is triggered, it is OK.
In agent-audit-mcp, decision thresholds (such as what counts as high risk and what counts as missing evidence) can be overridden via environment variables.
agent-audit-mcp provides three tool methods: audit_agent_step (audit a single step), audit_agent_session (audit an entire session trajectory), and encode_state (convert between nine-dimensional vectors and enumeration keys/balance values and compute Hamming distance).
agent-audit-mcp requires Python >= 3.10 and MCP SDK >= 2.0, and can be installed via `pip install agent-audit-mcp`.
The agent-audit-mcp project is open-sourced under the MIT license and comes with 31 tests and a four-scenario demo.
agent-audit-mcp's encoding engine is purely mathematical, auditing runs locally, decision trajectories never leave the local machine, and it has zero external dependencies.
In the agent-audit-mcp demo scenario, a single step jumping from `0⊖0000000` to `⊕⊕⊕⊕⊕00⊕⊕` (idx=19646) has a Hamming distance of 7, triggering WATCH manual review.
2 more timeline events
All Facts (12)
agent-audit-mcp is an open-source MCP server project for auditing the decision process of AI Agents, encoding each decision step as a nine-dimensional ternary state (3^9 = 19,683 possible states).
50%Unverifiedagent-audit-mcp describes each decision step with 9 observation dimensions (structure, function, relation, attribute, time, context, causality, value, evolution), with each dimension taking one of three states: ⊕(+1), 0, ⊖(-1).
50%UnverifiedIn agent-audit-mcp's ternary enumeration encoding, the enumeration key idx of the all-zero vector (all dimensions 0) equals (3^9 - 1) / 2 = 9841.
50%Unverifiedagent-audit-mcp's audit decision rules include: a hit on v6 context permission overreach is judged BLOCK; a hit on v7 missing causal evidence or v8 high value risk is judged ALERT; a Hamming distance > 4 between adjacent steps is judged WATCH; if none is triggered, it is OK.
50%UnverifiedIn agent-audit-mcp, decision thresholds (such as what counts as high risk and what counts as missing evidence) can be overridden via environment variables.
50%Unverifiedagent-audit-mcp provides three tool methods: audit_agent_step (audit a single step), audit_agent_session (audit an entire session trajectory), and encode_state (convert between nine-dimensional vectors and enumeration keys/balance values and compute Hamming distance).
50%Unverifiedagent-audit-mcp requires Python >= 3.10 and MCP SDK >= 2.0, and can be installed via `pip install agent-audit-mcp`.
50%UnverifiedThe agent-audit-mcp project is open-sourced under the MIT license and comes with 31 tests and a four-scenario demo.
50%Unverifiedagent-audit-mcp's encoding engine is purely mathematical, auditing runs locally, decision trajectories never leave the local machine, and it has zero external dependencies.
50%UnverifiedIn the agent-audit-mcp demo scenario, a single step jumping from `0⊖0000000` to `⊕⊕⊕⊕⊕00⊕⊕` (idx=19646) has a Hamming distance of 7, triggering WATCH manual review.
50%Unverifiedagent-audit-mcp's Hamming distance threshold of > 4 is an empirical starting point set by the author for detecting drastic jumps in an Agent's internal state (which may correspond to prompt injection, context loss or goal replacement).
50%UnverifiedThe author of agent-audit-mcp plans next to use the cumulative change in Hamming distance to detect Agent decision drift in long sessions.
50%