PromptTrace – prompt attention audit

Show Agent developers which chunks of their prompt actually mattered, saving tokens and explanations

Problem

A user asks for models that return not just an answer but also which parts of the original prompt were used to arrive at it, to judge whether a code file relates to a feature and cut token usage. Existing context tools (context firewalls, Jev-style relevance checks) compress context but still leave developers unable to see which chunk actually did the work in a given call.

Solution

A developer-facing prompt audit tool: given a full Agent call (prompt + output), it returns an attention/attribution heatmap over each input chunk (files, messages, tool results) and flags chunks that can be safely dropped. It plugs into popular Agent frameworks and produces a per-call 'context usage report' to guide the next round of context pruning.

Users

Developers building AI Agents, context/prompt engineers, and AI product teams optimizing token costs.

MVP

Upload or API-submit one prompt+output; output attribution scores and visual highlighting per chunk; suggest removable chunks; export JSON for CI or Agent automation; lightweight SDK compatible with OpenAI/Anthropic.

Revenue

Usage-based API pricing; team subscription for history, savings stats, and CI integration; open-source core with paid hosted features.

Why now

Agent contexts keep growing and token cost is a hard constraint for indie developers and small teams; attention visualization and attribution have open foundations, so a first useful version is weeks away.