# Giving an AI assistant access to captured HTTP traffic with mitmAny

> Source: <https://dev.to/dai_wang_ffb06b747d1aca76/giving-an-ai-assistant-access-to-captured-http-traffic-with-mitmany-4afk>
> Published: 2026-11-21 07:32:18+00:00

**Key takeaway:** mitmAny exposes captured HTTP/HTTPS flows to an MCP client read-only, letting the assistant cite request IDs and flag missing evidence rather than infer uncaptured behavior.

Disclosure: I maintain mitmAny, the open-source tool described here.

An HTTP trace can tell you what an application actually sent. Turning that trace into an explanation still takes work: identify the relevant host, compare requests, follow pagination, and separate observed behavior from assumptions.

mitmAny connects that investigation to an MCP-compatible AI client. It is a free, MIT-licensed HTTP/HTTPS inspector for macOS and Windows, with an Android USB capture workflow. The AI runs in your connected client; mitmAny supplies captured evidence.

*Illustrative product screenshot with synthetic data, not results from a live investigation.*

Start with a website or application you own or are authorized to inspect. Download mitmAny from the project's Releases page. For HTTPS inspection, install and trust its local CA certificate. Start capture manually, then perform a small sequence such as searching and loading the next page of results.

In Settings → Codex, install the bundled analysis skill and register the MCP connection. Reconnect your client or open a new session before requesting an analysis.

Keep the question narrow:

Analyze captured traffic for my API host. Compare the initial search and the next page of results. Identify the observed pagination parameters and response fields. Cite request IDs for each conclusion and explicitly list missing evidence. Do not infer behavior that was not captured.

Replace “my API host” with the exact hostname. A narrow host scope helps keep unrelated background traffic out of the investigation.

The useful output is an evidence-backed API outline: methods, paths, observed parameters, response shapes, and uncertainties. Ask the assistant to distinguish a field it saw from a constraint it has actually verified. Seeing a parameter once does not prove that it is required; seeing one successful response does not establish every possible schema.

Instead of repeatedly copying request bodies into a chat window, the client can summarize and search captured flows, inspect headers and bounded bodies, and wait for new traffic. You can perform another action in the application and ask it to compare the additional evidence.

The connection is read-only for network traffic. It does not replay requests, change proxy configuration, or bypass certificate pinning. Observation bookmarks organize an investigation but do not preserve records that have been evicted.

Clients must use the proxy and trust its CA. Some Android apps require a debug trust configuration. There is no QUIC/HTTP3 interception or WebSocket message inspector. Capture and retention limits mean a body can be incomplete or an older request unavailable; the analysis should say so.

Traffic is stored locally by mitmAny, but a connected AI client may send retrieved data to its model provider. Review what you capture and share. Automated redaction is not a guarantee that all secrets are removed.

All mitmAny features are free. Fees from your chosen AI client or model provider are separate.

Source and downloads: [https://github.com/wonderfulsuccess/mitmAny](https://github.com/wonderfulsuccess/mitmAny)

If you try it, a useful bug report includes your OS, client, setup step, and what happened, with sensitive data removed. Feedback on the first successful capture and MCP connection is especially useful.
