Anchorline MCP Server

Plan around what can't move.

Quickstart · Use it anywhere · How the engine works · Core workflow · Docs · Contributing

Anchorline is an open-source holiday planner for trips with things that can't move: a wedding on the 8th, three nights in Chiang Mai at least, a week on the beach, home by Sunday 23:00. Give it dates, interests and commitments. It recommends one route with its reasons and shows exactly what each edit would change before anything changes. It never invents a booking, a fare or a timetable. It runs in your browser with no account, in your terminal, and inside Claude and ChatGPT through an Agent Skill and an MCP server, all on the same tested engine.

Why Anchorline

• Commitments survive edits. Fixed events, locked dates, minimum stays, consecutive beach nights and the home-by deadline are hard rules. A change that would break one is blocked and explained, with concrete ways to relax it. Nothing is unlocked for you.

• Every change is previewed, and undo is real. You see the stays, nights, journeys, activities, budget and checks that would change, then choose to apply or keep. Undo and redo are new revisions that survive a reload. A second tab can't silently overwrite the first.

• Nothing is invented. Planning estimates, sourced facts, quotes you recorded and bookings you made are kept apart. Travel times are ranges, not made-up departures. An unknown price is unknown, not zero. Replacing a city invalidates the old route's evidence.

Related MCP server: wanderlog-mcp

Features

Use it anywhere

On Claude Code 2.1.275 or later, the plugin also installs in one step: /plugin install anchorline --marketplace IlgazTheFirst/anchorline.

Quickstart

Browser

Requires Node.js 20.19 or newer.

git clone https://github.com/IlgazTheFirst/anchorline
cd anchorline
npm ci
npm run dev            # http://localhost:5173

Choose Start a new trip, or Open the fictional demo. npm run build && npm run preview serves the offline-capable production build on http://localhost:4173.

Terminal

The CLI is one bundled file with no dependencies. Run it from a clone with node skills/anchorline/scripts/anchorline.mjs, or put it on your PATH with npm link.

anchorline places
anchorline plan --from London --tz Europe/London --depart 2031-02-01 --nights 14 \
  --home-by 2031-02-16T23:00 --must "chiang mai:3" --beach-nights 5 --out trip.json
anchorline check trip.json
anchorline preview trip.json change.json      # shows the diff, writes nothing
anchorline apply trip.json change.json --out trip-2.json
anchorline export trip.json --share --omit-dates

Trip from London: 14 nights, 3 stops: Bangkok 4 · Chiang Mai 5 · Phuket 5.
Style: A bit of everything, balanced pace.
...
Journeys (no times or fares yet)
  London → Bangkok: flight
  Bangkok → Chiang Mai: flight, about 1 h 5 min–1 h 25 min (planning estimate)
  Chiang Mai → Phuket: flight, about 2 h–2 h 20 min (planning estimate)
  Phuket → London: flight
...
Nothing here is booked. Travel times are planning estimates from documented routes, not timetables or fares. Prices are unknown unless you add them.

plan also reads a JSON trip spec from a file or stdin, and every command takes --json. The spec and command formats are in the skill reference and commands.md. Changes are untrusted suggestions unless you pass --as-user. Exit codes: 0 ok, 1 bad input, 2 infeasible plan or a broken commitment under check --strict.

Claude Code

/plugin marketplace add IlgazTheFirst/anchorline
/plugin install anchorline@anchorline

Then ask: "Plan 14 nights from London leaving 1 Feb 2031, at least 3 nights in Chiang Mai, 5 beach nights in a row, home by 16 Feb 23:00." The skill drives the CLI. The MCP tools plan_trip, check_trip, preview_change, share_export and list_places are available too. See integrations/claude-code.md.

Claude Desktop and ChatGPT

Both use the MCP server, anchorline mcp: over stdio for Claude Desktop, and over streamable HTTP with --http <port> for ChatGPT, which needs a public HTTPS URL that you host. See claude-desktop.md and chatgpt.md, which includes a Custom GPT instruction set.

How the engine works

flowchart LR
  A["Trip spec or edit"] --> B["Commitments become pins and minimums"]
  B --> C["Beam search over documented routes"]
  C --> D["Exact night allocation (dynamic programme)"]
  D --> E["Score: enjoyment + variety - moves - transfers"]
  E --> F["One plan, reasons, open questions, alternatives"]
  F --> G{"Preview"}
  G -- ready --> H["Apply: atomic commit, undo/redo"]
  G -- blocked --> I["Explain the broken commitment and relaxations"]

Fixed events, locked stays, your own places and protected rest become date pins. A beam search orders stops over a graph of documented connections. A dynamic programme then splits every night exactly, honouring pins, consecutive beach runs and full-day counts. Diminishing returns per night, a variety bonus, and penalties for moves, travel time and airport detours pick one route. The result, like every manual edit, goes through the same validated preview → apply pipeline. Same inputs and catalogue version, same plan. Every constant is in docs/ALGORITHM.md.

Core workflow

Create a trip → edit a stay → preview the impact → apply (saved) → reload → undo. It is automated in tests/e2e/core-flow.spec.ts at desktop (1440×900) and mobile (390×844) sizes, with made-up inputs.

Data and privacy

• Trips stay on your device. The web app makes no network requests beyond its own files; the CLI and MCP server make none at all.

• Every accepted change is saved atomically with its undo history. Imports are size-limited and fully validated before anything is written. Only http(s) links are kept, and text is never rendered as HTML.

• A shareable copy is built from an allowlist: traveller names and contacts, booking references, private notes and history never leave.

See SECURITY.md for the full data boundaries.

Coverage and limitations

• Automatic planning covers the Thailand & the Malay Peninsula pack: Bangkok, Ayutthaya, Chiang Mai, Ko Samui, Ko Pha-ngan, Ko Tao, Khao Sok, Phuket, Krabi (Ao Nang), Railay, Ko Lanta, Kuala Lumpur, George Town, Langkawi and Singapore. Sources were checked on 31 October 2026. Anywhere else, add your own places and plan by hand with all checks, budget and export features.

• No timetables, fares, live availability or booking. No maps. Season notes are approximate guidance.

Roadmap

• More destination packs (contributions welcome, see docs/DATA_PACKS.md)

• Publish the anchorline package to npm, and an optional hosted demo

• An optional server-side adapter for dated timetables and quotes, still entering through the preview pipeline as untrusted input

• Multi-region trips across packs

Contributing

Bug fixes, presets and destination packs are welcome. See CONTRIBUTING.md and the code of conduct. Please never include anyone's real itinerary, booking references or contact details in issues, fixtures or screenshots.

npm ci && npm run check && npm run test:e2e

License

MIT. Data, dependency and font notices are in NOTICE.md; the brand guide is in docs/brand.