Every agent you run. And every agent you ship.
Cohesor sits between agents and models. It doesn't care whether the person on the other end is one of your engineers or one of your customers — same routing, same compression, same budgets, same audit trail.
Two directions. One gateway.
Your employees on the other end. Attributed per seat.
Your customers on the other end. Attributed per end-user.
What changes.
The unit of governance. Internally it's a seat — one engineer, one budget, one roster row. In your product it's an end-user you've never met, and a runaway loop doesn't hit an employee's allowance, it hits your margin across every customer at once.
What doesn't.
The path. Every call is compressed, routed by complexity, checked against a budget and written to the ledger — whoever's asking. One endpoint, the same four things, no second integration to build.
The agents your own people run all day.
These never appear in a product demo. They run in the background of every working hour, resending enormous context, and they land as one opaque line on the provider bill.
AI coding assistants
The agents your engineers run all day.
e.g. Claude Code, Codex, Cursor, internal dev agents, CI bots
- ✓Huge resent context and tool results make compression the dominant lever
- ✓Trivial completions go to economy models, hard refactors to the strong end of the set
- ✓Spend attributed per engineer, not per org
Knowledge assistants
Answers pulled out of your own documents.
e.g. enterprise search, RAG apps, research and analyst agents
- ✓Retrieved context dwarfs the question, so compression carries the bill
- ✓Policy applies before egress, not after the answer comes back
- ✓Attribution rolls up per team or per key
Intelligent process automation
The back-office work nobody wants to do.
e.g. document processing, ticket triage, claims and ops agents
- ✓High volume of near-identical calls routes almost entirely to economy models
- ✓Per-key caps stop a stuck loop before it becomes an invoice
- ✓Every run traced, so a failed batch is a query, not an investigation
coding assistant · 3 keys
knowledge assistant · 1 key
ops automation · 2 keys
The agent on the other end is your customer.
When the agent ships inside your product, the unit of governance stops being an employee and becomes someone you've never met — and the thing at risk stops being a team budget and becomes your margin.
Customer service automation
The agent your customers talk to, not your engineers.
e.g. support, help desk, in-app chat, onboarding, virtual agents
- ✓Short turns at high volume make routing the dominant lever
- ✓A per-end-user cap means one abusive session can't move the month's number
- ✓Every conversation lands under its own end-user id
Agentic commerce & booking
Agents that transact on someone's behalf.
e.g. checkout, travel, concierge, scheduling
- ✓Tool calls go through the MCP gateway with per-agent scopes
- ✓Every action is traced, not just every message
- ✓Caps are enforced before a token is generated, not reconciled later
| Time | End user | Model | Tier | Cost |
|---|---|---|---|---|
| 09:41:02 | cus_8842 | deepseek-v4-flash | economy | $0.0041 |
| 09:41:02 | cus_1190 | kimi-k2.7-code | strong | $0.0113 |
| 09:41:01 | cus_8842 | deepseek-v4-flash | economy | $0.0022 |
| 09:41:00 | cus_4417 | — | capped | $0.0000 |
Ship it to a million customers.
Give every application its own key, and every key its own policy. Routing, compression and tool access are set per application; anything you don't override inherits the workspace default.
- ✓A service key per application — one for the widget, one for checkout, one for onboarding
- ✓A profile per key: its own model set, compression settings and connectors
- ✓Hard caps checked at the edge, before a token is generated
- ✓Every usage event carries its end-user id, so cost per customer is a query
support-widget
profile: economy-first · cap $0.50 / end-user
checkout-agent
profile: tools-scoped · cap $2.00 / end-user
onboarding-bot
profile: workspace default · cap $0.25 / end-user
Whoever's asking, the same four things happen.
Building something that isn't on this page? If it speaks OpenAI or Anthropic, it already runs through the gateway — no new integration, no rewrite.
Read the docs →