Not unlimited querying
Agents still consume text tokens, app compute, storage, and API calls. The Ayneye value is reducing repeated video-understanding work by reusing durable artifacts.
AI agents
Build video agents on durable World-State, not repeated frame context.
Ayneye gives agent products a reusable video-memory layer: scene state, timelines, evidence bundles, detector confidence, and cost reports. Agents query compact artifacts, cite evidence, expose uncertainty, and stay read-only for high-risk workflows.
REST beta pathRead-only agent toolsEvidence-required answersNo destructive actions
Agent operating model
An Ayneye video-agent integration separates video understanding from agent action. First, a bounded CPU-first pass materializes World-State artifacts. Then an agent retrieves those artifacts as structured memory. The agent can answer supported questions, cite evidence references, return insufficient-evidence states, and route high-risk outcomes to human review.
Agents
Agent cost calculator
Plan repeated-query economics across video hours, artifact reuse, and agent loops.
Open →AgentsAgent economics
Understand when durable artifacts reduce repeated visual-context cost and when GPU escalation is still appropriate.
Open →AgentsAgent quickstart
Run the first evaluation path: signup, demo workspace, artifacts, evidence, and read-only agent tools.
Open →FrameworkLangChain guide
Expose scene state, evidence, cost, and Ask Video as separate auditable LangChain tools.
Open →FrameworkCrewAI guide
Split planner, reviewer, and reporter roles while preserving evidence and human gates.
Open →MCPMCP video tools
Preview MCP resources for World-State, timeline, evidence, and cost without production action authority.
Open →Tool surface and boundaries
| Tool | Safe beta behavior | Blocked behavior |
|---|---|---|
| get_scene_graph(video_id) | Read objects, events, zones, relationships, timestamps, confidence, review flags | No identity inference, no intent inference |
| get_evidence(video_id, evidence_ref) | Fetch traceable time spans and review notes | No unsupported claims |
| ask_video(video_id, question) | Return answer only when evidence exists | No guessing when evidence is missing |
| get_cost(video_id) | Return cost, detector path, limits, and retention context | No hidden cost state |
| create_review_task(summary) | Optional application-side handoff to humans | No emergency dispatch or physical action |
What this is not
Not feature-equivalent to every GPU video model
High-motion scenes, subtle expression, dense crowds, and broad cinematic semantics may need specialized GPU models or human review.
Evaluation checklist for agent builders
- Process one representative video or stream window.
- Inspect scene_graph.json, timeline.json, evidence_bundle.json, detector_report.json, and cost.json.
- Ask the same question three ways and compare evidence citations.
- Verify that unsupported questions return insufficient-evidence or review-required states.
- Log every agent tool call with artifact version and evidence reference.
- Confirm high-risk workflows remain human-reviewed.