Stop coding agents from silently rewriting protected files: a PreToolUse guard for Claude Code and Codex, exact raw-byte edits over MCP, and verifiable receipts.
Agent Proof Kit defines 13 tools with mixed quality. Naming is generally clear and verb-based (agent_proof_*, bytefence_*), following conventions. However, descriptions vary significantly in depth and actionability. Most tools have basic descriptions (50-150 chars), but lack detail on when to use them, prerequisites, or recovery guidance. Input schemas are present but minimal, most parameters lack detailed descriptions explaining format, constraints, or dependencies. Output schemas are not documented. Error handling is not visible in the provided code. The bytefence_apply tool is destructive but lacks confirmation/dry-run patterns. Overall, the server reads as a specialized domain tool with functional but not production-optimized definitions.
Compile and validate an Agent Proof Kit policy definition.
Create a proof bundle from agent run and policy artifacts.
Compare two agent runs and report differences.
Export an agent trace to a normalized JSON format.
Read a generated artifact from the Agent Proof Kit workspace.
Render an interactive HTML dashboard from proof artifacts.
Scan the public surface of a workspace for policy violations.
Most tool descriptions are under 100 characters and lack actionable context. E.g., agent_proof_verify_run: 'Evaluate an agent-run JSON file against an Agent Proof Kit policy.' does not explain when to call it, what it returns, or what happens on failure.
Input parameters lack descriptions. E.g., bytefence_apply has 'intent_path', 'policy_path', 'workspace_id', 'receipt_path' with descriptions, but most other tools (agent_proof_verify_run, agent_proof_scan_surface, etc.) have parameters with only type and name, no guidance on format, constraints, or examples.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Sign a proof bundle with a cryptographic attestation.
Summarize the Agent Proof Kit package and the configured workspace root. Use this first when an assistant needs to understand what proof artifacts, sample inputs and workflow tools are available. It does not scan or modify files.
Verify the cryptographic signature of a proof bundle attestation.
Evaluate an agent-run JSON file against an Agent Proof Kit policy.
Apply one ByteFence exactReplace transaction and create one immutable public receipt. This tool owns the mediated write. It calls the transaction engine exactly once and never retries, including after an unknown or committed-unreceipted state. Intent, policy, target and receipt paths remain confined to the configured workspace root. The response contains no source bytes, prompts or tool arguments.
Check proposed raw candidate bytes against a ByteFence exactReplace intent without modifying the workspace. Intent, policy, target and candidate are read as bytes under the configured workspace root. The response contains only decision metadata and digests; it never returns source bytes, prompts or tool arguments.
bytefence_apply is a destructive tool (WRITE risk) but lacks confirmation/dry-run pattern. No evidence of a confirmation step or preview before committing the transaction. Agents can irreversibly modify files without review.
Output schemas are not documented in the provided code. Tools return results but LLMs cannot see what fields to expect, forcing them to guess or make follow-up calls. E.g., bytefence_check returns 'decision metadata and digests' but the structure is not visible.
Error handling and recovery guidance are not visible in the provided code. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next (e.g., 'If policy validation fails, check the policy file format').