Sutriva gives coding agents temporal memory -- a persistent, queryable record of what happened across a debugging session, live or recorded. MCP server implementation.
Sutriva demonstrates strong naming conventions (all tools start with action verbs: inspect_, get_, search_, analyze_, compare_) and comprehensive descriptions (average ~180 chars, well above the 34-char minimum). All 10 tools have clear, LLM-optimized descriptions explaining WHAT, WHEN, and WHY. Input schemas are present with typed parameters and descriptions. However, output schemas are not explicitly documented in the source, the tool descriptions reference what is returned (e.g., 'Returns session metadata and event count', 'Returns the compact temporal event timeline') but formal JSON Schema output definitions are not visible. Error handling guidance is minimal; tools do not describe recovery paths or categorize errors. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present despite all tools being read-only operations.
Dense, on-demand analysis over a narrow time range: densely re-samples frames in [startSeconds, endSeconds] and asks the vision provider a specific question. Use this when get_timeline's coarse sampling isn't enough detail for a specific moment -- it is more expensive than get_timeline/get_frame, so scope the range as narrowly as possible.
Compares two sessions side-by-side: returns a diff of events, highlighting what changed between them. Useful for understanding the impact of a code change or a user action across multiple runs.
Returns the current state of a live session: what is the browser showing, what network requests are in flight, what console messages have appeared, and what terminal commands have been run. Only populated while a session is live (see inspect_video's liveSession field). Use this to understand the present moment during an active debugging session.
Temporal rewind: returns evidence within a time window around a timestamp, so Claude can answer "what happened immediately before/after this?" Each evidence item has a confidence and a source reference (a frame artifact or a transcript segment) -- treat confidence as observed/likely/possible, not as proof of causality.
Retrieves a single targeted video frame near a given timestamp as an image, so Claude can visually inspect a moment identified in the timeline. Prefer this over inspect_video for follow-up visual questions -- it does not re-analyze the whole video.
Output schemas not documented. Tool descriptions mention what is returned (e.g., 'Returns session metadata and event count') but formal JSON Schema output definitions are not visible in source. LLMs cannot plan downstream tool calls or extract specific fields without knowing the response structure.
No tool annotations present. All 10 tools are read-only operations but lack readOnlyHint annotation. This prevents clients from understanding operation safety and optimizing caching/retry strategies.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Returns the compact temporal event timeline for a session created by inspect_video. Each event has a timestamp, type, description, and confidence. Use afterSeconds/beforeSeconds/limit to bound the result instead of requesting the whole session at once.
Returns the full transcript (if audio was present) for a session, with timestamps and speaker labels if available. Useful for searching dialogue or understanding what was said during a moment.
Snapshot of the current environment: git status, uncommitted changes, working tree diff, and available capabilities (browser, network, console, terminal). Use this to understand the codebase context before diving into a session.
Ingests a local MP4 (or other ffmpeg-readable video) into a Sutriva session: extracts metadata, samples frames, runs vision analysis, and builds a timeline. Returns session metadata and an event count -- call get_timeline next to see what happened, rather than requesting raw video data. Re-inspecting an unchanged file (by content hash) reuses the existing session instead of re-analyzing.
Full-text search over a session's timeline descriptions. Use this to jump straight to a moment (e.g. "error", "checkout", "500") instead of scanning the whole get_timeline output.
Error handling lacks recovery guidance. Tools do not describe what errors are possible, whether they are retryable, or what the LLM should do next. E.g., 'sessionId not found' should suggest 'Call inspect_video() first to create a session.'
Parameter constraints not fully specified. 'maxFrames' and 'intervalSeconds' in inspect_video lack min/max bounds. 'limit' in get_timeline lacks range. Unbounded numeric parameters let LLMs pass absurd values.
No pagination guidance for list-like tools. get_timeline accepts 'limit' but does not document max result size, whether results are sorted, or how to fetch the next page. Large sessions could return unbounded results.