Beta
Download

How to roll out AI coding tools to an engineering team

AutoRouter8 min read

Treat an AI coding rollout as six decisions to make before the first seat goes out: sign-in, how settings reach each machine, which models each team may use, what each team can spend, what you report on, and what happens to prompts and code. Anthropic, OpenAI and GitHub each give admins controls for their own tool, and the ones that hold are enforced by managed settings or the vendor's server. If your developers work in Codex or Claude Code, AutoRouter, the Mac app for Codex and Claude Code, has an Enterprise plan that puts allowed models, team budgets and zero data retention in one place and reports spend by team, project or ticket.

AutoRouter wrote this post. Vendor plans, prices and controls are as of October 22, 2026, and each one links its source.

What goes on the rollout checklist?

Identity comes first, because Anthropic applies zero data retention and server-side model restrictions only to people signed in to your organization:

  1. Identity. Require SSO, use SCIM where your plan includes it, and pin sign-in to your organization. Claude Code's forceLoginMethod and forceLoginOrgUUID managed settings do the pinning, and Anthropic says zero data retention covers only requests that sign in to the ZDR organization, so a session on a personal account isn't covered.
  2. Deployment. Push policy to every machine so developers have nothing to configure. Claude Code and Codex both read a macOS configuration profile that Jamf or another MDM can carry.
  3. Models. Decide which models each team may use, and enforce that list with an allowlist. A managed default model restricts nothing: Anthropic calls it "a starting default, not a lock", and OpenAI calls Codex workspace model choices "defaults, not permissions".
  4. Budgets. Set caps that stop usage when they're reached. GitHub warns that a Copilot budget without "Stop usage when budget limit is reached" sends a notification while charges keep accruing.
  5. Reporting. Pick the unit you'll be asked about (person, team or project) and check that each tool reports it. All three report per user; GitHub also reports by repository, and Claude Code can stream per-user token and cost metrics into your own stack over OpenTelemetry.
  6. Data retention. Decide whether you need zero data retention, and check which features it removes: under Anthropic's ZDR, features that store prompts, such as cloud sessions and Remote Control, are turned off.

How do you push the settings to every Mac?

Deploy each tool's managed configuration as a macOS configuration profile through your MDM, then confirm it on one test machine before the rest.

For Claude Code, put the keys in the com.anthropic.claudecode preferences domain, or in /Library/Application Support/ClaudeCode/managed-settings.json. Managed settings override every user and project setting, and as of October 22, 2026, Claude Code checks MDM policy for changes every 30 minutes. This example pins claude.ai sign-in to one organization and allows only Sonnet and Haiku models:

{
  "forceLoginMethod": "claudeai",
  "forceLoginOrgUUID": "xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx",
  "availableModels": ["sonnet", "haiku"],
  "enforceAvailableModels": true
}

enforceAvailableModels also keeps the picker's Default option inside the list, and it needs Claude Code v2.1.175 or later. Run /status on the test machine: its Setting sources line should show Enterprise managed settings. On Team and Enterprise plans, the claude.ai admin console can deliver the same keys instead, and Claude Code fetches them at startup and every hour.

For Codex, write a requirements.toml. OpenAI defines requirements as admin-enforced constraints that users can't override. This example from OpenAI's docs blocks --ask-for-approval never and --sandbox danger-full-access:

allowed_approval_policies = ["untrusted", "on-request"]
allowed_sandbox_modes = ["read-only", "workspace-write"]

Install it at /etc/codex/requirements.toml, or base64-encode it without line wrapping and deliver it in the com.openai.codex domain under the requirements_toml_base64 key. OpenAI names Jamf Pro, Fleet and Kandji as tools that can distribute it.

How much should you budget per developer?

Start from a pilot group's real usage. As of October 22, 2026, Anthropic says Claude Code averages about $13 per developer per active day and $150 to $250 per developer per month across enterprise deployments, with 90% of users under $30 per active day.

Agent work past a seat's allowance is billed as usage, so the cap matters as much as the seat price:

What do Anthropic, OpenAI and GitHub's own controls cover?

Each vendor covers SSO, model access, spend caps and reporting for its own tool, and most keep SCIM, audit and retention controls for the top tier, as of October 22, 2026:

ControlClaude Code (Anthropic)Codex (OpenAI)GitHub Copilot
SSOTeam and EnterpriseSAML SSO and MFA on BusinessThrough GitHub: SAML SSO or Enterprise Managed Users
SCIMEnterpriseEnterprise, Edu and HealthcareEnterprise Managed Users, provisioned from your IdP
Model allowlistavailableModels in managed settings; Enterprise can also disable models server-sideAdmins enable models for the workspace (Enterprise, Edu)Owners enable or disable models for Business and Enterprise seats
Spend capsSpend limits by organization, group or memberUsage limits and spend controls on workspace creditsBudgets by enterprise, cost center or user
Usage reportingSpend report by user and model; Analytics API on Enterprise; OpenTelemetryCodex analytics dashboard; Analytics APIMetrics dashboard and APIs, down to repository and user
AuditAudit logs and Compliance API on EnterpriseCompliance API for audit recordsAudit logs on Business and Enterprise
Data retentionCustom retention on Enterprise; ZDR for qualified Enterprise accounts, enabled by AnthropicRetention and residency controls on Enterprise; Work Cloud local computer access isn't strict ZDRIDE prompts not retained; GitHub.com, mobile and CLI prompts kept 28 days

Each set of controls stops at its vendor's tool, so a team that uses two of these tools keeps two sets of model lists, budgets and reports.

Where does AutoRouter Enterprise fit?

AutoRouter Enterprise fits teams on Macs whose developers work in Codex, Claude Code or both. AutoRouter, the Mac app for Codex and Claude Code, adds 20+ models to both apps, plus one more choice: AutoRouter, which picks a model for each turn and uses the ChatGPT and Claude plans people already pay for first. Every turn shows its model, what it cost and why. How AutoRouter picks a model explains the routing, and Does model routing make your results worse? answers the quality question.

Enterprise is available and priced by seats for the team. It includes everything in Pro, plus:

Two features are coming soon and aren't available yet: what each pull request cost (who shipped it, which models did the work and what it cost), and one skills library for a whole team. AutoRouter is in beta, runs on macOS 14 or later, and works with Codex (the app and the CLI) and the Claude Code CLI, so GitHub Copilot and machines that aren't Macs still need their vendors' own controls. To plan a rollout, email sales@autorouter.com.

Questions

Can developers override managed settings?

No, developers can't override a managed key in Claude Code or a requirement in Codex. In Claude Code, nothing in a developer's own files or the --settings flag overrides a managed key, apart from a few keys where the stricter value wins. In Codex, requirements can't be overridden, while managed defaults are only starting values.

How much do the seats cost?

As of October 22, 2026, Claude Team Standard seats cost $20 a month billed annually or $25 monthly, Premium seats cost $100 or $125, and Claude Enterprise costs $20 a seat a month billed annually, with a 20-seat minimum, plus usage at API rates. On the same date, ChatGPT Business cost $20 per user a month billed annually or $25 monthly, ChatGPT Enterprise was priced through sales, and Copilot Business and Copilot Enterprise cost $19 and $39 per seat a month. AutoRouter Enterprise is priced by seats through sales@autorouter.com.

Does AutoRouter see our prompts?

Requests pass through AutoRouter's servers on their way to each provider, and AutoRouter doesn't store transcripts. It keeps the message that opens each turn, up to 4,000 characters, to review routing, along with records of the model chosen, tokens and cost. Details are in AutoRouter's privacy policy; for how Enterprise zero data retention applies to your organization, ask sales@autorouter.com.

Sources

Each vendor page was read on October 22, 2026.