From request to running pipeline
No prompt handbook, no idle loop: a flow that tries the cheap path first and only gets expensive where there's genuinely something to decide — five steps, then it keeps running on its own.
Describe a topic or task
Describe what's needed in plain language — from a single task to a still-unformed topic ("keep an eye on the market for X").
Workflow or AI? The system decides
ThinkShip first checks whether a deterministic workflow is enough — cron job, trigger, data match, 0 tokens. Only when language understanding, judgment, or an actual decision is needed does it automatically assemble the right agent team.
Agents work, approvals apply
Where AI is genuinely needed, agents work in parallel or in sequence, following the configured three-tier approval model, with budget and token caps.
Results & audit trail
Results land on the taskboard, with a complete record of every action and every cost item — workflow steps and AI steps shown separately.
Keeps running passively, with no ongoing AI cost
Recurring topics keep running as a workflow in passive mode — 0 tokens until a real decision is due again. No one has to submit the request a second time.
A simple workflow
Here's step 2 in its smallest form: the same trigger can take three quite different paths, depending on whether there's anything to do at all.
The simple case: the trigger checks daily. With no hit, it stays a 0-token workflow. On a hit, AI assesses the situation — and, depending on the case, either hands back to the system automatically or to a human for approval.
A realistic workflow: invoice audit
But it also scales to something far more layered: depending on the outcome, the same trigger can take four quite different paths — including two tasks that run at the same time instead of one after another.
And the real case: a daily invoice audit. Most invoices match instantly for 0 tokens. On a mismatch, an AI assessment and a system budget check run at the same time — in parallel, not one after another — and depending on severity, it goes to a quick approval, a finance approval, or straight to a compliance escalation with an audit trail.
Technical foundation
Under the hood, ThinkShip runs as its own stack on your infrastructure: Postgres, Keycloak, NATS, a model broker, plus the backend and frontend services. An installer sets everything up automatically and repeatably — no manual click-marathon through admin panels.
Hardware guidance
A rough starting point for your own planning — three building blocks that scale independently.
Small core system
4 vCPU · 8–16 GB RAM · 100 GB SSD
Postgres, Keycloak, NATS, backend & frontend — enough for a pilot or a small organization on a single VM or mini server.
Workers
2 vCPU · 4 GB RAM per worker
Stateless and horizontally scalable. Rule of thumb: roughly 1 worker per 5 concurrently running agent tasks, added as load increases.
Optional local AI models
16–24 GB VRAM (small) · 48 GB+ (large)
Only relevant if you run models locally instead of in the cloud. Smaller, quantized models (up to roughly 14B) run on a single workstation GPU; larger models need proportionally more VRAM or multiple GPUs.
Illustrative figures for rough planning, not a sizing guarantee — actual load depends on user count, parallelism, and the models you choose.