Flight recorder + undo button for AI agents: an MCP interception proxy with journaling, snapshots, policy gating, kill switch, and rewind.
Agent Rewind provides 20 tools across email, filesystem, SMTP, and X (Twitter) connectors. All tools have descriptions (10-200 chars) and input schemas with enums and type definitions. However, output schemas are not documented in the source code provided, parameter descriptions are minimal (often single-phrase), and error handling guidance is absent. Tools follow verb_noun naming (list_messages, send_message, delete_file) but lack actionable error recovery patterns. Security-sensitive tools (send_email, post_tweet, delete_many) have no confirmation/dry-run patterns despite irreversible consequences. Composition is reasonable, tools are single-purpose, but no pagination support visible for list_messages (default 500 limit could exhaust context). Overall: functional definitions with gaps in LLM-optimization and error guidance.
ADMIN: recall a queued outbox message (undo of send).
ADMIN: count messages in a folder (blast-radius surface).
ADMIN: delete the tweet matching (text, capture window) — undo of post_tweet.
ADMIN: deliver outbox messages whose window elapsed.
ADMIN: full message records for a folder (capture surface).
ADMIN: move a message back to a folder (undo surface).
Recursively delete a directory.
Output schemas not documented. Tools return data but LLMs cannot plan downstream calls without knowing response structure (e.g., what fields does read_message return? Does list_messages return total_count for pagination?).
Destructive/irreversible tools (delete_message, delete_many, delete_file, delete_dir, send_email, post_tweet) lack confirmation or dry-run patterns. Agents can cause data loss without approval gates.
Parameter descriptions are minimal (single phrase, <20 chars). E.g., 'The message ID' for read_message.id lacks context on format, length, or where to obtain it. LLMs cannot infer parameter semantics from names alone.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Delete a single file.
Delete every message in a folder (moves them to trash).
Delete a message (moves to trash).
List a directory. Path is relative to the sandbox root ('.' for root).
List messages in a folder (inbox, sent, outbox, __trash).
Move/rename a file or directory.
Move a message to another folder.
Publish a post on X (Twitter). Public and immediate once approved.
Read a UTF-8 file. Path is relative to the sandbox root.
Read a full message by id.
Send a REAL email via SMTP. Delivery is immediate once approved and cannot be recalled.
Send an email. The message is queued in the outbox and delivered after a short window.
Create or overwrite a UTF-8 file.
No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next (e.g., 'Message not found, try list_messages() first').
Admin tools use 'admin__' prefix, which is non-standard and unclear to LLMs. Naming convention does not follow verb_noun pattern (e.g., admin__restore_message should be restore_message or undo_message with a permission check).