Pintle
The people
An issue-filing agent your CI calls, with prompts and logs you keep
Jonah Petrakis
Founder
Led a 40-engineer platform team whose incident triage lived in a vendor’s chat bot. When the vendor was bought he had two weeks, no API, and rewrote the rules from memory.
Meredith Calder
CTO
Built the job runner at a CI company for six years and never let a pipeline definition live anywhere but the customer’s repo. She wanted an agent that kept the same rule.
Colin Abernathy
CPO
Ran developer relations at an API company through two acquisitions and read the same forum thread both times: is there an export, and how long do we have.
On 9 September 2026 Meta said it was buying the company behind an agent many platform teams had plugged into Slack, Linear and GitHub. When CTOs asked what their teams could keep, the answer was nothing: no public API, with prompts and run history inside the vendor’s runtime.
Pintle runs agent skills, such as filing issues and drafting pull requests, only when something you own calls it. A pipeline step, cron job or Slack bot sends the task and a pointer to YAML and markdown prompts in your GitHub repo. Every run is logged to storage you chose, so if we shut down, the workflows and the record stay with you.
A language model plans each Pintle run, but most of the machine learning goes into matching incident threads to existing issues. Rules such as a quoted issue key or an identical error settle about two matches in five. The rest go to embedding search over that customer’s own issues, then a classifier that scores each candidate against the thread.
Below a confidence of 0.85, Pintle’s match classifier files nothing. The run returns ranked candidates, and a person picks one with a button in their Slack bot or a manual gate in CI. That choice is stored as a labeled pair, the candidates offered and the one picked, without the thread text, and tunes only that customer’s matching.
Pintle’s GPU hours pile up around a new customer’s backfill, a cluster of incidents and a prompt change. A backfill embeds twelve months of Linear issues, tens of thousands, in about an hour; a bad deploy at 17:00 can send forty runs in ten minutes; a prompt change replays five hundred stored runs before merging. Replays and index refreshes cost no credits, so two of those bursts go unbilled.
Pintle’s matching runs on cloud GPU capacity we control in Dublin because incident threads carry customer names and hostnames. A Swedish scale-up’s data agreement will not let those leave the EU for a model someone else hosts. Pintle’s own storage sits on cloud infrastructure in Dublin too. Only the planner’s text generation runs elsewhere, at one hosted model provider’s EU endpoint under zero-retention terms.
Pintle moves its planner onto the GPU capacity it controls in Dublin during 2027. By Q1 2027 incident triage and PR drafting run on Pintle’s fine-tuned open-weight planner; the other skills follow by Q3 2027. From then the hosted provider is opt-in per key, and a key without it runs every model in Dublin.
Pintle Ltd, an Irish company founded in February 2026, went live in September 2026. A new key starts with 5,000 sandbox credits, and paid packs are bought from the same page.