> ## Documentation Index
> Fetch the complete documentation index at: https://docs.barndoor.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# How Bridge works

> For admins and security reviewers: the parts of Barndoor Bridge, the path model and tool traffic takes, what every governed session has, what Barndoor sees, and what Bridge doesn't do.

This page is for admins and security reviewers who want to know what Bridge does with model and tool traffic before rolling it out. To set Bridge up, start with [Set up your organization](https://docs.barndoor.ai/how-tos/barndoor-bridge/set-up-your-organization.md).

## The parts

| Part | Where it runs | What it does |
| - | - | - |
| **Bridge app** | The laptop | The **Barndoor** window and menu-bar item (system tray on Windows), and the `barndoor` command. Signs people in, keeps configuration synced, launches each client in a Barndoor-managed configuration, and runs Barndoor's local tools. |
| **LLM Gateway** | Barndoor platform | Receives every model request from a governed session. Applies Model Access, data protection, and audit, then forwards to the provider. |
| **MCP servers and policies** | Barndoor platform | Your organization's connected MCP servers, and the policies that decide, per call, whether a tool may run and whether it needs approval. |
| **Admin UI** | Barndoor platform | Where admins set Model Access, client settings, policies, and MCP servers, and read usage and audit. |

Bridge doesn't govern anything on its own. It routes each client through your Barndoor organization, and the organization's settings hold the rules.

## Model traffic

**1. The client sends its requests to Bridge**

A governed client sends its model requests to Bridge on the same laptop, never directly to Anthropic or OpenAI.

**2. Bridge forwards them to the LLM Gateway**

Bridge forwards only to your organization's Barndoor LLM Gateway. Each session uses a short-lived Barndoor credential, tied to the person and the client, that Bridge renews about once an hour and revokes when the client exits. The credential can reach only the models the person is allowed.

**3. The gateway applies your rules**

[Model Access](https://docs.barndoor.ai/how-tos/use-llm-controls.md#model-access) decides which models the request may use. Data protection runs on the request and the response. The call is audited with the person, the client, and the agent.

**4. The gateway calls the provider**

The Model Route decides where the request goes: Anthropic, OpenAI, Amazon Bedrock, Google Vertex AI, or another configured provider.

## Tool traffic

Every governed client gets exactly one MCP server, Barndoor (none when Bridge governs models only; under MCP only, a person's own Claude Desktop keeps its other servers too). It gives the agent a small set of tools for finding and calling everything else. Behind it are:

| Tools | Where they run | Policy | Data protection | Audit |
| - | - | - | - | - |
| **Remote MCP servers** the organization connected in Barndoor (Jira, GitHub, Salesforce, Slack, and so on) | Barndoor platform | Yes | Yes | Yes |
| **Built-in local tools** (Shell, Filesystem, Git, GitHub, AWS CLI, Kubernetes, Docker, Terraform, and others) and bundled skills. An admin turns each one on | The laptop | Yes | Yes | Yes |
| **The organization's own local MCP servers**, signed by the organization and pinned with MDM | The laptop | Yes | Yes | Yes |

**Policy** is checked on every call, at call time. For tools that run on the laptop, Bridge asks Barndoor before running anything. If Barndoor can't be reached, Bridge falls back to a signed copy of the policy for the tools the session has already used, for up to 24 hours. Any other tool is refused.

**Approvals** come from your organization's policy in the [Access Control Center](https://docs.barndoor.ai/how-tos/manage-policies.md). A tool set to **Needs Approval** pauses until the person approves the call; a tool set to **Always Allow** runs; a tool the policy doesn't grant is refused. Approval is per call: the person sees exactly what will run, with its arguments, and approves that call only. Apart from Bridge's own bundled skills and utilities, which ask before a destructive action, Bridge adds no approval rules of its own. See [Tool confirmations](https://docs.barndoor.ai/how-tos/tool-confirmations.md).

**Data protection** applies to every tool call, wherever the tool runs. Remote MCP servers are scanned on the Barndoor platform as their calls pass through it. For a tool that runs on the laptop, Bridge sends the tool's input to Barndoor for scanning before the tool runs, and its output before the client sees it, then applies the verdict: allow, redact or tokenize (the tool or the client gets the rewritten text), or block. Scanning happens on the platform, not the laptop, so verdicts, wording, and audit match remote calls.

* Only your organization's [data protection policies](https://docs.barndoor.ai/how-tos/data-control-center/mcp-policies.md) act. With none for a server, nothing is sent for scanning.
* **It fails closed.** If a call can't be scanned, for example because Barndoor is unreachable, the call is refused or its output withheld rather than delivered unscanned. The one exception: if Barndoor is unreachable and your organization had no data protection policies when Bridge last checked, in the past 24 hours, local calls run unscanned.
* A skill's steps are checked and scanned one tool at a time, and a block stops the skill.
* Bridge's own helpers, such as memory across sessions and session history, aren't checked against policy or scanned. They don't reach outside Bridge.
* [Field controls](https://docs.barndoor.ai/how-tos/data-control-center/field-controls) aren't evaluated on local servers. If you set one on a local server, that server's calls are withheld.

**Shell commands** that the agent runs through Barndoor's shell tool are checked against policy and audited as `shell/shell_exec`. The audit record names the tool and the policy decision, not the command. When a data protection policy covers the Shell server, the command is also scanned before it runs. Admins can also block commands outright with [client settings](https://docs.barndoor.ai/how-tos/barndoor-bridge/client-settings.md#shell).

### Organization-published local MCP servers

A local MCP server your organization publishes runs only if three independent parties agree:

| Who | Holds | Can't do on its own |
| - | - | - |
| **Your publisher** | The private signing key. Signs each release. | Get code onto a laptop whose MDM profile doesn't pin the key. |
| **Your IT, through MDM** | The public key, and the hosts builds may download from. | Change what a release contains. |
| **Barndoor** | The signed release, the installation, and the policy. | Change the release: any edit breaks the signature. |

Bridge checks the signature against the pinned key, downloads the build only from an allowed host, checks its SHA-256, and checks it again on every launch. If any check fails, the server doesn't run. See [Publish a local MCP server](https://docs.barndoor.ai/how-tos/barndoor-bridge/publish-local-mcp.md).

## What Bridge governs

Everything above assumes Bridge governs **models and MCP**, the default. Admins can instead choose **models only**, where the model path applies and Bridge adds no Barndoor MCP server, or **MCP only**, where the tool path applies and models run on each person's own Claude or OpenAI sign-in, outside the LLM Gateway. See [Choose what Bridge governs](https://docs.barndoor.ai/how-tos/barndoor-bridge/set-up-your-organization.md#choose-what-bridge-governs), and [Govern tools only (MCP only)](https://docs.barndoor.ai/how-tos/barndoor-bridge/mcp-only.md) for what that mode guarantees.

## What every governed session has

The client's own built-in tools (file edits, shell, web fetch, browser) run on the laptop and don't go through Barndoor's tool policy. Admins decide which of them stay on with [client settings](https://docs.barndoor.ai/how-tos/barndoor-bridge/client-settings.md); by default every one is on, exactly as the client ships. Barndoor's own tools work the other way: they are denied by default, and a tool runs only when a [policy](https://docs.barndoor.ai/how-tos/manage-policies.md) grants it.

Whatever the client settings say, every session has these properties when Bridge governs **Models and MCP** or **Models only**. Under **MCP only**, models and the client's own tools aren't governed, and the list is different; see [Govern tools only (MCP only)](https://docs.barndoor.ai/how-tos/barndoor-bridge/mcp-only.md#what-it-guarantees).

* Model traffic goes only to the LLM Gateway.
* Barndoor is the only MCP server. MCP servers the person configured locally, and plugins' MCP servers, don't connect. Two exceptions: Claude Desktop still runs servers from plugins your administrators deploy to its organization plugins folder, and under **Models only**, where Bridge adds no MCP server, Claude Code can load a project's own `.mcp.json` servers once the person approves each one.
* The person can't widen the settings from the command line.
* Usage and audit records carry the client settings version: the one at launch, then the current one after each hourly renewal.
* Launches fail closed. If the installed client isn't a version Barndoor supports (unless you [allow unreviewed Codex Desktop versions](https://docs.barndoor.ai/how-tos/barndoor-bridge/client-settings.md#where-to-find-them)), a workspace carries a competing configuration, or the person is allowed no model the client can use, Bridge refuses to launch rather than start the client ungoverned.

## What Barndoor sees

| Barndoor sees | Where admins find it |
| - | - |
| Every model request and response from a governed session, with user, client, agent, model, tokens, and the client settings version | [LLM Usage Dashboard](https://docs.barndoor.ai/how-tos/read-llm-usage-dashboard.md), [audit log export](https://docs.barndoor.ai/how-tos/log-export.md) |
| Every call to a Barndoor tool, remote or local, with the server, the tool, and the policy decision | [MCP Usage Dashboard](https://docs.barndoor.ai/how-tos/read-mcp-usage-dashboard.md), audit log export |
| The input and output of local tool calls, when a data protection policy covers that server, for scanning | Data protection verdicts in the audit log |
| Calls to Barndoor's shell tool, with the tool and the policy decision (not the command text) | Audit log export |
| Launches that failed, with the reason and the user | **Client settings → Recent failed launches** |
| Bridge's own diagnostics, sent on a best-effort basis | Barndoor support |

Barndoor doesn't see what the client's own built-in tools do on the laptop (a native shell command, a native file edit, a native web fetch) except where it shows up in the conversation sent to the model. It doesn't read files on the laptop except through a tool call.

## What Bridge doesn't do

Bridge governs the sessions it launches. On its own it doesn't stop someone opening Claude, ChatGPT, or another AI tool directly. The [Codex policy](https://docs.barndoor.ai/how-tos/barndoor-bridge/provider-lockdown.md#codex-policy), on by default, makes Codex run only through Barndoor; [provider lockdown](https://docs.barndoor.ai/how-tos/barndoor-bridge/provider-lockdown.md#provider-lockdown), off by default, blocks direct access to provider APIs and does the same for Claude.

Sign-ins stored on the laptop, such as a Claude or ChatGPT sign-in, stay readable by the person who owns the account on that machine.

### What each control covers

Each row is something a person could otherwise do, and the control that covers it. The defaults are for **Models and MCP**.

| What a person might do | By default | Covered by |
| - | - | - |
| Add their own MCP server, or a plugin's, to a governed session | Blocked | Built in. Barndoor is the only MCP server. |
| Turn a restricted capability back on with a command-line option | Blocked | Built in. |
| Update or downgrade Bridge themselves on a managed machine | Blocked | Built in. IT's MDM decides the version. |
| Run Codex outside Bridge | Blocked | The [Codex policy](https://docs.barndoor.ai/how-tos/barndoor-bridge/provider-lockdown.md#codex-policy), installed unless IT opts out. |
| Run Claude outside Bridge, or call a provider API directly | Possible | [Provider lockdown](https://docs.barndoor.ai/how-tos/barndoor-bridge/provider-lockdown.md#provider-lockdown). |
| Change Claude Desktop's Barndoor configuration, for example with **Configure Third-Party Inference…** | Possible. Bridge writes a fresh one at each launch, and with provider lockdown the updater puts back the profile Claude Desktop uses when opened directly. | The [Claude Desktop managed profile](https://docs.barndoor.ai/how-tos/barndoor-bridge/provider-lockdown.md#claude-desktop-managed-profile). |
| Use an AI desktop app Bridge doesn't launch, such as Cursor | Possible | [Santa rules](https://docs.barndoor.ai/how-tos/barndoor-bridge/provider-lockdown.md#santa-rules). |
| Use a provider's website in a browser, such as `chatgpt.com` | Possible | Your web filtering. Bridge and provider lockdown don't block these sites. |
| Use the client's own shell, file, or web tools, which act outside your tool policy | Possible | [Client settings](https://docs.barndoor.ai/how-tos/barndoor-bridge/client-settings.md): route them **Through Barndoor** or turn them off. |
| Reach the internet from a native shell when Network is governed | Possible | Set Shell to **Through Barndoor** and govern the shell tool with policy, or block outbound traffic at the device or network layer. |
| Read credential files in their home folder, such as `~/.aws` | Possible | Not covered by Bridge. **Secrets: Scrubbed** removes environment variables only. |
| Stop Bridge starting at login | Blocked on a supervised Mac; possible on others | The managed login item profile, part of the [standard install](https://docs.barndoor.ai/how-tos/barndoor-bridge/install-with-mdm.md). Its lock needs a supervised Mac. |
| Under **MCP only**: remove the Barndoor server from their own Claude Desktop, or use their own MCP servers beside it | Possible | Not covered in Claude Desktop's chat. For Claude Code, the [managed Barndoor server](https://docs.barndoor.ai/how-tos/barndoor-bridge/mcp-only.md#managed-barndoor-server-for-claude-code). |

## Install and update integrity

* **macOS:** a signed, notarized installer (`Developer ID Installer: Barndoor AI, Inc.`, team `547SWHX99F`). The MDM updater accepts packages only from `downloads.barndoor.ai`, checks the SHA-256, the package signature, and Gatekeeper, and never downgrades.
* **Windows:** an MSI installer, with the MSI and both programs code-signed by Barndoor AI, Inc. See [Install on Windows](https://docs.barndoor.ai/how-tos/barndoor-bridge/install-on-windows.md).
* **Version control belongs to IT.** On a managed machine, Bridge refuses to update itself; IT's MDM decides the version.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.