trvl

Ask your AI assistant to plan a real trip, and it actually can. trvl gives Claude, Cursor, Windsurf, Codex, or any MCP-compatible client one smart tool — the travel router — with live access to flights, hotels, rental cars, trains, buses, ferries, price alerts, award sweet spots, weather, and destination intel. Free for noncommercial use, no personal API keys for default search, no signup. One binary.

You: I have €300 and a free weekend. Surprise me.

Illustrative response shape (prices are examples): Dubrovnik, Croatia 🇭🇷 — ✈️ Ryanair HEL→DBV €167 RT · 🏨 Old Town Studios 4.6★ €84 · 🌡️ 26°C, sunny. 📊 Naive €350 → optimized €251 → saved €99 (28%) by flying Friday and splitting airlines.

▶ Try it live, no install: socialistic.ai/trvl-travel-mcp (community-hosted).

trvl is a single binary that runs on your machine. Your AI client talks to it over MCP; it queries flight, hotel, transport, weather, and places sources, then merges, de-duplicates and optimizes the results before handing back one answer. Most sources are free public endpoints, so there is nothing to sign up for. The ones behind bot protection may reuse the browser session you already have, which is the one thing worth reading about before you install: see What trvl reads, and what it keeps.

trvl speaks the two latest MCP revisions: 2026-07-28 and 2025-11-25. An initialize with no _meta is answered as 2025-11-25, whether the handshake names 2025-11-25 or 2026-07-28, and that answer includes ping and resources/subscribe. _meta of 2026-07-28 on an initialize selects the 2026 shape. A client that sends 2026-07-28 in _meta["io.modelcontextprotocol/protocolVersion"] on each request gets server/discover, resultType, cache hints on the list and read endpoints, and subscriptions/listen instead of resource subscribe. Any other _meta protocol version returns JSON-RPC -32022 with data.supported and data.requested. On HTTP the MCP-Protocol-Version header has to match that field; a 2026 header alone is -32020. A 2026 request must also send Mcp-Method (and Mcp-Name for tools/call, resources/read, and prompts/get).

• Whole journey, door to door. It plans the entire trip across modes — home to airport, flight, arrival transfer, hotel, onward train — and prices each leg in its real mode. Most tools stop at one flight, one hotel.
• No personal API key for default search. There is no Amadeus key to apply for, subscription, or per-call bill on the default paths. Optional providers switch on when you supply their credentials; none is required for the default flight, hotel, or ground searches.
• Your assistant, your machine. One local binary, any MCP client, not locked to a vendor. Searching sends the query to the providers being searched, the same as any travel site would: route, dates and traveller count go to Google, Kiwi, Booking and the rest. What trvl keeps for itself stays on your machine, apart from any webhook you configure yourself. A released build sends one daily heartbeat to https://telemetry.revaluator.ai/v1/heartbeat, and you can switch that off. TRVL_TELEMETRY_ENDPOINT replaces that URL. Both are spelled out below.
• It optimizes, not just lists. Shift-day pricing, split-airline routing, hidden-city checks, award sweet spots, round-trip fares. It hands back the cheaper option and shows what it saved.
• It is honest when a source fails. Typed statuses and labelled estimates, never an empty result dressed up as "nothing found."

Full head-to-head against Google Flights, KAYAK, Skyscanner, Kiwi, and other travel MCPs: docs/COMPARISON.md.

Yes if you already plan trips with an AI assistant and want it to search real flights, hotels, trains, buses, ferries, and transfers instead of guessing — or if you're building an app that needs travel intent without a paid travel API.

Probably not if you just want to book on a website (use Google Flights), or you want a hosted product with an account and dashboard. trvl is a tool you run, not a service you log into.

Full positioning: docs/POSITIONING.md.

Let your AI do it — paste into Claude Code, Cursor, Windsurf, or Codex:

Read https://raw.githubusercontent.com/MikkoParkkola/trvl/main/AGENTS.md and set up trvl

It installs the binary, wires the MCP server, installs the skill, and verifies everything. Under a minute.

Or by hand:

brew install MikkoParkkola/tap/trvl   # install
trvl mcp install                       # auto-detects your AI client

Restart your client. trvl mcp install --client <name> targets a specific one (10 supported: Claude Desktop/Code, Cursor, Windsurf, Codex, VS Code, Zed, Gemini, Amazon Q, LM Studio).

# Direct binary (no Homebrew)
TRVL_VERSION="$(curl -fsSLI -o /dev/null -w '%{url_effective}' https://github.com/MikkoParkkola/trvl/releases/latest | sed 's#.*/v##')"
curl -fsSL "https://github.com/MikkoParkkola/trvl/releases/download/v${TRVL_VERSION}/trvl_${TRVL_VERSION}_$(uname -s | tr '[:upper:]' '[:lower:]')_$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/').tar.gz" | tar xz -C /usr/local/bin trvl

# Go
go install github.com/MikkoParkkola/trvl/cmd/trvl@latest

# Docker
docker run --rm ghcr.io/mikkoparkkola/trvl flights HEL NRT 2027-06-15

# Build from source
git clone https://github.com/MikkoParkkola/trvl.git && cd trvl && make build

# Claude Code CLI
claude mcp add trvl --transport stdio -- trvl mcp

# Manual JSON (Claude Desktop, Cursor, Windsurf, etc.)
# { "mcpServers": { "trvl": { "command": "trvl", "args": ["mcp"] } } }

Paste any of these to your assistant once trvl is wired:

Find the cheapest realistic trip from Helsinki for the long weekend of July 1–5, nonstop, hotel near the center.
Compare award sweet spots HEL→LHR business on Aug 15 with 80k Amex MR + 20k Virgin points.
Create a mistake-fare watch for HEL→BCN, July 1–8, and alert me below €90.

More starter prompts and what good answers look like: docs/DEMO.md.

An AI agent acts on trvl's output without a human checking every result, so the bar is correctness, not just coverage.

• It tells you when it can't. A blocked or rate-limited provider returns a typed status (AKAMAI_BLOCK, RATE_LIMITED, BOOKING_COOKIES_MISSING) with a fix hint, instead of a fake "nothing found." Estimated values are labelled; currency-mismatched totals are skipped, not faked.
• Tested more than it is written. More test code than source, race-checked on macOS, Linux, and Windows. A smoke gate runs the packaged binary before any release — a build that doesn't run can't ship.
• It degrades gracefully. Providers run concurrently with per-provider timeouts; one source failing returns partial results instead of aborting the search.
• It's observable. trvl status (or the local /dashboard in HTTP mode) shows per-provider success rate, latency, freshness, and circuit-breaker state.

On hotel prices specifically. Hotel metasearch exposes list-level rates first; some are real, some only firm up after the property detail page reveals the room/tax/cancellation matrix. So trvl separates discovery (search_hotels — fast, lead-in prices) from decisions (search_accommodations — verifies room-level offers before ranking) and drill-down (search_hotels_with_details, hotel_rooms). It provides booking links for manual handoff but never books, holds, or guarantees a rate. Detail: docs/PROVIDERS.md.

Two things happen without you asking for them. Neither is obvious, so both are stated here rather than left to a linked page.

It reads your browser's cookies, automatically. Hotel and rail sites put bot protection in front of their search APIs, and trvl gets past it by reusing the browser session you already have — that is why searches work with no API key. The reads start when trvl launches, before any search. No flag turns them on. On macOS, browser cookie stores are encrypted, so reading them means Keychain access and you should expect a Keychain prompt. What is read is your own session cookies for the site being searched, and they go into the request to that same site. Every provider definition capable of receiving cookies is reviewed in source and embedded in the binary; runtime provider JSON is not executable. The room lookup's caller-supplied URL is checked against Booking.com before cookies are attached, and a test fails if that stops being true. If a site redirects trvl to a different host, the cookies do not follow: Go's HTTP client refuses to carry them across a change of host. That check compares hosts and not schemes, so a site redirecting its own https:// address to plain http:// would keep them — a site downgrading its own traffic to cleartext is the one case that would put a session on the wire unencrypted. trvl reports your cookies to no endpoint of its own.

Optional provider definitions are source-only. trvl providers enable <id> can enable only a reviewed definition shipped in the current binary. The runtime state file records consent, enabled state, and health; it cannot replace endpoints, headers, authentication, request templates, or response mappings. Older files under ~/.trvl/providers are left in place for rollback or manual migration, but trvl does not load them. New definitions must arrive through a reviewed source change (or a user-maintained fork).

It keeps working state under ~/.trvl: saved trips, preferences and traveller profile, price watches, search history, cached cookies and provider tokens, a provider health log, a Google rate-limit cooldown, upgrade and provider self-heal bookkeeping, and a random install id. That state is local. A released build sends the install id, and nothing else from this directory, at most once a day to https://telemetry.revaluator.ai/v1/heartbeat. TRVL_TELEMETRY_ENDPOINT replaces that URL. TRVL_NO_TELEMETRY=1 stops the send. A price watch you give a webhook URL to POSTs that watch's route and price data to the address you supplied, which is the point of a webhook.

JSON state files are written to a temp file and then renamed over the target, so a crash cannot leave a half-written JSON document behind. Price watches and their history use transactional watch.db storage; the first migration backs up the legacy JSON before committing the database. Temp-file replacement can still leave orphaned temp files from interrupted writes, each a full copy of the JSON file it was about to replace. trvl does not delete them on its own, because the orphan is occasionally the only surviving copy of the target. trvl tempfiles reports what is there with sizes and ages; trvl tempfiles --delete removes only the ones whose writing process is provably gone.

You can decline either behaviour:

export TRVL_NO_BROWSER_COOKIES=1   # never read your browsers or the sessions in them, and never open a window in your real browser
export TRVL_NO_TIER2_CDP=1         # never start a headless browser of its own

Both controls trade coverage for isolation. Cookie access is enabled by default because protected hotel and rail sources rely on existing sessions; turning it off can materially reduce results. The headless fallback starts an installed Chrome, Brave, or Edge with an empty profile only when a challenged source needs fresh cookies. It does not read your existing browser sessions, show a window, or take focus. Turning it off can leave a challenged source empty. A visible-browser recovery path exists for a small number of providers, but it is separate, provider-specific, and asks before opening; disabling the headless fallback does not enable it automatically.

trvl honours HTTP_PROXY, HTTPS_PROXY and NO_PROXY. It checks and pins both hops — the proxy and the destination — so a redirect or a changed DNS answer cannot move the connection somewhere the URL never named.

By default trvl refuses to connect to private, loopback and link-local addresses, which is what keeps a hostile redirect away from your internal network and from cloud metadata endpoints. A corporate proxy is almost always on a private address, so reaching one needs an explicit opt-in:

export TRVL_ALLOW_PRIVATE_PROXY=1   # the PROXY may be on a private address

That relaxes the proxy hop only. Destinations are still refused on private addresses. The broader TRVL_ALLOW_LOCAL_PROVIDERS=1 allows both, and exists for pointing trvl at a mock provider on your own machine — do not reach for it just to use a proxy, because it also switches off the destination guard.

Authenticated proxies (http://user:pass@host) are not supported.

trvl calls travel sites directly, and many of them treat cloud and datacenter IP addresses as bots. From a server, the same search can return fewer options than from home. trvl reports this instead of hiding it. Each provider in provider_statuses carries a reason when it did not answer: blocked (refused, usually HTTP 403 or a bot check), rate_limited, unavailable (the provider's own 5xx error), dns, network or timeout. A search where some providers failed is marked partial, never "nothing found".

One user measured this from a Kubernetes cluster in October 2026: SNCF, Trainline and Rome2Rio answered 403, Wizz Air answered 503, and Italo's host did not resolve on some DNS resolvers. Only the 403s point to IP blocking; a 503 or a DNS failure can have other causes. A home or office connection, or the proxy support above, usually improves coverage but cannot guarantee it.

Price watching is experimental. It works, and it keeps data — your watches and their whole price history live under ~/.trvl. Treat that history as something you could lose. The store is backed up before any migration and the legacy files are kept afterwards, so a bad outcome is recoverable; but the feature is younger than the rest of trvl and is being changed more often.

Upgrading to 1.21.0 can delete price history. The first run converts the store to a transactional database, and if you have lowered TRVL_WATCH_MAX_POINTS_PER_WATCH it now honours that limit during the conversion — earlier versions ignored it there while applying it everywhere else, so a lowered limit was quietly not in force. Points above your limit are removed. Run trvl watch migrate --dry-run first to see the real number: the preview used to under-report it, which is the reason this warning exists.

Watch price history is capped, or it grows without bound: one real store reached 320,028 points in 41MB, which cost every running trvl process about 686MB of memory. Three limits bound it, and all three can be changed.

export TRVL_WATCH_MAX_POINTS_PER_WATCH=1000   # points kept per watch
export TRVL_WATCH_MAX_POINTS_TOTAL=50000      # points kept across all watches
export TRVL_WATCH_ROUTE_TTL_DAYS=90           # how long a dateless route watch keeps being checked

Those are the defaults, and the cost of moving each one:

An unusable value is refused, not adjusted: trvl will not start against TRVL_WATCH_MAX_POINTS_TOTAL=0, rather than quietly using the default and leaving you to believe your setting took effect.

To see whether any limit is actually binding on your store, run trvl watch migrate --dry-run. It reports the size on disk, the spread of points per watch, and how close you are to each cap, without changing anything.

Full mechanism — what is read at which point, the exact difference between those two switches, the headless-browser fallback, and the AF-KLM credential rules: docs/ARCHITECTURE.md.

One command publishes deliberately, because that is what you ran it for: the calendar helper writes an event to your Google calendar. trvl share does not — it prints the trip card, or copies it to your clipboard, and you decide who receives it. It used to upload the card as a public GitHub gist; that was removed in #527, because a card carries destinations and dates, and those together say when your home is empty.

Every source trvl uses by default is free and needs no account. These extras switch on only if you supply a key of your own, and stay silent otherwise:

Every one of these reads its key from the environment. AF-KLM is the single exception, and only when you ask for it: under an explicit --provider afklm it may also read the macOS Keychain or 1Password. Nothing in this table reaches a credential manager during an ordinary search — the reasoning is in docs/ARCHITECTURE.md.

A released build POSTs at most one heartbeat a day to https://telemetry.revaluator.ai/v1/heartbeat. Set TRVL_TELEMETRY_ENDPOINT to send that same body somewhere else. The body is only:

• a fixed project tag (trvl) and event name (heartbeat)
• the trvl version
• the Go runtime string (OS, architecture, Go version, e.g. darwin/arm64/go1.26.9)
• a random install id, created on the first such send and stored in ~/.trvl/install-id

No hostname, no username, no search queries, no travel data. The JSON body does not contain your IP. Cloudflare terminates TLS for the default host, so Cloudflare can see the connection address. The receiver does not write that address down. From the connection, Cloudflare’s geolocation supplies a city name and a country code, and those two values are stored with the heartbeat. They are approximate: a VPN, a relay, or a mobile network can place the city at the network exit rather than where you are, and some connections have no city. Coordinates and request headers are not stored. The install id stays the same across days, so those requests can be linked to each other, including the city stored with them. It is random and contains nothing about you. The request has a 3-second timeout and a failure is ignored.

Development builds (go build / go run, version dev), CI, and tests do not send. To turn it off, set any one of these before running trvl:

export TRVL_NO_TELEMETRY=1   # trvl-specific switch
export NO_TELEMETRY=1        # common convention
export DO_NOT_TRACK=1        # cross-tool Do-Not-Track signal

Local stdio is the default and safest transport. trvl mcp --http binds to 127.0.0.1, requires a bearer token, and generates one at startup if unset. Remote exposure, scoped read/write tokens, and OAuth 2.1 introspection: docs/REMOTE-MCP-OAUTH.md.

• No tools showing? Restart your AI client after trvl mcp install; confirm which trvl is on $PATH.
• Empty flight results? Some routes have no Google Flights data — try a major pair like trvl flights HEL LHR 2027-07-01.
• Ground transport times out? Rail/ferry providers throttle; retry after 30s or pass --timeout 3m.

Full troubleshooting: docs/CLI.md.

Glama · LobeHub · Smithery · MCPHub · Cursor Directory · PulseMCP · MCP Market · pkg.go.dev

Third-party directories cache their own descriptions and may show older tool or command counts. Use this README, the changelog, and the official MCP Registry for the current surface and release version.

Independent coverage: Roberto Reale's Budget Travel Pipeline — an independent build-and-test that surfaced real fixes and shaped the v1.10 trust roadmap.

Part of a suite of MCP tools: mcp-gateway (universal gateway) · nab (web extraction with anti-bot) · axterminator (macOS GUI automation).

trvl is a personal-use tool that reads public-facing web APIs (Google Flights, Google Hotels, and others). It does not bypass authentication or circumvent rate limits; request patterns are throttled to look like manual browsing. Automated access may violate some providers' Terms of Service — you are responsible for compliance in your jurisdiction.

trvl flights and trvl hotels accept an optional --stealth flag that routes the fetch through trvl's Chrome HTTP/2 fingerprint transport. It is off unless you pass it, activates only for hosts you list in TRVL_STEALTH_ALLOWLIST, and does nothing but log one line for any host not on that list — an empty allowlist means it never activates. Only flight and hotel search honour it. Using stealth against sites whose terms prohibit automated access is your responsibility.

export TRVL_STEALTH_ALLOWLIST=".google.com"
trvl flights HEL NRT 2027-09-01 --stealth

Licensed under PolyForm Noncommercial 1.0 — free for personal and noncommercial use. Commercial use (company-internal, hosted service, embedding in paid platforms) requires a separate license: EUR 500/month per named project via GitHub Sponsors, see COMMERCIAL.md and LICENSING.md.

Built on fli, utls, and SerpAPI's parameter reference.

If trvl saved you a browser tab or an API subscription, a star helps other travellers (and their assistants) find it. That's the whole ask.