The change feed for always-on coding agents.
Always-on agents like ChatGPT dots and Claude Duet keep working while you are away, but they only learn what changed by checking. Chappe watches OpenAI, Anthropic, Google and 10 more vendors for them and pushes only the changes that touch your code, so your agent wakes up to open the fix.
Every check costs tokens, so they look every few hours and miss what happens in between.
“Something urgent can arrive between checks.”
The one change Chappe found that week for the demo repository. It also missed a Google change on Nov 2 while that source was down.
A real run on Oct 30–31, 2026, for a demo repository that calls OpenAI, Anthropic and Gemini, with a ChatGPT dot as the agent.
Times are minutes and seconds from the push to the pull request. Anthropic’s change was found the evening it was posted; the OpenAI and Google retirements come from their deprecation pages. The repository was connected after the changes were found, so Chappe replayed the stored events to it.
A sample, trimmed from the event our ChatGPT dot received for the Anthropic change. Field names may change.
{
"event_type": "track.created",
"vendor": "anthropic",
"title": "Fast mode for Claude Opus 4.8 will be removed",
"deadline": "2026-11-30",
"source_url": "https://platform.claude.com/docs/en/release-notes/overview",
"repo": "your-org/your-repo",
"relevance": "affected",
"recommended_action": "open_draft_pr",
"branch": "chappe/anthropic-sunset-fast-mode-for-claude-opus-4.8",
"what_to_do": "… migrate it in a draft pull request on branch
chappe/anthropic-sunset-fast-mode-for-claude-opus-4.8, or update the
pull request you already opened for this track_id …"
}What changed, and by when. Taken from Anthropic’s own release notes, with the sentence and the link.
Why it is this repository’s problem. The agent can ask for the matched lines: catalog.py:13 sends Opus 4.8 with fast mode.
What to move to. The vendor’s migration step comes with it: Claude Opus 5.5 keeps fast mode.
Where to put the fix. A fixed branch per change, so later news updates the same pull request instead of opening another.
Vendors change their minds. An always-on agent stays subscribed, so it gets those too.
Chappe pushes the correction and the agent updates the same pull request, a minute later in testing.
Your agent tells you both dates and leaves the code alone until one is confirmed.
The pull request is closed with a note, 65 seconds after the retraction in testing.
Tested end to end with a ChatGPT dot and test events on Oct 30–31, 2026.
The polling side is measured: an open-source always-on agent (OpenDots) searching the web with Parallel on GPT-6.1 Sol, 11 checks against the same repository on Nov 2, 2026.
Checks the vendors itself on a timer, for each repository.
Every vendor we watch, pushed to your agent only when a change touches your code.
Chappe reads most vendor pages every two hours, once for everyone. A polling agent that checks less often cuts its bill, not the misses below, and every change reaches it later.
Our demo repository, a small CLI that calls OpenAI, Anthropic and Gemini. Both sides got the same scan of it: SDK versions, model IDs and methods, with file and line.
OpenDots’ Dot agent, reproduced from its source (commit 8091324): its prompt, its Parallel search and page-reading tools, and its limit of five model turns per check. GPT-6.1 Sol. Run Nov 2–3, 2026.
“Which announced vendor changes affect this repository and still need action?” Four are true: Anthropic removes fast mode on Claude Opus 4.8, OpenAI removes gpt-5-mini, Google shuts down gemini-3.1-flash-lite, and Google deprecates gemini-2.5-flash (Nov 2).
“What did OpenAI, Anthropic or Google announce on this day that affects this repository?” for each day from Oct 27 to Nov 2. Two days have a true answer: Oct 30 and Nov 2.
Within its limits the agent answered 5 of 11 checks. Given one extra turn to answer, it found at best 3 of the 4 current changes, once proposing a replacement that retires the same day, and got 6 of 7 days right, missing Nov 2. Chappe on the same tasks: 3 of 4 and 6 of 7, missing the same Nov 2 change while its Google source was down.
Both at list price: the model at OpenRouter’s $2 per million input and $10 per million output tokens, Parallel at $1–5 per 1,000 searches and $1 per 1,000 pages read. Only the agent’s own turns count, not the extra answer turn we added for scoring. A check took about a minute.
The checks ran after the fact, which favours the always-on agent. Chappe also missed Google’s Nov 2 change until its source was fixed. Pricing is planned and may change.
Add Chappe to your ChatGPT dot, a Claude Code routine or Codex, or give it a webhook.
It records SDK versions, model IDs, endpoints and methods, with file and line.
Only the ones that touch your code. Everything else stays quiet.
Your source code never leaves your repository. Chappe sees identifiers and file locations only. The edit is made by your own agent, in your own repository, as a draft for you to review. How we handle your data
We are onboarding the first teams building on OpenAI, Anthropic and Gemini. Sign in with GitHub or Google and your API key is ready on the spot.
Get early access