ChatGPT and Claude can review code you paste in; TinyAgent is the AI that opens GitHub and does the repo chores. It runs or checks GitHub Actions, creates repos and commits files, and tells you what changed.
Free to start. $8 in Wallet funds on us, no card needed.
Prompt
Open the Actions tab of my repo tinyfish-demo. Tell me the status of the latest CI run, its job, each step's result and how long it took. Read-only: do not re-run or change anything.
ChatGPT or Claude
Install the TinyFish plugin in ChatGPT or Claude. Public repos need no sign-in; for private repos or changes, sign in to GitHub once through TinyProfiles.
Example instruction
Use TinyFish to open the Actions tab of my repo acme/web. List the failed workflow runs from the last 7 days with the failing job and step, then re-run the latest failed run of the deploy workflow only. Do not change any files.
TinyAgent API
Call TinyAgent with each repository every morning and get open issues, stale pull requests and failed Actions back as JSON.
Triage a GitHub repo from one call:
from tinyfish import TinyFish, EventType, RunStatus
client = TinyFish()
with client.agent.stream(
url="https://github.com/tinyfish-io/tinyfish-cookbook",
goal="""Read-only triage of this repository. Return JSON with open issues and open pull requests (number, title, author, last_updated), and the 3 that most need a maintainer's attention with a one-sentence reason each. Do not comment, label or close anything.""",
) as stream:
for event in stream:
if event.type == EventType.COMPLETE and event.status == RunStatus.COMPLETED:
print("Result:", event.result_json)Runs are billed at $0.016 per agent step.Read the docs
These are the GitHub jobs people give TinyAgent most often. Public repos need no sign-in, and TinyAgent changes only what your prompt allows.
For small engineering teams
No release engineer to watch the builds? TinyAgent triggers the workflow you name, waits for it to finish and reads the result or the failure. In the run above it read every step of the latest CI run without changing anything.
Try
Run the test workflow on main in acme/web, wait for it to finish and tell me which job and step failed, if any.
You get: The run's status and duration, plus the failing job and step if there is one.
See TinyAgentFor founders and builders
Starting a project from a chat? TinyAgent creates the repository with the name, visibility and settings you give it, so your code has somewhere to go.
Try
Create a private repository called acme-landing with a README and a Node .gitignore.
You get: The repository link and the settings it used.
See TinyAgentFor maintainers
A version bump or a README fix should not need a local checkout. TinyAgent creates or edits the file and commits it to the branch you name. Check the Actions result in a separate run.
Try
In acme/web, set .nvmrc to 20 and commit it to the branch chore/node-20 with the message “Bump Node to 20”.
You get: The commit link and the files it changed.
See TinyAgentFor reviewers
Know where a pull request stands before you open it. TinyAgent reads its checks, and marks it ready for review or updates its description when your prompt allows it.
Try
Open pull request #42 in acme/web and tell me which checks failed. If all passed, mark it ready for review.
You get: Each check with its status, and any change it made to the pull request.
See TinyAgentChatGPT can read code through its GitHub connector. TinyAgent adds the clicking: it opens GitHub in a real browser to triage issues, check or re-run Actions and make the small changes you allow.
Only if your prompt says so. Ask for read-only and it only reads and reports, as in the run above.
Yes, through a TinyProfiles browser profile signed in to your GitHub account. Public repos need no sign-in.
Yes. Name the workflow and it can re-run it and report the result. Changing workflow files is better done in a pull request you review.
Copilot helps you write code in your editor. TinyAgent handles repo chores on github.com, such as triage, Actions checks and settings, from ChatGPT, Claude or the API.
Yes. In ChatGPT and Claude you can schedule the task to run on a timer. From code, call TinyAgent once per row from a loop or a queue; accounts start at 2 concurrent runs.
TinyAgent is billed from your Wallet at $0.016 per agent step, not per run. A typical GitHub task in our runs takes around 8 steps, about 13 cents. New accounts get $8 in Wallet funds. Accounts start at 2 concurrent runs.
Free to start. $8 in Wallet funds on us, no card needed.
