In-App Assistant

The assistant is a chat panel in the API lab → Assistant tab. You bring the model: it talks to any OpenAI-compatible endpoint (a hosted API, a local gateway, or a local server) using tool calling, and it drives the same tools the MCP server exposes.

Configure the model

Open Settings → AI model and set:

• Base URL — the OpenAI-compatible root, e.g. https://api.openai.com/v1. The assistant posts to {base}/chat/completions with stream: true.
• API key — sent as a Bearer token. The field is masked unless Reveal sensitive values is on (/docs/en/security/secrets/).
• Model — the model id, e.g. gpt-4o-mini.

Settings persist in the browser’s local storage (traffic-jam:ai-settings). Nothing leaves your machine except the chat requests to the endpoint you configure.

Use it

1. Optionally pick a session under Attach current session — its id and name are appended to your message as context.
2. Ask a question. The assistant orients with list_sessions / get_session, then drills in with the read and analysis tools, and only replays when it needs live evidence.
3. Watch the tool cards inline: each shows the tool name, arguments, and (expandable) result. Mutating tools are badged live traffic.

Confirming live-traffic tools

Five tools send real traffic: replay, replay_batch, replay_minimize, mutation_probe, and run_recipe. With Confirm before live-traffic tools enabled (the default), the assistant pauses and shows Run / Deny on each one before it executes. Deny returns a “denied by the user” result to the model so it can adjust.

What it can do

The assistant has 25 tools:

• Read — list_projects, list_sessions, get_session (digest), get_request, search_requests, export_har, list_replay_history, list_client_certs, list_fingerprint_profiles, list_transports, list_decode_scripts.
• Replay — replay, replay_batch, replay_minimize.
• Analysis — decode_body, conformance_diff, token_lifecycle, signature_candidates, signature_crack, mutation_probe, run_recipe, evaluate_computed_template.
• Authoring — save_decode_script, save_decode_pipeline, save_recipe (persist parser artifacts to the project; no live traffic).

The analysis tools are the same ones behind the Collector workspace (/docs/en/collector/): conformance diff, signature cracking, mutation probes, and token lifecycle.

[!NOTE] The system prompt instructs the model to keep secrets out of its prose — it references credentials by name, not value — and to prefer the minimum necessary replays. Replay responses still redact sensitive headers unless a tool passes includeSensitiveValues.