Rolling Out AI Company-Wide with Per-Person Budgets and Access - Part 1

Employees in every department use AI tools, and most companies give them access through a shared API key or individual subscriptions. A company-wide rollout gives each person their own key, approved models and budget, on the company's own provider account, from a policy written once for each group of people.

Rahul Singh, Founder, Multi-Router 7 min read

AI tools are used in every department

AI tools have proliferated into every department. Engineers use Claude Code and Codex to write, review and test code. Marketing uses Claude Desktop to draft campaigns and briefs, sales uses it to research accounts before a call, and support uses it to summarize long ticket threads. Each of these tools sends requests to a model provider, and each request is billed to someone's account.

Companies usually provide access in one of three ways. A shared API key gives everyone the same credential, so every request looks the same to the provider, spend cannot be attributed to a person, one heavy user can exhaust the account's rate limits, and removing one person means replacing the key for everyone. Individual subscriptions give each person their own account, but the spend is spread across personal plans and expense reports, and the company has no say in which provider or region serves the traffic. A provider key for each person makes spend attributable, but every key has to be created, configured on a laptop, replaced when it leaks and removed when the person leaves, and model choices and spending limits are set one key at a time.

A company-wide rollout gives each person their own access

A company-wide rollout gives each person their own key, their own list of approved models and their own budget. The traffic runs on the company's own Anthropic, AWS Bedrock or Google Vertex account, in the region the company chooses. The rules are written once for each group of people, and each person sets up their own tools.

Multi-Router provides this with two parts. An admin writes a tools policy that defines what each tool does for the people on it, and each person sets up the tools their policy grants from a web page called the Portal.

A tools policy defines each tool for a group of people

A tools policy is an organization's set of rules for each AI tool it grants. A policy can grant Claude Code, Codex and Claude Desktop, and sets these controls for each tool:

Control Effect
RouteThe provider that serves the tool, such as Anthropic, Bedrock or Vertex.
RegionThe region the provider serves it from, such as the US, on routes with regions.
CredentialsThe organization's provider key, which stays in Multi-Router.
ModelsThe models the tool can use. A model that is switched off is refused when requested.
Per-person capThe most each person can spend per day, week, month or lifetime.
Per-model capsA cap for one model, per person, inside the per-person cap.
Budget rulesNotifications and blocks at chosen points of the cap.
GuardrailsChecks on prompts and answers, such as masking personal data or blocking secrets, before they reach the model.

A company typically writes one policy for each kind of work. A Standard Engineering policy grants Claude Code and Codex with a $200 monthly cap per person, Claude Opus 5 capped at $120 of that, and a guardrail that blocks prompts containing access keys and other secrets. A Go-to-Market policy grants Claude Desktop with a $40 monthly cap per person and a guardrail that masks customer names, email addresses and card numbers. A budget rule on each policy emails the team lead when a person reaches 80% of their cap and pauses the tool at 100% until the period resets.

People · Engineering

Standard Engineering

policypolicy:standard-engineering
Add peopleEdit policy
People
4
3 set up
AI Agents
1
Claude Code
Spend this period
$288 of $600
resets Oct 1 · 14 days
Revision
3
everyone set up is on 3
AI Agents
Claude Code
Anthropic API · $200 per person, monthly · 3 of 3 models · 2 rules
3 of 4 set up
Route
Anthropic API
Credentials
The company's credentialsAnthropic production
Budget
$200 per person, monthly · $288 of $600 spent
anthropic/claude-opus-5$120 per person
Budget rules
NotifyAt 80% of the per-person capemail priya@example.comFired this period
BlockAt 100% of the per-person capblock requestsLive
Models3 of 3 served
ModelWire idStatusPrice / 1M in · out
Claude Opus 5 anthropic/claude-opus-5claude-opus-5Served$5.00 / $25.00
Claude Sonnet 4.5 anthropic/claude-sonnet-4.5claude-sonnet-4-5Served$3.00 / $15.00
Claude Haiku 4.5 anthropic/claude-haiku-4.5claude-haiku-4-5Served$1.00 / $5.00
People4 on this policy
Search this page…
NameAI AgentsSpendRev
Priya Shah
priya@example.com
Code $64.20/$200$64.203
Dana Lee
dana@example.com
Code $182.40/$200$182.403
Marco Ruiz
marco@example.com
Code $41.75/$200$41.753
Aisha Khan
aisha@example.com
Nothing set up——
Showing 4 of 4
PreviousNext
Organization
Engineering
Created
Jun 12, 2026
Created by
Priya Shahpriya@example.com
Revision
3 · Sep 2, 2026
Spend this periodOpen in Budgets
$288of $600
Claude Code$288 of $600
3 of 4 on this page have spent.
History
  • Revision 3 · Claude Code
    Sep 2, 2026 · Priya Shah (priya@example.com)
  • Revision 2 · Claude Code
    Jul 21, 2026 · Priya Shah (priya@example.com)
  • Policy created
    Jun 12, 2026 · Priya Shah (priya@example.com)
The Standard Engineering policy: Claude Code's route, credentials, budget, budget rules and models, the people on the policy and the revision each person is on.

An admin assigns a policy through an invitation, from the account roster or from the policy page. A directory group can also assign it automatically, which Part 3 covers.

Each person sets up their own tools in the Portal

The Portal is a web page, separate from the dashboard, where people set up the AI tools their policy grants. A person signs in with their work email, through single sign-on or a password, and sees the tools available to them.

When a person selects Set up on a tool, such as Claude Code, Codex or Claude Desktop, Multi-Router creates a router and a key for that person and that tool. The Portal then shows the steps to install the tool, the settings to paste into it, and a check that confirms requests reach the model. For Claude Code, the settings come as a settings file for macOS and Linux, a settings file for Windows, environment variables, and a managed settings file that IT can deploy to every machine. For Claude Desktop, the Portal lists each field of the Third-Party Inference settings, including the gateway address and the gateway key, with a copy button beside each one.

Claude Desktop

2 models approved · $40 per month
set up

Install, set up and troubleshoot Claude Desktop

Install, point it at Marketing’s router, check it works, and what the error messages mean.
1

Install Claude Desktop

Install it, then come back here.

Download Claude Desktop from claude.ai/download and sign in with your work account. Then turn on Developer Mode: Help → Troubleshooting → Enable Developer Mode. The app restarts with a Developer menu.

2

Point it at Marketing’s router

Enter these values in the app. The router header is what routes you; without it the key works but nothing answers.
Developer → Configure Third-Party Inference… → Connection
Reveal keyCopy
Inference provider
Gateway
Gateway base URL
https://api.multi-router.ai
Credential kind
Static API key
Gateway API key
sk-mr-••••••••••••••••
Gateway auth scheme
Bearer
Custom inference headers
X-MR-Router = rt_4Tm9Lc2XzB
The key is hidden. Copy still copies the working settings.
3

Check it works

Quit and reopen Claude Desktop (it reads these settings once, at launch), open a new chat and say hello. The model picker lists the models your organization allows.
If you get an answer, you are routed through Marketing and the request shows in Spend by day above within a few seconds. If not, see Troubleshooting below.

Troubleshooting

What Claude Desktop shows, and what it means.
“invalid api key”
The key in your settings is wrong or was replaced. Paste the current one, or issue a new key from Your key above.
“budget … is exhausted for this period”
You reached this period’s cap. It resets on the date shown above; nothing to change on your side.
“model ‘…’ is disabled on this router”
You asked for a model that is not on your policy. Use one under Allowed models.
Claude Desktop in the Portal for a person in Marketing: the install steps, each Third-Party Inference field with a copy button, and the check that it works.

Each person's router and key are their own, so Traffic attributes every request to the person and the tool that sent it, and each per-person cap counts that person's spend alone.

People see their own budget and request more

Each tool has its own page in the Portal. The page shows the person's spend against their cap, the amount remaining, their spending pace over the last three days, the date the cap will be reached at that pace and the date it resets. It lists the models the person can use, with the spend and cap for each, and the rules the organization has set.

Your AI agents

Claude Code

3 models approved · $200 per month · last used Sep 17
set up

Your budget

on track
Monthly · resets Oct 1, 00:00 UTC (in 14 days)
$64.20of $200
32%
Left
$135.80
Your pace
$4.52/day
average over the last 3 days
Cap reached
beyond reset
at this pace
Resets
in 14 days
Oct 1, 00:00 UTC

Need more room this month?

Ask Engineering for a higher limit until the month resets.
Request more budget

Spend by day

this month
Sep 1Sep 17

Set by Engineering IT

Cap
$200 per month, per person
Models
3 allowed, Opus 5 capped at $120
At 80%
notifies you
At 100%
requests paused until reset

Allowed models

3 models on your policy · 1 with its own cap
Claude Opus 5on track
anthropic/claude-opus-5
$41.30 of $120 this month, your own cap34%
last used Sep 17
Claude Sonnet 4.5
anthropic/claude-sonnet-4.5
$19.85 this month · counts toward the $200 total, no separate cap
last used Sep 16
Claude Haiku 4.5
anthropic/claude-haiku-4.5
$3.05 this month · counts toward the $200 total, no separate cap
last used Sep 9
Asking for any other model is refused. A model with its own cap pauses on its own when it reaches it; the others keep working until the $200 total is reached.
Your key …7Qx2 · issued Sep 2
Shown once when it was issued. Issuing a new key deletes this one immediately.
Issue a new key Remove

Install, set up and troubleshoot Claude Code

Install, point it at Engineering’s router, check it works, and what the error messages mean.
1

Install Claude Code

Install it, then come back here.
Install Claude Code using Anthropic's official installer.
2

Point it at Engineering’s router

Add these values to your settings. The router header is what routes you; without it the key works but nothing answers.
macOS / LinuxWindowsEnvironment variablesManaged by IT
~/.claude/settings.json
Reveal keyCopy
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.multi-router.ai",
    "ANTHROPIC_AUTH_TOKEN": "sk-mr-••••••••••••••••",
    "ANTHROPIC_CUSTOM_HEADERS": "X-MR-Router: rt_8Kd2Pq6VnW"
  }
}
The key is hidden. Copy still copies the working settings.
3

Check it works

Run Claude Code and ask it to say hello.
If you get an answer, you are routed through Engineering and the request shows in Spend by day above within a few seconds. If not, see Troubleshooting below.

Troubleshooting

What Claude Code shows, and what it means.
“invalid api key”
The key in your settings is wrong or was replaced. Paste the current one, or issue a new key from Your key above.
“budget … is exhausted for this period”
You reached this period’s cap. It resets on the date shown above; nothing to change on your side.
“model ‘…’ is disabled on this router”
You asked for a model that is not on your policy. Use one under Allowed models.
Claude Code in the Portal: the person's budget and pace, the models their policy allows with Claude Opus 5's own cap, their key, and the setup steps.

A person who needs more budget for the current period selects Request more budget and enters a new limit with an optional message. Approvers receive an email and approve or deny the request in the dashboard. An approved request raises that person's cap for the current period, and the policy stays the same for everyone else.

A person can also issue themselves a new key. The old key stops working immediately, and the budget and usage carry over to the new one.

Policy changes apply to everyone on the policy

Each save of a policy creates a new revision and applies it to everyone on the policy. Each person keeps their router and key and receives the new route, models, budget and guardrails. Moving the engineering team to another region or switching off a model is a single edit to the policy, with no change on anyone's machine.

The policy editor asks for confirmation before changes that affect tools in use. Removing a model or a tool that someone used in the last seven days requires the admin to review the people affected and confirm. Lowering the per-person cap requires an acknowledgment. The policy page lists each revision with who saved it and when, and shows which people are on an older revision.

Admins manage each person's tools from the dashboard

Each person's page in the dashboard lists the tools they have set up, with each tool's status and when it was last used. From that page an admin can replace a person's key, show the setup settings, open the person's traffic, or retire a tool, which deletes its key and router. The account's activity log records Portal access granted and removed, policy changes, tools set up and retired, and keys replaced, with the person who made each change and the time.

Next: single sign-on and directory sync

Assigning policies by hand works for a small team. As a company grows, access has to follow the company directory as people join, change teams and leave. In Part 2 of this series we'll cover single sign-on, which signs people in through the company's identity provider, and in Part 3 directory sync, which assigns each person's organization, role and tools policy from their directory groups. The Self-Service Provisioning docs cover tools policies and the Portal in detail.

Start for free

Create a router, set your SDK base URL to Multi-Router and see the cost of every request.