Trip Planner — two agents, one enforced boundary

A trip planner built on the Model Context Protocol. You name a place and a number of days; a planner agent researches it and builds a real day-by-day itinerary in SQLite.

The interesting part isn't that it plans trips. It's how the two agents are kept apart.

The boundary

Local Scout  ──▶  tavily-mcp          can search the web, cannot touch the trip
Planner      ──▶  itinerary_server    can write the trip, has no internet

The Planner has no search tool. Its only route to the outside world is calling research_place — which is the Scout, attached to it as an ordinary tool via .as_tool(). The Scout, in turn, has no itinerary tools, so it structurally cannot write to the trip no matter what it decides to do.

Neither restriction is a prompt. Each agent is handed a different set of MCP servers, and that's the whole enforcement mechanism. Prompts can be talked around; a tool that isn't in the list cannot be called.

Related MCP server: wanderlog-mcp

Guard rails live in the server

add_activity rejects a day outside the trip's length:

> add_activity(trip_id=1, day=99, title="...")
Rejected: Day 99 is outside this 3-day trip - use a day from 1 to 3

That check is in db.py, not in the instructions. The agent gets a real error back and has to correct itself — which it does.

Setup

uv sync
cp .env.example .env      # add your keys

Two keys are needed: OPENAI_API_KEY (or another provider via PLANNER_MODEL) and TAVILY_API_KEY for the Scout's search. npx must be on your PATH — the Tavily MCP server runs through it.

Run it

uv run main.py              # create a trip, plan it, print it
uv run main.py --list       # list existing trips
uv run main.py --show 1     # re-print one itinerary

Point it at a different model with one env var:

PLANNER_MODEL=gpt-4o-mini

What a run looks like

Two days in Porto, planned from scratch:

=== Trip 2: Porto, Portugal (2 days) ===

Day 1
  - São Bento Railway Station
      Iconic tiled station hall and an easy first stop in the historic centre.
      Allow 15–20 minutes. No booking needed.
  - Sé do Porto (Porto Cathedral)
      Main old-city cathedral and a strong historic anchor. Allow 30–45 minutes.
  - Ribeira
      Porto's classic riverfront old quarter. Allow 1–2 hours.
  - Palácio da Bolsa
      Ornate 19th-century interiors; one of the best paid visits in the centre.
      Book ahead for guided tours.
  - Torre dos Clérigos
      The classic bell tower and city panorama. Booking ahead helps at busy times.

Day 2
  - Dom Luís I Bridge
  - Port wine cellars in Vila Nova de Gaia
  - Jardim do Morro
  - Jardins do Palácio de Cristal

Day 1 is the historic centre, Day 2 is Gaia and the west side — grouped so a day doesn't zig-zag across the city.

The MCP server

It speaks stdio, so the agent launches it as a subprocess. The Scout's server, tavily-mcp, is filtered down to just tavily_search — hiding crawl, map and extract keeps its tool list small and its behaviour predictable.

Layout

db.py                 SQLite schema and access — trips, activities, guard rails
itinerary_server.py   MCP server: the only way the trip gets written
planner.py            Scout on Tavily, wrapped via .as_tool(); Planner on itinerary
main.py               CLI

Four files, about 330 lines.

Notes

• The project sets link-mode = "copy" and passes UV_LINK_MODE=copy to every spawned MCP server, so it works on a cloud-synced drive (OneDrive, Dropbox), where uv's default hardlinking fails with os error 396.

• main.py reconfigures stdout to UTF-8 — accented place names would otherwise print as mojibake in the Windows console.

Verified

Both parts were tested for real, not assumed:

• The MCP server over stdio: tool listing, an itinerary read, and the day guard rail correctly rejecting day=99 on a 3-day trip.

• A full end-to-end run on a 2-day Porto trip — the Scout searched, the Planner wrote nine activities, grouped geographically, with booking notes on the two that need them.