Proxygen

Proxygen’s MCP server runs on localhost and is reached through a small bridge binary bundled with the app. It requires a per-session authentication token so only agents launched from your machine can use it.

The MCP is designed to prepare a workspace for you to inspect and continue from, not to replace the UI. When an agent creates an editor, a rewrite rule, a test, or a transform recipe, the result is visible in the corresponding Proxygen tool so you can review, adjust, and keep working with it.

- Ask an agent to explore traffic, summarize an API, or pick out interesting requests
- Have an agent draft rewrite rules or scripts based on a natural-language description
- Generate test configurations and payloads from example requests in history
- Extract values from responses across many messages
- Reproduce or modify requests in an editor
- Configure TLS interception, history filters, and recording settings

Enable the server fromPreferences → MCPby turning onEnable MCP server.

Everything a tool returns ends up in the agent’s context, and usually in whatever transcript logging that agent keeps. Proxygen therefore replaces credentials with a fingerprint before they leave the app.

The fingerprint is stable, so an agent can still answer the question it usually needs answered: two requests showing the same fingerprint are using the same token, and different fingerprints mean different tokens. It is a salted SHA-256 prefix and cannot be turned back into the original value. The salt is generated once per installation, so fingerprints are comparable within your machine but not across machines.

Searching and filtering match the redacted text as well. This matters more than it first appears: if a rule were matched against the real value, an agent could use it as a yes/no probe — a regex on a URL or header, repeated — and recover a secret one question at a time without the value ever being returned. Matching what the caller can already see closes that off, and searching for a fingerprint becomes the way to find every message using one credential.

Redaction is controlled byRedact secrets from tool callsinPreferences → MCP, which is on by default. Tools accept arevealSecretsparameter to return values in the clear, and it only works once you turn that preference off; while it is on the tool returns an error telling the agent to ask you first. This is deliberate — without the preference, an agent could simply opt itself out of redaction.

Message bodies are not redacted.A token in a JSON response body, such as an OAuthaccess_token, still reaches the agent in full. Redaction covers the structured places credentials live, not arbitrary payload content.

Bodies are usually the bulk of what an agent reads. Tools that return one acceptbodyMaxSizeto cap how much comes back,bodyOffsetto choose where the window starts, andbodyJsonPathto project a single value out of a JSON body inside Proxygen instead of returning the body at all.

bodyOffsetmakes it possible to read the middle of a large body without first pulling in everything before it. The response reportsbodyOffsetandbodyLengthfor the range actually returned, plusbodyTotalSizefor the whole body, so the agent can work through it in pieces. Windows that fall in the middle of a multi-byte character are nudged to the nearest character boundary, and the reported range reflects that. The sockets tools usemessageMaxSize,messageOffsetandmessageJsonPathfor the same thing.

OpenPreferences → MCP. TheAgent Setuppopup and thesettings.jsonbutton generate the configuration needed to register Proxygen with common agents.

The popup offers one-line registration commands for three agents. Selecting an entry copies the command to your clipboard; paste it into a terminal to register Proxygen with that agent.

- Claude Code

claude mcp add proxygen --transport stdio -- <path-to-proxygen-mcp>

gemini mcp add proxygen <path-to-proxygen-mcp>

codex mcp add proxygen -- <path-to-proxygen-mcp>

The command shown in Proxygen contains the absolute path to the bundledproxygen-mcpbridge binary.

For agents that are configured through a JSON settings file rather than a CLI command, use thesettings.jsonbutton. It copies a ready-to-paste snippet to your clipboard:

{ "mcpServers": { "proxygen": { "command": "/path/to/Proxygen.app/Contents/MacOS/proxygen-mcp" } } }

Merge this into your agent’s MCP settings file.

Proxygen must be running and the MCP server must be enabled for agents to connect.

Once Proxygen is registered with an agent, you can drive it with natural language. The examples below show the kinds of requests the tools are designed to handle — the agent picks the right tools and parameters and leaves results visible in Proxygen for you to continue from.

- “Give me a summary of the last hour of history — which hosts, status codes, and content types are present.”
- “Build a sitemap of everything underapi.example.comand highlight endpoints that only appear once.”
- “Look through messages 400–650 and flag any request that took longer than 2 seconds, adding a note with the elapsed time.”
- “Find every 5xx response in the last 200 messages, group them by host, and show me the first example of each.”