Chrome extension: Nea Agora Recorder (developer mode)
A local-only Chrome extension that records your LLM / agent sessions on the web and exports agent_service_record JSON artifacts.
It does not send data to any server.
It does not compute scores or rankings.
It only turns raw interaction history into a structured record you control.
Patent Pending
Certain concepts and methods implemented or described in this repository are covered by one or more provisional patent applications filed by the authors.
Open-source usage is permitted under the MIT license.
This notice is provided for disclosure purposes only and does not imply certification, endorsement, or warranty.
This repository accompanies the article “Inside a Real ChatGPT Service Record”
It documents real service records captured while dogfooding the recorder during coding, creative, and research sessions.
The goal is not evaluation or scoring, but observability: what actually happened during human-AI interaction.
Today, LLMs and agents operate like black boxes.
- You can’t see how often they fail.
- You can’t see when a human had to step in.
- You can’t compare “how well” different agents behave over time.
The Nea Agora Recorder is a small missing piece:
A behavioral/service record generator that sits between raw chats and any future trust, ranking, or enforcement layer.
It focuses on observed outcomes and human interventions, not on opaque “trust scores”.
All data stays local to your browser.
The extension:
- Observes your chat sessions on
https://chatgpt.com/(for now). - Groups events into sessions.
- Tracks:
outcomeper session:"success""abandoned""escalated_to_human"
neededHumanOverride(boolean)- basic timing data (
startedAt,endedAt,approxDurationMs)
- Exports everything into a single
agent_service_recordJSON file.
There is no backend:
- No sync
- No telemetry
- No remote analytics
Just a local artifact you can inspect, archive, or feed into your own tooling.
Right now this is a developer-only MVP.
-
Clone the repo:
git clone https://github.com/neaagora/nea-agora-recorder.git cd nea-agora-recorder -
Build or prepare the extension (if needed):
# if you have a build step npm install npm run build # after build, the extension lives in: # apps/chrome-extension/dist
-
Load it in Chrome:
- Open
chrome://extensions - Enable Developer mode
- Click Load unpacked
- Select
apps/chrome-extension/dist(the folder withmanifest.json)
- Open
You should now see Nea Agora Service Recorder in your extensions list.
-
Open
https://chatgpt.com/and use ChatGPT as usual. -
The extension's content script will observe your prompts and group them into sessions.
-
Use the extension UI (popup) to:
- Inspect recorded sessions.
- Adjust the outcome if needed:
"success""abandoned""escalated_to_human"
- Toggle
neededHumanOverridewhen you had to manually fix something.
-
Export a service record:
- Click Export in the popup.
- This downloads a JSON file with:
recordType: "agent_service_record"subjectinfo (which agent/surface)observerinfo (the recorder + environment)- an array of
sessionswith outcomes and flags.
See examples\service_record__chatgpt_minimal.json for a concrete example. These example service records are taken from real usage.
Some fields differ across versions (v0.4 vs v0.5), reflecting the evolution of the recorder during dogfooding.
Discrepancies are intentional and documented in the accompanying article.
The real sample in examples/ contains multiple sessions, including:
successabandonedescalated_to_humanwithneededHumanOverride: true
This is not a trust score. It is raw, structured history that other systems can interpret later.
This project is not:
- a trust or reputation system
- a ranking engine
- a centralized analytics platform
- a cloud service
It is only:
A tool to record what happened and when a human had to intervene.
Any trust scoring, aggregation, or enforcement should live in separate tools or services.
Early MVP:
- Chrome only
- ChatGPT-only (
https://chatgpt.com/*) - Developer-mode install
- UI is functional but minimal
- Data model may evolve
Breaking changes may happen as we learn.
- Better session visualization
- Import / merge multiple ServiceRecord files
- Optional CSV export
- Per-agent filters
- Safer defaults around what gets recorded
Issues and PRs are welcome. If you’re building:
- trust ranking
- enforcement
- service-record visualization
...and want to hook into this recorder, open an issue and describe your use case.
MIT. See LICENSE.