$ cat ./what-your-end-users-see.md
ship the workflow experience your end-users actually want.
without writing it for them.
this is the view your B2B customer's end-users get when you embed Flowget — they describe what they need in chat, an agent (in alpha) composes from your typed catalog, and you keep the data, the bill, and the audit trail. their tenant never leaves your VPC.
the boundary your customers ask about
what your customer's data sees vs. what our control plane sees
your tenant sees everything
- raw customer data, every field
- secrets, tokens, credentials
- full business logic
- per-node inputs and outputs
- the entire workflow history
our control plane never sees
- customer plaintext — we hold ciphertext, with the codec wired
- secrets — never leave your worker
- your business logic — it runs in your VPC
- raw inputs / outputs of any node
- chat content — the LLM the agent calls runs worker-side (agent in alpha)
we orchestrate. with the codec wired, we never see the bytes. when your customer's legal team asks what crosses the line, you can answer with the full list — and, with the codec wired, it's metadata, not content.
the agent builds from your catalog.
your customer says "when an order comes in, check risk score; if low, charge — if high, escalate." the agent (in alpha) composes the graph from operations you declared. the engine rejects any node not in your catalog — no hallucinated steps, no made-up fields, catalog-grounded by construction.
nothing fires until they approve.
generated workflows land disabled. your customer reviews the graph, tests in sandbox, then enables. no silent prod changes — the agent (in alpha) never ships an automation behind their back.
refine in chat. review the diff.
"change the threshold to 750." the agent (in alpha) updates the graph; your customer sees a one-line diff. no drag-drop required — though the visual builder is there if they want it.
our cluster never sees plaintext.
workflows run inside your environment. with the codec wired, payloads cross our cluster as ciphertext. the LLM the authoring agent calls runs in your worker (authoring agent in alpha; the provider boundary ships today) — the same BYO-worker boundary that protects domain data extends to the model. (the agent calls the LLM you wired in, on your key — what crosses your VPC to that provider is yours to log, redact, or route.)
every run, itemized.
opt-in per-node cost on successful executions — you emit LLM tokens, API calls, compute, storage. if a number changes, the breakdown changes with it. trust by default. (or hide it entirely — that's your call as the host product.)
every decision, replayable.
every run produces a deterministic history. every approval carries the actor and timestamp, ready for your audit log. when the regulator asks what happened on March 14th, you have the record to show them.
the experience you ship. the data you keep.
this is what we ship for your customers — in alpha, choosing partners carefully. talk to us first; ship together.