Repos, artifacts, and the reading list behind the design.
This page is for developers. plnt.chat runs on plnt, an open-source agent runtime (Apache-2.0) you can read, run, and build agents on. The prior art that shaped the design is linked, not assumed.
Four repositories, one story.
The open-source agent runtime that powers plnt.chat. Temporal sagas, sandboxed execution, JSONL audit logs. Apache-2.0, usable standalone.
Pluggable agent workflow recipes. Each ships as a versioned spec + code. The skill-bundle format the marketplace agents use.
Earlier product experiment: an AI copilot for Google Maps profile owners. Superseded by plnt.chat, kept for reference.
Astro static site + Preact playground. Marketing pages + developer docs + interactive demo. Deployed to Vercel.
The things you can click, curl, and pull.
Pick a workflow, watch the orchestration timeline, invoke it. Browser UI backed by the playground API.
HTTP endpoint. /v1/workflows, /v1/workflows/<id>/invoke (SSE), /healthz.
Multi-arch python:3.12-slim, ~150 MB, non-root, health-checked. Published to GHCR.
Prior art the design is built on.
Sagas, compensation, retry policies. The playbook for orchestration workflows that survive process restarts.
The reason retry budgets sit at the edge (Envoy), not per-hop. Chapter 22 of the SRE book.
The reference K8s CRD + controller for inference workloads. plnt's WorkflowRun borrows heavily from the shape.
DAG-of-steps orchestration on Kubernetes. Inspiration for the workflow spec shape (steps + deps + tool bindings).
Chart-per-runner packaging, values.yaml discipline, hooks, tests. The reason plnt does not invent its own package manager.
Traffic weighting for canary → stable transitions. plnt writes VS weights during promote / rollback.
Two minutes on a local kind cluster.