PXEngine exists because AI chat doesn't survive contact with a team.
Most AI tools get used once, impress someone, and then get abandoned because nobody can explain exactly what happened or make it happen the same way again. We built PXEngine to fix that specific problem — not to make a smarter chatbot.
A workflow a company actually depends on needs to survive someone leaving, a prompt being tweaked, or a model upgrade — none of which a chat transcript can promise.
So PXEngine treats the process as the product. A workstream pins down which agents run, which tools they call, and what the input and output look like at each step. The model still does the reasoning — summarizing, scoring, writing, deciding — but it does that reasoning inside a structure the team defined, not a structure it invents fresh every time someone asks.
How it's built
Four layers, each with one job. Nothing routes around the layer above it.
A single Next.js application. The sidebar adapts to role instead of routing people to a separate admin app, builder app, and viewer app — an org member sees workstreams and projects; an admin also sees agents, widgets, and settings.
A FastAPI backend running Google's Agent Development Kit. Any agent module dropped into the agents directory is auto-discovered — its routes exist without anyone registering them by hand.
Claude models reason inside a workstream's defined steps, calling the specific tools and skills assigned to that step. The orchestrator decides per task whether one agent is enough or whether the work should split across specialists.
PostgreSQL holds session state, MongoDB holds workstreams and agent data, Pinecone holds embeddings for retrieval. Every organization's data stays isolated from every other's.
What we optimize for
Deterministic AI, not random chat
PXEngine doesn't ask the model to "figure it out" on every run. A workstream defines which agents run, which tools they use, and the shape of the input and output — the model brings intelligence within a controlled, repeatable frame, not outside of one.
One interface, not ten apps
There's no separate admin panel, no separate builder app, no separate consumer app. Everything lives in one interface where the sidebar adapts to what a person is allowed to do.
Built for companies, not individuals
Workstreams exist so a team follows the same process through an acquisition, an onboarding wave, or a department scaling up — with multi-tenancy keeping every organization's content isolated from the next.
From transcript to persistent interface
Most AI tools end at a chat log. We're building toward structured views of workstream data that live beyond the conversation — a live dashboard or a pipeline an agent keeps managing over time, not a widget that finishes once and stops.
What's next
In active design or development, roughly in the order we expect to ship.
Smart Tiles & Projects
Persistent, structured views of workstream data that live beyond the chat — turn any completed workstream into an ongoing, managed interface.
CRM & Outreach Modules
Built-in components for entity relationships, communication tracking, and deal pipelines — assembled from the same atomic component system.
Workstream Versioning
Personal tweaks to a shared workstream template without affecting the master version everyone else runs.
Performance Instrumentation
Telemetry to track load times, token usage, and system performance across every component.
Embeddable Widgets
Copy-paste snippet embeds, similar to Power BI or Tableau, so a workstream's output can live anywhere a team already works.
Want to see it on your own process?
Bring one real workflow to Early Access and we'll help you turn it into a workstream.