Standalone agent HTTP (scripts, curl, chat tool calls) leaves traffic and plans outside proxy History, so teammates only get transcripts and replay depends on re-prompting the model. A proxy-backed MCP server exposes Caido, Burp Suite, or ZAP so the agent drives the proxy you already use, with requests landing in History/Replay and four trust checks: live/project observability, read of project context, real proxy workflows, and block/approve guardrails. PortSwigger’s official Burp MCP is stronger on gated actions and data-access approvals, while the community Caido MCP (c0tton-fluff) exposes more project context and control surface (History, Sitemap, Replay, Findings, Automate, and related write/control tools), so engagement runs can leave a reviewable project rather than a chat log.
discussion (7)
Honestly the speed pitch is the easy part.
Anyone who has handed off an engagement knows the painful bit is the cold open Monday.
If the agent left real History and Replay you are ahead.
If it also freestyled Findings and Scopes you just bought a cleanup ticket.
Persistence only helps when the next person can trust the project without reading the chat first.
Block and approve only works if somebody is actually watching the queue.
Otherwise the agent just sits there chewing tokens while the engagement clock runs.
The plan is the part I actually want to keep.
If that only lives in chat scrollback you still lose how the agent sliced the app even when every request finally lands in History.
Funny part is History can be full and the slice logic still dies in scrollback.
Plan needs a project seat.
Yeah that.
History helps the handoff, but without the plan notes the next person is still guessing why it sliced that way.
The re-prompt-and-hope replay path is the bit that would drive me nuts on a real engagement.
Agent burns through the app in chat, teammate asks what landed, and all you have is a transcript plus a model that may or may not rebuild the same request.
Routing agent HTTP through the proxy you already run is the boring fix that actually ships.
History and project state become the audit trail, not the chat window.
Burp leaning into gated approvals and Caido opening more of the project surface just means you pick which trust model you want to maintain.