PromptSpend

@promptspend/mcp

LLM pricing for coding agents — every price carries its source and the date it was last confirmed.

There are already MCP servers that will tell you what a model costs. This one tells you where the number came from, when it was last checked, and which numbers two sources disagree about. And it will cost a described workload rather than just quoting a rate, because the arithmetic between "$3 per million tokens" and "$4,955 a month" is where nearly every estimate goes wrong.

claude mcp add promptspend -- npx -y @promptspend/mcp

No API key. No account. MIT.

---

Why another pricing server

Every price comes with its paperwork.

// Real output, not an illustration: `get_price` for gpt-5 on 2026-08-03.

{ "model": "gpt-5", "input_per_million_usd": 1.25, "output_per_million_usd": 10, "provenance": { "source": "vendor", "source_description": "Hand-verified against the provider's own published pricing page.", "last_verified": "2026-08-03", "disputed": false, "upstream_stale": false, "confidence": "Confirmed against the provider's own pricing page on 2026-08-03.", "verified_url": "https://developers.openai.com/api/docs/pricing", }, }

last_verified moves as rows are re-read, so the date here will age; the shape will not. What is worth checking is that the block is present at all — a price without one is a price you cannot audit.

That matters more to an agent than to a person. Someone reading a web page sees the interface around the number and forms their own view of how much to trust it. A model handed a bare figure has nothing, and will repeat it with whatever confidence the phrasing implies. Given the date and the source it can say _"as of 1 August, per OpenAI's pricing page"_ instead of stating a number as though it were timeless.

And it says when it does not know. Where two sources disagree and no human has adjudicated, the response is marked disputed: true with both figures, rather than quietly picking one:

DISPUTED — two sources disagree on this price and no human has adjudicated it.

The figure shown is the primary feed's number, last confirmed 2026-08-02. Treat it as indicative and say so.

The tool a price lookup cannot provide

estimate_cost runs promptspend.com's own cost engine — imported, not reimplemented, so a number this server reports and a number the website shows cannot drift apart. A test asserts it.

It accounts for the things hand-rolled estimates miss:

• Conversation history compounds. Turn _N_ re-sends turns 1…*N*−1 as input, so cost grows with the square of the turn count.
• Cache _writes_ cost more than input — 1.25× at both OpenAI and Anthropic. Counting only the cheaper reads reports a saving your invoice will not have.
• Long-context tiers apply per request, not per conversation.
• Reasoning tokens are billable even though you never see them.
• Output is priced separately, typically 3–5× input.

_"What would a support assistant cost at 4,000 conversations a day on GPT-5

versus DeepSeek V3.2?"_

Context footprint

MCP tool definitions are loaded into your context on every turn, for as long as the server is connected. A heavy server can add thousands of tokens to every message, which is why the advice going round is to prefer a CLI for read-only data.

A pricing server that quietly taxes every turn, to tell you about token costs, would be an easy and deserved joke. So:

This server adds ~897 tokens per turn. Three tools, short descriptions,

budgeted at 1,200 and checked in CI by npm run check:footprint.

If that number ever stops being true, the build fails — the README figure is asserted against the actual manifest, not typed once and trusted.

Tools

`get_price` with `as_of: "2026-09-15"` returns the rate PromptSpend published on that UTC day, read from the price history — useful for checking an old invoice. A day before PromptSpend began tracking the model is answered "not tracked", never with the first rate seen, and a later correction (PromptSpend had been reading the wrong page) is returned with the figure, so an agent does not present a misreading as what the vendor charged.

`estimate_cost` with `usage` prices measured totals — say a month from a usage export — instead of a described workload: { "input": 1200000, "output": 300000, "cache_read": 4000000, "cache_write": 200000 }, where input excludes cache reads. A model without a published cache rate is billed at its full input rate (and the response says so), and a model with a long-context tier is flagged, because totals cannot show which requests crossed it.

find_cheaper returns candidates to test, never a recommendation — the capability index behind it is an illustrative estimate, not a benchmark, and the response says so.

Install

Claude Code

claude mcp add promptspend -- npx -y @promptspend/mcp

Cursor / Claude Desktop / Windsurf — in mcp.json:

{ "mcpServers": { "promptspend": { "command": "npx", "args": ["-y", "@promptspend/mcp"] } } }

No install — hosted at https://promptspend.dev/mcp. Same tools, same answers, imported from this package rather than copied. In Claude.ai or ChatGPT, add it as a custom connector with that URL; in Claude Code:

claude mcp add --transport http promptspend https://promptspend.dev/mcp

The hosted server receives your tool arguments (a model name, a workload's token counts) to answer them, and stores none of them; it needs no key or account. Like the rest of the pricing API at promptspend.dev, it has Cloudflare's request logging switched on: each entry records the time, the address (/mcp), request headers and connection details such as an IP address, kept for up to seven days. PromptSpend's code never logs the arguments themselves. Run the npm package instead if a workload description should never leave your machine.

Or as an Agent Skill, which costs context only when a price question comes up rather than on every turn: see skills/.

What it will not do

• It will not pass off a stale price as current. The catalog is fetched, never bundled. A retained copy may be used for at most 24 hours when the live file cannot be reached, and every affected tool response carries an explicit catalog_status warning written for the agent to relay. Past that bounded grace period, the tools return an error rather than an old number.
• It has no benchmarks, latency or endpoint data. Other servers do. Three tools is a deliberate choice about your context budget.
• It has no telemetry. Neither the npm package nor the hosted endpoint sends analytics or usage reports, and neither needs a key or an account. The hosted endpoint's request logging is described under Install.

Scope of the prices

Standard-tier, global-endpoint list prices in USD. Not modelled: regional and data-residency premiums, priority tiers, server-side tool-call fees, fine-tuning, and negotiated or committed-use discounts.

---

The same catalog is also in free native apps for iPhone and iPad and Android.

Part of PromptSpend — MIT. Found a wrong price? That is the most serious class of bug this project can have. Open an issue.