Assist
The inference control plane
Your team's AI, grounded in your team's expertise.
A general model answers from memory and hopes it is right. Assist is the one governed door every team's inference crosses: it grounds each request in department knowledge and focuses the conversation down to what matters, then routes it to the right model, so a fast, inexpensive model answers as though it already knew your company.
Tuesday, 9:14 AM. A pharmacy counter, a customer waiting.
"Beyond-use dating is 14 days refrigerated."
No source. A revision out of date.
"Per your USP 797 SOP, this preparation is dated 9 days refrigerated."
Cited to sterile-compounding-sop.md
Same model. The difference is the expertise Assist put in front of it.
Sound familiar?
The question your team asks that AI keeps getting wrong.
The most dangerous AI answer is the confident wrong one, cited to a source that does not exist. A general model is a brilliant writer with no librarian: it writes first and never checks. And the real answer lives in your own documents, not the model's memory.
Compounding pharmacy
Assist answers from your own USP compliance SOPs and formulation records, quotes the line it used, and never reaches a model this technician is not cleared for.
A general model answers from whatever it absorbed in training, confidently and with no source, and its number can be a revision out of date.
Legal
Assist grounds every line in the firm's own precedent bank and matter files, cited, so the paralegal verifies in seconds instead of re-researching from scratch.
The model fills the gap with a plausible-looking citation, a case or a clause that was never yours, and the error is invisible until someone checks.
Financial services
Assist scopes the answer to that desk's approved product and suitability docs, and records every change, so an auditor can trace exactly what informed it.
A generic answer risks a compliance breach, and nobody can later say which rule it followed or which desk it came from.
What a confident wrong answer costs
A compliance event
In a regulated shop a fabrication is not an embarrassment; it is an incident, with a paper trail.
Cracked trust
The first confident wrong answer teaches the team to distrust every answer that follows it.
The re-verification tax
Staff re-check everything by hand, so you pay for the AI and for the manual work it was meant to replace.
One control point. Every team's own expertise.
The difference, quoted
"Per your USP 797 SOP, this preparation is dated 9 days refrigerated."
cited to sterile-compounding-sop.md, the answer your SOP already holds, at the counter, at 9:14 AM.
Context refinement
Ground the request, then focus the conversation.
A data-plane gateway routes traffic without ever reading it. Assist reads and reshapes each request in two opposite motions, at one governed door, scoped by the same group permissions, and invisible to the client.
Ground
Add the signal
A group-scoped brain search injects the department's knowledge, library material, and behaviors relevant to the request, so a smaller model answers as if it had read the handbook.
Focus
Remove the noise
A long conversation is trimmed to the turns that matter, blending relevance with recency, capped to a budget, with tool round-trips folded into plain prose. The turn a user references is pulled back into focus instead of buried.
Both are deterministic and model-free: same input, same output, no embeddings and no extra dependency.
Why it works
A smaller model, grounded, beats a bigger one guessing.
A frontier model is expensive partly because it must already know everything, and it drifts when a long history buries the thread. Assist grounds each request in your context and focuses it down to what matters, so you pay for a focused prompt, not the whole transcript resent every turn, and a fast, inexpensive model answers from your sources instead of from what a giant model memorized.
Answers from sources, not guesses
Pay for a focused prompt, not the transcript
Less long-context drift
Answers your org controls
What Assist grounds in
A brain: your expertise, modeled on a mind.
A brain holds what it has looked up, what it understands, and what it knows how to do. Each department reaches only its permitted brains, and Assist folds the right memory into each request.
Semantic memory
Knowledge
Its own authored understanding, the opinions and syntheses written into the brain and weighted as its own.
Procedural memory
Behaviors
Procedures it applies whole, a way of driving a tool or carrying out a task, invoked as a skill rather than recited.
Extended mind
Library
External sources it fetches and cites but never claims as its own, the reference shelf it consults on demand.
How it works
One control point between your agents and your expertise.
Auth, grounding, focus, routing, and access control move out of every agent and into one control point, so your harnesses stay thin and the hard parts are solved once. Every request from every harness crosses the same path.
One conversation, end to end
-
Your client sends its conversation
Any harness: Claude Code, the Claude app, Cowork, an OpenAI client.
-
Authenticated and authorized
The request proves who it is and reaches only the models its group permits.
-
Grounded in governed expertise
Your department's knowledge is folded in, scoped by the same access rules.
-
Focused to what matters
The conversation is trimmed to the turns that carry the answer.
-
Routed to the right model
Sent to the model that fits the task, behind a stable alias you control.
-
Answered, authoritative and transparent
The client gets a grounded, cited answer, never the machinery behind it.
One governed path, and the client never sees the rewrite.
Drop-in for the clients you already run
Point Claude Code, the Claude app, Cowork, or any OpenAI client at the gateway by setting its base URL and sending your token. No SDK, no code change.
POST /v1/messages
POST /openai/v1/chat/completions
Clients name a stable model alias; you repoint it to the upstream that serves it, with no client change.
Governed at one gate
The same path that grounds a request also controls it, so access lives in one place instead of scattered across every team's harness.
Authentication is mandatory
Every request needs a valid, unrevoked token. There is no flag to turn it off, because a control that can be left off gets left off.
Model access fails closed
A caller only reaches the models their groups permit. With no group context the request is rejected, not waved through to a default.
Every change is on the record
Who may call, which models and expertise they may reach, and every change to the rules, all on the record.
The brain augments, it never blocks. A slow crawl, a failed search, or an unreachable source degrades to no context, and your model's answer still returns.
A control plane, not a data plane
An ordinary gateway forwards. Assist decides.
A data plane routes, meters, and logs traffic without ever reading it. A control plane reads each request and reshapes it against org policy. The blank cells below are the difference.
| Capability | Data plane | Assist |
|---|---|---|
| Model routing and failover | ||
| Auth, rate limiting, cost, logging | ||
| Access control over models | ||
| Deep observability, semantic cache, guardrails | Roadmap | |
| Grounding: per-request department knowledge | — | |
| Focus: relevance and recency windowing | — | |
| Access control over knowledge, not just models | — |
Model routing is table stakes; deep observability, caching, and guardrails are on the roadmap. The blank cells are why Assist is a control plane rather than a pipe.
Get started
Give every team the AI it can trust.
Point your agent at the gateway and give each team exactly the expertise it should have, grounded, cited, and under one access layer.