Start with a recorded execution
English | 简体中文
Open the online demo to try the graph, Agent session and saved files without installing anything. The same site includes the full project overview, capture and deployment guides, analysis commands, integration references and release history.
For offline replay, open the demo library from the extracted archive:
./agentprov demo
Choose Open replay to explore a signed execution or Read guide for its illustrated walkthrough. Nine replays and eleven guides are included. No Go, Docker, agent account or API key is needed for replay.
Investigate, repair and compare
./agentprov demo investigation-workflow
A real DeepSeek session delegates code review, runs seven tests and encounters denied optional diagnostics. It then continues the same session, disables those diagnostics and successfully produces the report. Both signed runs load together.
In Compare, set A to Diagnostic repair: before and B to
Diagnostic repair: after. Read settings.json and the reports under
File evidence; Configuration shows Harness settings.
Open Related sessions to inspect delegation and continuation. Search the recorded error, inspect AppArmor/seccomp decisions, compare the saved files, then export selected evidence as a read-only page. Task, configuration and collection details remain available from the overview.
The CLI checks the evidence signature and imports into a temporary store. Replay works offline and never runs the recorded commands. Ctrl-C closes the server and removes its temporary data.
More examples
./agentprov demo --list
./agentprov demo deepseek-context
./agentprov demo multiagent-provenance
./agentprov demo k8s-cross-pod-a2a
./agentprov demo --no-browser
--no-browser prints a local URL. For a remote machine, forward that local port
to your browser. The demo/ directory contains the original bundles, public
keys, scripts and guides. Live capture requirements are listed in each guide.
LLM Judge and Jev are optional Python integrations. Opening their guides does not call a model. To try the LLM Judge integration with preset offline results:
AGENTPROV_BIN="$PWD/agentprov" python3 demo/llm-judge/judge.py run --offline
Jev requires Python 3.9+; live evaluation needs credentials and permission to send
selected evidence. Follow demo/jev-judge/README.md. Completed studies can be
reopened offline.
Language
Web pages follow the browser language on first visit, falling back to English.
The English / 中文 switch saves your choice. CLI output defaults to English;
commands, paths, IDs, JSON and original evidence keep their original text.
The existing --lang zh-CN option selects Chinese help and supported output,
and opens web pages in Chinese.
Platforms and verification
- Linux archives include the optional native sensor. Kernel capture needs the appropriate kernel, cgroup access and BPF/perf permissions.
- macOS supports application recording and replay. The CLI is not Apple Developer ID signed or notarized.
- On Windows, use the Linux archive inside WSL.
SHA256SUMSand.sha256files check download integrity. The example public keys verify evidence signatures. Collection status separately describes what the original recording captured.
Embedded guides and their images work offline. Following external links requires network access.