AI Bridge

Bridge is a smart proxy for AI. It acts as a man-in-the-middle between your users' coding agents / IDEs and providers like OpenAI and Anthropic. By intercepting all the AI traffic between these clients and the upstream APIs, Bridge can record user prompts, token usage, and tool invocations.

Bridge solves 3 key problems:

1. Centralized authn/z management: no more issuing & managing API tokens for OpenAI/Anthropic usage. Users use their Coder session or API tokens to authenticate with coderd (Coder control plane), and coderd securely communicates with the upstream APIs on their behalf. Use a single key for all users.
2. Auditing and attribution: all interactions with AI services, whether autonomous or human-initiated, will be audited and attributed back to a user.
3. Centralized MCP administration: define a set of approved MCP servers and tools which your users may use, and prevent users from using their own.

When to use AI Bridge

As the library of LLMs and their associated tools grow, administrators are pressured to provide auditing, measure adoption, provide tools through MCP, and track token spend. Disparate SAAS platforms provide some of these for some tools, but there is no centralized, secure solution for these challenges.

If you are an administrator or devops leader looking to:

• Measure AI tooling adoption across teams or projects
• Provide an LLM audit trail to security administrators
• Manage token spend in a central dashboard
• Investigate opportunities for AI automation
• Uncover the high-leverage use cases from experienced engineers

We advise trying Bridge as self-hosted proxy to monitor LLM usage agnostically across AI powered IDEs like Cursor and headless agents like Claude Code.

Setup

Bridge runs inside the Coder control plane, requiring no separate compute to deploy or scale. Once enabled, coderd hosts the bridge in-memory and brokers traffic to your configured AI providers on behalf of authenticated users.

Required:

1. A premium licensed Coder deployment
2. Feature must be enabled using the server flag
3. One or more provider API keys must be configured

Activation

To enable this feature, activate the aibridge experiment using an environment variable or a CLI flag. Additionally, you will need to enable Bridge explicitly:

If you have other experiments enabled, separate them by commas.

Providers

Bridge currently supports OpenAI and Anthropic APIs.

API Key:

The single key used to authenticate all requests from Bridge to OpenAI/Anthropic APIs.

• CODER_AIBRIDGE_OPENAI_KEY or --aibridge-openai-key
• CODER_AIBRIDGE_ANTHROPIC_KEY or --aibridge-anthropic-key

Base URL:

The API to which Bridge will relay requests.

• CODER_AIBRIDGE_OPENAI_BASE_URL or --aibridge-openai-base-url, defaults to https://api.openai.com/v1/
• CODER_AIBRIDGE_ANTHROPIC_BASE_URL or --aibridge-anthropic-base-url, defaults to https://api.anthropic.com/

Bridge is compatible with Google Vertex AI, AWS Bedrock, and other LLM brokers. You may specify the base URL(s) above to the appropriate API endpoint for your provider.

---

Collected Data

Bridge collects:

• The last user prompt of each request
• All token usage (associated with each prompt)
• Every tool invocation

All of these records are associated to an "interception" record, which maps 1:1 with requests received from clients but may involve several interactions with upstream providers. Interceptions are associated with a Coder identity, allowing you to map consumption and cost with teams or individuals in your organization:

These logs can be used to determine usage patterns, track costs, and evaluate tooling adoption.

This data is currently accessible through the API and CLI (experimental), which we advise administrators export to their observability platform of choice. We've configured a Grafana dashboard to display Claude Code usage internally which can be imported as a starting point for your tooling adoption metrics.

We provide an example Grafana dashboard that you can import as a starting point for your tooling adoption metrics. See here.

Implementation Details

coderd runs an in-memory instance of aibridged, whose logic is mostly contained in https://github.com/coder/aibridge. In future releases we will support running external instances for higher throughput and complete memory isolation from coderd.

See a diagram of how Bridge interception works

sequenceDiagram actor User participant Client participant Bridge User->>Client: Issues prompt activate Client Note over User, Client: Coder session key used<br>as AI token Client-->>Bridge: Sends request activate Bridge Note over Client, Bridge: Coder session key <br>passed along Note over Bridge: Authenticate Note over Bridge: Parse request alt Rejected Bridge-->>Client: Send response Client->>User: Display response end Note over Bridge: If first request, establish <br>connection(s) with MCP server(s)<br>and list tools Note over Bridge: Inject MCP tools Bridge-->>AIProvider: Send modified request activate AIProvider AIProvider-->>Bridge: Send response Note over Client: Client is unaware of injected<br>tools and invocations,<br>just receives one long response alt Has injected tool calls loop Note over Bridge: Invoke injected tool Bridge-->>AIProvider: Send tool result AIProvider-->>Bridge: Send response end end deactivate AIProvider Bridge-->>Client: Relay response deactivate Bridge Client->>User: Display response deactivate Client

MCP

Model Context Protocol (MCP) is a mechanism for connecting AI applications to external systems.

Bridge can connect to MCP servers and inject tools automatically, enabling you to centrally manage the list of tools you wish to grant your users.

Bridge makes use of External Auth applications, as they define OAuth2 connections to upstream services. If your External Auth application hosts a remote MCP server, you can configure Bridge to connect to it, retrieve its tools and inject them into requests automatically - all while using each individual user's access token.

For example, GitHub has a remote MCP server and we can use it as follows.

See the diagram in Implementation Details for more information.

You can also control which tools are injected by using an allow and/or a deny regular expression on the tool names:

In the above example, all tools containing _gist in their name will be allowed, but create_gist is denied.

The logic works as follows:

• If neither the allow/deny patterns are defined, all tools will be injected.
• The deny pattern takes precedence.
• If only a deny pattern is defined, all tools are injected except those explicitly denied.

In the above example, if you prompted your AI model with "list your available github tools by name", it would reply something like:

Certainly! Here are the GitHub-related tools that I have available:

1. bmcp_github_update_gist
2. bmcp_github_list_gists

Bridge marks automatically injected tools with a prefix bmcp_ ("bridged MCP"). It also namespaces all tool names by the ID of their associated External Auth application (in this case github).

Tool Injection

If a model decides to invoke a tool and it has a bmcp_ suffix and Bridge has a connection with the related MCP server, it will invoke the tool. The tool result will be passed back to the upstream AI provider, and this will loop until the model has all of its required data. These inner loops are not relayed back to the client; all it seems is the result of this loop. See Implementation Details.

In contrast, tools which are defined by the client (i.e. the Bash tool defined by Claude Code) cannot be invoked by Bridge, and the tool call from the model will be relayed to the client, after which it will invoke the tool.

If you have the oauth2 and mcp-server-http experiments enabled, Coder's own internal MCP tools will be injected automatically.

Troubleshooting

• Too many tools: should you receive an error like Invalid 'tools': array too long. Expected an array with maximum length 128, but got an array with length 132 instead, you can reduce the number by filtering out tools using the allow/deny patterns documented in the MCP section.

• Coder MCP tools not being injected: in order for Coder MCP tools to be injected, the internal MCP server needs to be active. Follow the instructions in the MCP Server page to enable it.

• External Auth tools not being injected: this is generally due to the requesting user not being authenticated against the External Auth app; when this is the case, no attempt is made to connect to the MCP server.

Known Issues / Limitations

• Codex CLI currently does not work with Bridge due to a JSON marshaling issue: https://github.com/coder/aibridge/issues/19
• Claude Code web searches do not report correctly: https://github.com/coder/aibridge/issues/11

Supported APIs

API support is broken down into two categories:

• Intercepted: requests are intercepted, audited, and augmented - full Bridge functionality
• Passthrough: requests are proxied directly to the upstream, no auditing or augmentation takes place

Where relevant, both streaming and non-streaming requests are supported.

OpenAI

Intercepted:

• /v1/chat/completions

Passthrough:

• /v1/models(/*)
• /v1/responses (Interception support coming in Beta)

Anthropic

Intercepted:

• /v1/messages

Passthrough:

• /v1/models(/*)

Troubleshooting

To report a bug, file a feature request, or view a list of known issues, please visit our GitHub repository for Bridge. If you encounter issues with Bridge during early access, please reach out to us via Discord.