Production · v1.4 · 2026

Operate AI agents
in production.

An 8-layer SaaS platform for multi-tenant agent products. TypeScript-first with Mastra, every operation triple-exposed as library + MCP + CLI, outcome-based billing wired through Metronome and Stripe, EU AI Act audit chain on every action.

Sonnet 4.6 defaultVoice p95 < 500ms ~$5/mo dev environment
Live run logprod
14:32:08.184marketing-contentDrafted brief · 1.4s · 0 tools
14:31:51.012revenue-analystSnapshot Q1 · 2.1s · pulled mrr+nrr
14:31:33.876support-triageP1 · routed → billing · human review
14:31:09.402sdrLead tier HOT · 87/100
14:30:44.110competitive-intelInspected acme.com · 3 signals
5 agents · 1 tenant · 24 instances~14ms p95
What you get

The boring parts of agent infra — done.

Memory, evals, billing, compliance, observability. Wired, tested, and on a CI gate so you can spend your time on the agent, not the plumbing.

01

Mastra agents

First-class TypeScript agents with tools, memory, and model routing as primitives. Durable workflows coordinate them with checkpointing and human-in-the-loop pauses.

02

Operation shims

Every external integration is one library function exposed three ways — in-process import, MCP server, CLI binary. Orchestrators import; external agents speak MCP; humans run the CLI.

03

Postgres + pgvector memory

Neon Serverless branches per developer. HNSW vector indexes, append-only event log with hash chain, RLS on every multi-tenant table.

04

Outcome-based billing

Emit a typed outcome event on every billable success. Metronome rates them, Stripe invoices them. No reconciliation drift.

05

EU AI Act ready

Hash-chained audit log (Article 12), human-handoff nodes on every workflow (Article 14), kill switches per-agent and per-tool, PII redaction at the boundary.

06

Evals as a CI gate

Promptfoo regression suite blocks deploys when any prompt drops more than 3 points. Langfuse traces every step. Hamming for voice.

Architecture

Eight layers, no hand-waving.

Each layer has a defined contract, a working reference implementation, and an ADR explaining the call.

  1. 01Data & Context
  2. 02Tool Layer (Operations)
  3. 03Orchestration
  4. 04Reasoning / Models
  5. 05Interaction Surfaces
  6. 06Observability & Evals
  7. 07Compliance & Trust
  8. 08Outcome & Pricing

Ready to look inside?

The full control plane — instances, workflows, memory, HITL queue, catalog, upgrades — is one click away.