mcpcost

mcpcost

Is that MCP server alive, and what does it cost your context window?

Live board: mcpcost.ulehla.dev

mcpcost probes every remote server in the official MCP registry and records, for each one:

• whether it answers initialize and tools/list, and how fast
• the protocol version it negotiates
• whether it needs auth, which scheme (from WWW-Authenticate), and how a client gets in: OAuth with dynamic client registration, OAuth with a client id from the publisher, or a key, read from the server's public OAuth metadata
• whether it asks for payment (HTTP 402)
• how many tools it has and how many tokens their definitions cost

mcpcost: what do your own servers cost?

npx mcpcost                     # reads Claude Code, Claude Desktop, Cursor, VS Code, Windsurf and Gemini CLI configs
npx mcpcost --config .mcp.json  # one file
npx mcpcost --budget 30000      # exit 1 when the total is above the budget, for CI
npx mcpcost --claude            # exact counts from the Anthropic count_tokens endpoint (free, needs ANTHROPIC_API_KEY)
npx mcpcost mcp                 # the same measurements as an MCP server over stdio, for your agent

It starts every configured server the way your client does, lists the tools and prints the token cost per server and in total. See packages/mcpcost.

Servers that need auth

A server that answers 401 is reachable but not measured until it gets credentials. Store them once; every later measurement of that URL uses them.

npx mcpcost auth https://mcp.linear.app/mcp                                   # OAuth: register, consent in the browser, store the tokens
npx mcpcost auth https://api.githubcopilot.com/mcp/ --header "Authorization: Bearer ghp_…"   # an API key instead
npx mcpcost auth --list                                                       # what is stored, without the secrets

The OAuth flow follows the MCP authorization spec: discovery, dynamic client registration, PKCE, a callback on 127.0.0.1. The consent screen is the one step that needs a person. After it, the tokens live in ~/.mcpcost/credentials.json (mode 0600, or $MCPCOST_CREDENTIALS), npx mcpcost sends them for that URL, and a refreshed token goes back to the file.

The board does the same: npm run credentials:push copies the file into the D1 table credentials; the daily run reads it before the probe, writes the refreshed tokens back after it, and marks the server "measured with credentials". A refresh token that a service revokes or lets expire needs the consent step again.

Without credentials, the board still shows a tool set for many gated servers: tool definitions do not depend on the token, so when the publisher's repository holds a package on npm or PyPI, probe/derive-packages.ts finds it and the weekly container run measures it with placeholder values. The server page then says "from the publisher's package". probe/derived-aliases.json maps the few servers whose registry entry names no repository.

In CI, as a GitHub Action

- uses: ondraulehla/mcpcost@v1
  with:
    config: .mcp.json   # one or more config files, one per line
    budget: 30000       # fail when the tool definitions cost more than this

The job summary gets a table with every server, its result and its cost, plus the most expensive tools. A pull request that adds a 40 000-token server fails the check. Set claude: true and put ANTHROPIC_API_KEY in env for exact counts from the free count_tokens endpoint; the default is the o200k_base estimate. Outputs: tokens, servers, over-budget.

On a pull request the action also measures the same config files on the base branch and posts one comment with the difference: servers added, removed and changed, the token delta per server, the new total and the budget verdict. The comment is updated on every push, never duplicated. The job needs pull-requests: write for it; without that permission the step warns and the check still runs. Set comment: false to turn it off. A complete workflow is in .github/examples/mcp-budget.yml.

In Claude Code, as a plugin

In a Claude Code session:

/plugin install mcpcost --marketplace ondraulehla/mcpcost

Or from the shell: claude plugin marketplace add ondraulehla/mcpcost, then claude plugin install mcpcost@mcpcost.

The plugin adds one skill, /mcpcost:measure, which runs npx mcpcost on your configs and reads the result, and the board's MCP server, so Claude can look up or probe a server before it goes into a config. What the plugin itself costs: one skill description in the context, and four tools that Claude Code defers behind tool search.

The board

mcpcost.ulehla.dev is rendered on the edge from a D1 database and refreshes every night. The only scripts it loads are a theme switch and the Cloudflare Web Analytics beacon, which counts page views without cookies.

• Overview: the share of servers that answer, token cost percentiles, and a form to check any server
• Check: probes one URL live, with the same steps as the daily run; the exact token count runs in the browser
• Statistics: results by cause, token cost and tool count percentiles, the heaviest tool sets, protocol versions, auth schemes, transports, biggest publishers
• Servers: full-text search over name, title, description and host (word prefixes, so git copilot finds the GitHub Copilot servers), filters by result, protocol and transport, seven sorts, and every filter is a plain URL you can share
• Server page: status, latency, protocol, capabilities, the most expensive tools, servers with the same tool set, history, the badge, and e-mail alerts: a mail when the server stops answering, comes back, or its tool set or token cost changes
• Track: a server that is not in the registry can be tracked from the check page; the daily probe then measures it under the name url/<host>/<path>
• Hosts: the 500 biggest publishers with how many distinct tool sets they actually serve

For agents and scripts

• JSON: /api/s/<name>.json, /api/servers.json?q=…&status=ok&sort=tokens, /api/check.json?url=…, /api/summary.json. The list API takes at most 6 requests in 10 seconds from one address; its answers are cached for three hours.
• MCP server over Streamable HTTP, no auth: https://mcpcost.ulehla.dev/mcp with the tools lookup_server, search_servers, check_server, registry_summary. The manifest for the registry is site/server.json.
• /llms.txt lists all of it. /tokens answers the general question: how many context tokens does an MCP server cost.

Each server on the board gets a page and a badge:

[![mcpcost](https://mcpcost.ulehla.dev/badge/io.github.you/your-server.svg)](https://mcpcost.ulehla.dev/s/io.github.you/your-server)

First full run, 2026-10-03

The registry held 38 942 latest entries. 24 294 of them have a remote endpoint, and the probe connected to all of them in 40 minutes.

Of the 12 975 servers that answered:

• the median tool set costs 1 727 tokens, the 90th percentile 7 429, the 99th 28 200
• 208 servers cost more than 20 000 tokens, 46 more than 50 000, and 3 more than a whole 200k window
• the heaviest is com.replynodes/mcp: 66 tools, 431 925 tokens
• the median server has 8 tools, the largest 1 385
• median initialize latency is 455 ms, the 90th percentile 1.5 s
• 60 % negotiate protocol 2025-11-25, 24 % still 2025-03-26, 4 % 2024-11-05
• 122 servers answer with protocol 2026-07-28, which the current TypeScript SDK rejects

The registry is lumpy. 16 161 of 16 534 hosts publish one server. Five hosts publish more than 6 000 entries between them: one Workers subdomain with 2 365 near-identical entries, a gateway with 1 716 entries of which 1 321 return a JSON-RPC error on initialize, and a host with 276 entries that all answer 402.

How it works

registry ──▶ probe/run.ts ──▶ data/raw/probe-run.json
                                   │
                        probe/build-data.ts
                                   │
               data/latest/{servers,summary,hosts}.json
                                   │
                           probe/to-sql.ts ──▶ import.sql ──▶ D1
                                                                │
                              site/ (Astro on Workers) ◀────────┘

• packages/mcpcost/ the library and the CLI mcpcost: registry client, remote and stdio probe, config discovery, token counting
• probe/ the batch runner and the data pipeline
• site/ the board: Astro 7 with the Cloudflare adapter on Workers, rendered from D1 with FTS5 search, SVG badges, no client JavaScript beyond the theme switch and the analytics beacon
• .github/workflows/probe.yml runs the whole pipeline daily: probe, counts, build-data, to-sql, the D1 import. The run's summary.json, servers.json and hosts.json go up as a workflow artifact; the next run diffs against it and imports only the rows that changed, inside the free tier's 100 000 row writes a day
• .github/workflows/deploy.yml deploys the site when site/ changes

Methodology

• Transport. The probe uses the official TypeScript SDK as a client. Streamable HTTP first, SSE when that is the only remote.
• No GET stream. The probe answers 405 to the SDK's standalone GET, so it never opens a server-to-client stream. The probe needs no server-initiated messages, and some servers stall tools/list while that stream is open (DeepWiki held it for 15 s and never answered the POST meanwhile).
• Exact token count. The daily run asks the Anthropic count_tokens endpoint for every alive server, and the Gemini countTokens endpoint when GEMINI_API_KEY is set (the CLI: --gemini). Anthropic first: one request with the server's tools plus a marker tool, one with the marker tool only, and the difference is what the tools cost in claude-opus-5-5. A second count uses Claude Code's mcp__<server>__ prefix on each tool name. The endpoint is free and calls no model. A server is counted again only when its tool definitions change. About 5 % of servers publish a schema the endpoint refuses; they show the estimate and the reason.
• Estimate. Each tool is serialised as compact JSON with name, description and input_schema, and counted with the o200k_base tokenizer. The exact count is about 1.9 times the estimate at the median. The estimate needs no key and is the same for every server.
• Same tool set. A hash of the sorted tool names and schema sizes marks servers that serve identical tools under different names.
• Timeouts. 15 seconds per request. At most one probe runs against a host at a time.
• Getting access. For a server that answers 401 or 403, the probe reads the OAuth protected-resource metadata (RFC 9728) and the first authorization server's metadata (RFC 8414, then OpenID discovery): authorization servers, scopes, the publisher's guide, dynamic client registration (RFC 7591), PKCE, grant types. The server page shows it as "Getting access", the JSON API as authKind and authMetadata. Two to four small requests, no token, no consent; the probe result does not change.
• Identification. Every request carries the User-Agent mcpcost/<version> (+https://github.com/ondraulehla/mcpcost). To exclude a server, open an issue with its registry name.

Run it

npm install
npm run probe -- --sample 200 --per-host 2        # a sample, results in data/samples/
npm run probe -- --all --out data/raw/probe-run.json
node --import tsx probe/build-data.ts --in data/raw/probe-run.json --out data/latest
node --import tsx probe/to-sql.ts --in data/latest --out data/latest/import.sql
npm test --workspaces --if-present

Site, locally against a local D1:

cd site
npm run db:schema:local && npm run db:import:local
npm run dev

The alert mails need a sending domain: either Cloudflare Email Sending (Workers Paid plan, then the send_email binding in site/wrangler.jsonc) or Resend (verify the domain there and npx wrangler secret put RESEND_API_KEY). Without one, the watch form says that the confirmation mail could not be sent.

Deploying needs a Cloudflare account with a D1 database named mcpcost (wrangler d1 create mcpcost, then put the id in site/wrangler.jsonc). The workflows expect the repository secrets CLOUDFLARE_API_TOKEN (Workers Scripts: Edit, D1: Edit) and CLOUDFLARE_ACCOUNT_ID. Both workflows apply site/schema.sql and the pending files in site/migrations/ on every run. The two indexes on the shown token count (migrations 007 and 008) are the exception: each build writes a row per server, so the probe workflow applies one a day before the import and gives the import the rest of the write budget. A new database needs them applied by hand once.

The edge rules

The Workers Free plan counts every request the Worker answers, a 403 as much as a page. On 2026-10-22 and 23 scrapers with Chrome user agents on hosting addresses (Meta's cloud, Alibaba Cloud, RapidSeedbox) made 4 000 list variants an hour and spent the account's 100 000 requests a day, although the Worker refused them. What stops them must run before the Worker: site/edge/rules.mjs writes the zone's WAF rules through the API, idempotently.

• a custom rule that blocks those networks on the whole host, verified bots excepted
• a custom rule that gives every list variant (/servers?…) from anything but a verified bot a managed challenge; a browser passes it once, a script does not. The list API (/api/servers.json) is not in it: scripts and agents cannot pass a challenge, so the API stays open under the rate limiting rule
• a custom rule that blocks a verified bot on the list variants: robots.txt disallows them for everyone, search engines never ask, and Meta's crawler (meta-externalagent behind a Chrome agent string) walked 3 000 of them an hour on 2026-10-23
• the one rate limiting rule of the Free plan: more than six list variants in ten seconds from one address is blocked for ten seconds

CLOUDFLARE_API_TOKEN=… node site/edge/rules.mjs --dry-run   # shows what it would write
CLOUDFLARE_API_TOKEN=… node site/edge/rules.mjs

The Edge rules workflow (.github/workflows/edge.yml) runs the script when the file changes on main, or by hand, with the CLOUDFLARE_ZONE_TOKEN secret: a token with Zone: Zone WAF: Edit and Zone: Zone: Read on the zone. The probe's token has neither. The middleware in site/src/middleware.ts keeps the same refusals as a second layer, for the workers.dev host the zone rules do not see.

The watch

site/edge/watch.mjs says who read the board in a window of time and what it cost: the Worker's requests and D1 rows per hour from the account's analytics, and the networks, agents and URL shapes behind them from the Worker's logs. The Watch the edge workflow (.github/workflows/watch.yml) runs it every hour over the hour before. A threshold crossed (5 000 Worker requests an hour, 800 an hour from one network whose agents do not name a crawler, 300 list variants an hour from one network whatever it calls itself, a million D1 rows an hour; the WATCH_* variables) turns the run red and opens the issue labelled edge-watch, or adds the hour to the open one. The tables name the network; its ASN goes into SCRAPER_ASNS in site/edge/rules.mjs, and the Edge rules workflow blocks it before the Worker. The report also names the D1 queries that read the most rows in those hours. With Zone: Analytics: Read on the zone token, it shows what the zone stopped before the Worker as well.

CLOUDFLARE_API_TOKEN=… CLOUDFLARE_ACCOUNT_ID=… node site/edge/watch.mjs                                   # the last hour
CLOUDFLARE_API_TOKEN=… CLOUDFLARE_ACCOUNT_ID=… node site/edge/watch.mjs 2026-10-22T22:00:00Z 2026-10-23T06:00:00Z   # one night

The old host

The site was mcp-pulse.ulehla.dev until 2026-10-15. The old host stays a route of the Worker, and the middleware in site/src/middleware.ts answers every request on it with a 301 to mcpcost.ulehla.dev.

Static assets are served before the Worker, and the Astro adapter serves prerendered pages before the middleware too, so a file in site/public/ or a prerendered page answers 200 on the old host. run_worker_first in wrangler.jsonc does not change that. So robots.txt, llms.txt and the methodology page are routes, and every page a crawler reads goes through the middleware. Only the favicon, the fonts under /_astro/ and the search engine verification file remain static.

To redirect those as well, with no Worker invocation at all, add a Redirect Rule to the ulehla.dev zone in the Cloudflare dashboard (Rules, Redirect Rules, Create rule):

• When incoming requests match: Hostname equals mcp-pulse.ulehla.dev
• Then: Dynamic redirect, expression concat("https://mcpcost.ulehla.dev", http.request.uri.path), status code 301, preserve query string on

Redirect Rules run before the Worker and before static assets, so the rule covers every path. The middleware stays as the fallback for the workers.dev host, which the zone rule does not see.

Licence

MIT