Execution provenance for AI agents and software
Open-source TypeScript libraries that record the steps, state and decisions behind a result as connected notes, and keep the gaps visible.
Built in the pattern, with recorders on, each note is connected as it is written.
One question, and you reread them all.
So stitch in code: bind its events into one action.
The flow runs in the pattern. Your trainer stays yours and leaves a receipt.
On-prem devices: join what shares an ID, and give every other gap its reason.
Code threads through the notes where they lie. A note it can't reach keeps the reason.
Two tracks, one kind of record.
Only the notes behind it light up.
A labelled gap beats a confident guess.
“No warranty applied.”
“Denied: 45 days¹ is past the 30-day policy². Warranty: unknown³, no order ID to connect it.”
Flow for steps. Why for causes. Story for people.
This page does the same. Try the switch above.
Stitched by default, or in code with a reason for every gap.
And a side effect: an audit trail for security monitoring.
Recording does not need an LLM. Enable the evidence you need before the run; missing evidence stays unknown.
footprintjs is a family of open source TypeScript libraries for execution provenance. They record the steps, state and decisions behind a result as connected notes, so people and coding agents can walk back from it, replay it and audit it.
Run the work and keep the record.
Follow a recorded application decision
Inspect tracked reads, writes and explicit decisions behind an outcome, where the required evidence was captured.
Investigate a supported saved run
foottraceQuery execution records, follow supported dependencies and inspect reported gaps without running an AI model.
Connected as the work happens.
The run keeps its own edges.
Investigate an agent’s answer
Inspect the recorded context, model calls, tools and decisions. Test suspected causes rather than treating an explanation as proof.
Compose reusable platform capabilities
footprint-blueprintA modular platform framework with focused import paths and shared execution evidence.
Configure commerce for your shop
smallbusiness-footprintReusable commerce workflows built on Blueprint, configurable for different shop types.
Make a film with a making-of record
footprint-storyreelBuild teaching films from a storyboard and recipe, with a seekable timeline, captions and a recorded production workflow.
For what you cannot rebuild. Every gap keeps its reason.
You can't rebuild React, so bind its events.
Make applications agent-operable
hcifootprintGive agents a typed map of the app's real capabilities while preserving permissions, confirmation gates, and the signed-in user's control.
Preserve recorded exploration paths
Keep supported human and agent branches of an analysis, compare paths and replay retained findings. Recording settings and declared statistical obligations still matter.
Hybrid: the flow is in the pattern, the trainer stays yours.
Keep the evidence behind a model release
FootPrint coordinates the workflow and your trainer keeps its code. Check matching evaluation evidence, then reopen the release decision later. No npm release yet.
The same idea at the collector, for on-prem devices.
Keep the work behind a data result
footprintjs-datapipelineCollection and derivation workflows built on FootPrint, with traceable execution evidence.
Whichever way it was stitched.
Read the record.
Inspect a recorded run
footprint-explainable-uiExplore recorded steps, state and supported variable traces. Choose a recorded subflow invocation where the record format supports it.
Inspect context and tools
agentfootprint-lensCompare recorded messages, prompts, tool activity, decisions and cost while investigating an agent run.
Follow the recorded journey
agentthinkinguiPresent a supported agent record as a story a person can replay and inspect. The story explains captured evidence, not a model’s hidden reasoning.
Related tools from the same project.
One story, several ways to read it
footprint-storydeckTurn structured content into an article, a scrolling story or a presentation. FootPrint integration is optional.
Spot conflicting claims in your context
contextfootprintCompare supplied assertions by identity and find conflicts. It does not decide which one is true.
Use FootPrint in an ordinary application. Your coding agent can inspect its recorded run without putting an AI model inside the application.
Use named stages, tracked state and explicit decisions.
Run the application with the required recorders enabled.
Follow the wrong result back through its recorded dependencies.
Add a regression test, repair the code and record another run.
Share the relevant trace slice rather than an entire log dump. The context needed depends on the workflow and tooling; no measured token or speed saving is claimed.
footprintjs is a family of open source TypeScript libraries for execution provenance. Recorders capture the steps, tracked state and decisions behind a result as connected notes, so people and coding agents can walk back from it, replay it and audit it.
It is a record of how a result was produced: which steps ran, what they read and wrote, and which decisions they took, joined by the edges between them. Logs give you facts. Provenance also gives you the links between them.
Logs record messages, and traces record order. A footprintjs record also keeps the decisions and the data behind them, so you can follow one result back through its dependencies instead of searching everything.
Build the agent with AgentFootprint, which records its context, tool calls, model calls and decisions as it runs. Start at the answer and follow the recorded evidence backward, or inspect the run in Why Lens. An LLM can explain those findings, but it does not invent the history.
No. Software built in the pattern is stitched by default. For what you cannot rebuild, you stitch in code: HACI Footprint and VizFootprint bind a framework's events into actions, Datapipeline does the same at the collector level, and footprintjs-ml coordinates workflows around your existing trainer.
It stays a labelled gap with a reason, such as not-collected or no-match. When backward evidence is incomplete, Foottrace withholds the root cause and reports its gaps instead of guessing.
No. An ordinary application can keep an execution record without an LLM. You enable the recorders you need before the run.
Yes, as a side effect. The same record that lets you walk back and replay gives security monitoring a trail to read. It is evidence of what was observed. It does not by itself prove correctness, compliance or the cause of an outside event.