Skip to main content
Barndoor Bridge is in Early Access. These pages are not listed in the navigation. Overview: /how-tos/barndoor-bridge/overview · Using Bridge: /how-tos/barndoor-bridge/using-bridge · Install with MDM: /how-tos/barndoor-bridge/install-with-mdm · Client settings: /how-tos/barndoor-bridge/client-settings · Publish a local MCP: /how-tos/barndoor-bridge/publish-local-mcp
A governed coding session, started from Bridge or with barndoor run, is configured by two things:
  • Model Access decides which models the developer may use. It is the only thing that grants or denies a model.
  • Client settings decide which of the client’s own native capabilities stay on: files, shell, network, browser, secrets, skills, and subagents.
Neither one decides which MCP servers or Barndoor tools a session can use. Every governed session gets exactly one MCP server, Barndoor, and what it exposes is decided by policies and tool confirmations. A developer’s own locally configured MCP servers are not loaded.
Client settings replace launch profiles. There is no longer a profile to pick or a CLI id to pin: see Moving from launch profiles.

Where to find them

Open LLM Management, then Client settings under Models. The page has one tab per client: There is one set of settings per client for the whole organization. They apply to every developer in it; to treat groups or users differently, target them with Model Access policies.
The Client settings page: Claude Code and Codex tabs, the Claude Code settings heading, and the What the client can do capability tiles at the client default, with a Save button
An organization that has never saved client settings uses the client default: every capability on, exactly as the client ships. You do not have to save anything before the first launch.

Which models a session gets

A launch never picks models from a list you maintain. At launch, the gateway works out every Model Route the developer is allowed to use and fills the client’s model choices from that set. The allowed set is every enabled Model Route that:
  • the developer’s Model Access policies allow, and
  • can serve a request right now. A route whose providers are all unhealthy is left out until the next launch or renewal.
Routing policies the developer is allowed to use count too. The session’s credential is scoped to the whole allowed set, and the gateway refuses every other model name.

Claude Code

Claude Code has a model for each Claude family plus a default. Each is filled with the newest allowed route of that family, judged by the Claude model behind the route’s primary target: When two routes carry the same model generation, the gateway prefers a route that has at least one API-key, Bedrock, or Vertex target over a route that uses only Claude logins. After that, a route whose primary is the Claude login wins. 1M context is worked out for you. When every model on a route is a known 1M-context Claude model, Claude Code is told so and compacts at 1M tokens. Otherwise it compacts at 200k. Any other allowed route, including a non-Claude model, can be selected in Claude Code with /model <route name>.

Claude Desktop

Claude Desktop asks the gateway for its model list. The picker shows every allowed Claude route, labelled by family, with the session’s Default first and a 1M option where the route supports it. Routes for non-Claude models are not shown in Claude Desktop.

Codex

Codex starts on the newest allowed GPT route (a route whose primary model is gpt- followed by a version). With none, it uses the first allowed routing policy. When two routes carry the same generation, a route on a ChatGPT sign-in wins over one on an API key. In Codex Desktop the developer can switch to any other allowed model that the installed Codex supports.

Seeing which route fills what

LLM Management → Model Routes shows this before anyone launches. Under each route’s name, badges such as Claude Code · Opus show which model choices the route fills. The badges describe an organization member whose Model Access allows every route; a developer allowed fewer routes may get a different one.
Model Routes list: gpt-5.5 badged Codex · Default; claude-haiku-4-5 with an amber warning badge Claude Code · Haiku; claude-sonnet-4-6 badged Claude Code · Sonnet; claude-opus-4-8 badged Claude Code · Default, Opus and Fable
A route whose every target uses a Claude login gets an amber badge: Anthropic accepts a Claude login only from Claude Code, so Claude Desktop, Codex, and API clients get a 429 on that route. Add an API-key, Bedrock, or Vertex target to the route as a fallback.
Tooltip on the amber Claude Code · Haiku badge: Claude Desktop can't use this route: every target uses a Claude login, which only Claude Code can use. Add an API-key, Bedrock or Vertex target.
When you create a route or add targets to one, the dialog warns about a Claude-login-only route and says what will change for Claude Code and Codex.
Create Model Route dialog with one Claude login target selected: a warning that every target uses a Claude login (seat) token, and the note Claude Code will use this route for Default and Opus instead of claude-opus-4-8
On the Claude Code tab of Client settings, Create Claude routes (shown when you have an Anthropic login provider) creates direct Model Routes for your Anthropic provider’s current Claude models, first enabling the newest model in any family that has none enabled. It also gives each Claude-login route a fallback target on your first API-key, Bedrock, Vertex, or Foundry provider that carries the same model. It lists any routes it could not give a fallback and any Model Access allowlists that would still block the new routes, with Add to these policies to fix those in one step.
The Create Claude routes dialog: an explanation of the routes and fallbacks it creates, an Anthropic login provider picker, and a Create routes button
Claude routes ready: models enabled, routes created and already available, a fallback added per route, a warning listing claude-haiku-4-5 as working only for Claude Code, and a Model-access policies need an explicit update box with an Add to these policies button

When no model fits

If the developer is allowed no route that fits the client, the launch is refused:
The fix is a Model Access change, not a client settings change. The launch also appears under Recent failed launches.

When changes take effect

  • Saving client settings creates a new version. New launches use it, and the CLI prints the version it launched with, for example barndoor: governed claude session using Claude Code settings v3.
  • Running sessions keep the capabilities and model choices they started with. The developer restarts the client to pick up an edit.
  • Session credentials renew about once an hour. At renewal the allowed set is worked out again, so a route you removed from someone’s Model Access stops working within the hour, and a route you added becomes usable. The model choices themselves do not move until the next launch. If nothing the developer is allowed fits the client any more, renewal fails and requests return barndoor: governed session renewal failed — check network and retry.

What the client can do

Every client starts at its own defaults: files, shell, network, browser, inherited environment, skills, and subagents all on, exactly as the client ships. Governance only takes capabilities away. Click a tile to govern one capability, or use Governance → Apply a bundle to set several at once. Each departure from the default appears as a restriction you can remove. Click Save to apply. Some things apply to every governed session, even one with no restrictions:
  • Model traffic goes only to your LLM Gateway, with Model Access, data protection, and audit applied to every request.
  • The session runs on a short-lived Barndoor credential that is tied to the user and the client’s agent and can reach only that client’s model API.
  • The developer’s own client settings files and MCP servers are not loaded. The only MCP server is Barndoor.
  • Usage and audit records carry the client settings version the session launched with.
What client defaults do not cover is anything the native tools do on their own. Native shell commands, native file edits, native web access, and credentials in the developer’s environment act outside Barndoor tool policy.
Through Barndoor means the native tool is removed and the client is told to use the Barndoor MCP server for that kind of work. Whether a Barndoor tool for it exists, and whether it needs confirmation, is decided by your policies, not by client settings.

Files

Native file tools: Read, Edit, Write, Glob, and Grep, plus NotebookEdit on Claude.
The Files panel: Read & write, Read only, and Through Barndoor, with chips to block Read, Edit, Write, Glob, Grep, or NotebookEdit individually
Files does not sandbox the shell. A native shell can read and write files whatever this setting says, so a native shell requires Read & write. Lowering Files moves a native shell to Through Barndoor. barndoor run --worktree also requires Read & write.

Shell

The Shell panel: Native, Through Barndoor, and Off, a chip to block Bash, and the Deny and Pre-approved command pattern boxes
Through Barndoor and Off remove the same native tools. Neither removes the Barndoor shell tool itself. To stop commands running through Barndoor, deny or require confirmation for the shell tool in your policies.

Command patterns (Claude only)

These appear in the Shell panel and apply to Claude’s native Bash tool. Write one pattern per line, in Claude Code’s permission syntax, for example git push --force * or npm run test:*.
  • Deny command patterns. Matching Bash commands are refused. They are not applied to commands run through the Barndoor shell tool; govern those with policy.
  • Pre-approved command patterns. Matching Bash commands run without a Claude permission prompt. Only available while Shell is Native.
Both require a native Bash tool. Codex does not support command patterns, and settings that have patterns but no native shell cannot be saved.

Network

Native web tools: WebFetch and WebSearch on Claude, web search on Codex.
The Network panel: Native, Through Barndoor, and Off, with chips to block WebFetch or WebSearch
On Claude Code, Network governs the client’s web tools, not the machine’s network. A native shell can still reach the internet with curl, git, or a package manager. To keep commands off the network as well, set Shell to Through Barndoor and govern the shell tool with policy, or block outbound traffic at the device or network layer.
Governing Network turns a native Browser off.

Browser

The Browser panel: Native or Off
A native browser requires native network. Claude Desktop ignores this setting, and Codex Desktop always runs with its browser off; see Desktop clients.

Secrets

What the client process, and everything it starts, can see in the developer’s environment.
The Secrets panel: Inherited or Scrubbed
Scrubbed removes environment variables only. Credential files under the developer’s home directory, such as ~/.aws or ~/.config/gh, stay readable to anything that can read files. On Claude Code, Scrubbed also means Claude’s plan usage is not shown in the session.

Skills / Subagents

The Skills and subagents panel: a switch each for Skills and Subagents, with chips to block the Skill and Task tools

Blocking individual tools

Each panel lists the native tools in that area as chips. Click one to block that tool alone without changing the rest of the capability, for example block WebSearch but keep WebFetch.
The Codex Skills and subagents panel, with chips to block Skills, Subagents, Apps, Plugins, Hooks, and ImageGeneration
Codex: governing any capability turns off Computer Use, Apps, Plugins, Hooks, and Image Generation. These features have no tile. They stay on while every capability is at the client default, even with tools blocked, and switch off as soon as any tile is changed or a bundle is applied.

Bundles

The Governance menu: Apply a bundle with Guarded development and Locked down, and Clear all governance
Applying a bundle sets every capability above. Blocked tools and denied command patterns are kept; pre-approved patterns are removed. Because both bundles move the shell off native, settings that keep denied command patterns cannot be saved: clear them first. Change any tile after applying a bundle and the page simply shows the resulting restrictions. Clear all governance returns the client to its default and removes every rule.
Claude Code settings after applying Guarded development: five governed tiles and one removable restriction row for each of Files, Shell, Network, Browser, and Secrets

Desktop clients

Some things are always off in a governed desktop session, whatever the settings allow. Settings can only turn more off.
  • Claude Desktop: computer use and Cowork’s workspace shell are off, local MCP servers are off, and Barndoor is the only MCP server.
  • Codex Desktop: browser use, computer use, Apps, Plugins, Hooks, and Image Generation are off, Codex’s own approval prompts are off, and Codex Desktop can only change settings Barndoor does not control.

Rules the editor enforces

Settings cannot be saved when:
  • Shell is Native and Files is not Read & write. The editor narrows the shell for you when you lower Files.
  • Browser is Native and Network is not Native. The editor turns the browser off for you when you govern Network.
  • Command patterns are set and there is no native Bash, or the client is Codex.
  • A tool is both blocked and allowed.
The editor never turns a capability back on for you. Developers can narrow their own session but never widen it. barndoor run --posture governed routes a native shell through Barndoor, and --posture paranoid applies the Locked down settings, whatever the client settings say. No flag turns on a capability the settings turn off.

Recent failed launches

The bottom of the Client settings page lists launches that failed in the last 7 days, grouped by client. Each row shows the time, the user, the reason (for example Model access denied, Model route unavailable, or CLI out of date), how many times it repeated, and the message the developer saw. It is the first place to look when someone reports that Bridge will not launch.
Recent failed launches: two Claude rows with the reason Model access denied and the message None of the models you're allowed to use fit Claude Code, and one Codex row with the reason CLI out of date

Moving from launch profiles

Client settings replaced launch profiles in October 2026. For an existing organization:
  • Your default launch profile for each client became that client’s settings, with the same capabilities. Other profiles were archived, not deleted.
  • Models now come from Model Access. The routes a profile listed are no longer read. If a developer could launch before because their profile named a route, make sure their Model Access policies allow it.
  • barndoor run --profile and the MDM default_profile key are ignored. The CLI prints barndoor: --profile is deprecated and ignored; your organization's Claude Code settings apply (or the same line for the MDM setting). Remove --default-profile the next time you re-render your MDM profiles; see Install with MDM.
  • Usage reporting groups governed usage by Client settings, labelled with the settings name and version.
  • The admin API has new GET /admin/client-settings, GET /admin/client-settings/{client}, and PUT /admin/client-settings/{client} endpoints, where {client} is claude or codex. The /admin/agent-runtime-profiles endpoints are deprecated and still served for at least 90 days.