Cheap First
MCP server that answers with a cheap model and escalates to a premium one only when
Answered our latest tool-list check yesterday with 2 tools. It will be marked verified after 3 successful checks in 7 days.
Cheap First is a freemium MCP server for agent tooling. It runs as a hosted remote server, so there is nothing to install locally.
- Pricing
- Freemium
- Platforms
- API, CLI, Web
- Runs
- Remotely
- Version
- 0.1.1
- Listing updated
Similar MCP servers
Install
Run in a terminal
claude mcp add --transport http cheap-first https://cheap-first.vercel.app/mcpAdd to Settings → Connectors → Add custom connector
https://cheap-first.vercel.app/mcpAdd to ~/.cursor/mcp.json
{
"mcpServers": {
"cheap-first": {
"url": "https://cheap-first.vercel.app/mcp"
}
}
}Add to .vscode/mcp.json
{
"servers": {
"cheap-first": {
"type": "http",
"url": "https://cheap-first.vercel.app/mcp"
}
}
}Add to ~/.codeium/windsurf/mcp_config.json
{
"mcpServers": {
"cheap-first": {
"serverUrl": "https://cheap-first.vercel.app/mcp"
}
}
}Add to Grok Bot desktop app → Connect apps → Add a custom connector
https://cheap-first.vercel.app/mcpEnter a Name (for example "Cheap First"), choose Type: Remote HTTPS, paste this as the Server URL and choose Add plugin.
Add to your client's MCP settings (most use this shape)
{
"mcpServers": {
"cheap-first": {
"type": "http",
"url": "https://cheap-first.vercel.app/mcp"
}
}
}Using several servers? Build one config file with the MCP config generator.
About Cheap First
The README describes a pipeline where a low-cost model answers first, a separate checker model scores that answer from 0 to 100, and a frontier model is called only if the score falls below a threshold. The team reports a benchmark of 143 questions where the cheap path cost $0.27 against $1.33 for premium-only answers, with the checker escalating 4 questions.
For MCP clients, the server exposes an ask tool with strict, balanced and budget modes that returns the answer, whether it escalated, the check score and the cost. A stats tool returns the published benchmark numbers. The same service also offers an OpenAI-compatible chat endpoint and a standalone output verifier for checking another agent's answer.
It suits developers who send many routine questions through an agent and want to cut model spend without giving up accuracy on the questions that really need a stronger model. Free demo calls are rate limited, and paid use runs per answer over x402 or with a prepaid key.
Tools this server provides (2)
Read directly from the server.
1 of 2 tools says it's read-only. 1 of 2 tools says it can delete or change data.
From the server's own tool annotations; not verified by us.
- ask_cheap_firstSays it can change data
- Answer a question the Cheap First way: Dolphin answers, SERV Reasoning scores it, and a premium model answers instead only if the score misses the pass mark. Returns the answer, whether it escalated, the check score, the cost and the estimated saving. Free calls use the public demo limits (600 characters, 10 an hour per IP, 300 a day). Send Authorization: Bearer cf_... or an x402 payment of $0.005
- cheap_first_statsSays read-only
- Published benchmark figures from the saved result files. The in-house run, the harder run graded by code, and the public GSM8K and ARC-Challenge run when that file is complete. The public summary includes accuracy, cost, escalation, latency, and where the checker missed. Does not call a model.
What it does
- Cheap-first routing
- Each question is answered by a low-cost model before any premium model is considered.
- Automatic checker escalation
- A separate checker scores the cheap answer and triggers a premium model when the score is below the pass mark.
- Selectable strictness modes
- Strict, balanced and budget modes set how high the pass mark is, trading cost against escalation frequency.
- Per-answer cost reporting
- Each response reports the check score, the cost of the call and an estimated saving against premium-only use.
- OpenAI-compatible endpoint
- A chat completions route lets OpenAI-style clients switch by changing the base URL.
- Standalone output verifier
- The checker is also sold separately to grade another agent's answer before payment is released.
Good for
- Lowering model costs for high-volume routine questions in an agent workflow
- Adding a cheap-first answer tool to Claude or Cursor without choosing a model per query
- Gating payment to a hired AI agent on a pass or fail verdict from an independent checker
- Comparing answer quality and cost between cheap and premium models on a fixed question set
Pricing
Freemium
Free demo with rate limits; paid answers at $0.005 each via x402 or a prepaid key.
Prices change. Check cheap-first.vercel.app before you buy.
Cheap First FAQ
How do I connect Cheap First to an MCP client?
The README gives the endpoint https://cheap-first.vercel.app/mcp and config examples for Cursor, Claude Desktop and Claude Code. The free demo works without headers, and a prepaid key can be sent as a bearer token.
What does a single answer cost?
The paid API charges $0.005 per answer in USDC on Base via x402. The README notes that an escalation to the premium model can cost more than that, and the server caps premium spend per instance.
Are the tools listed in the MCP registry entry?
The registry entry lists the ask and stats tools with their parameters, and the README describes the same two tools.
Which models does Cheap First use?
The README names DeepSeek V4.1 Flash via the Dolphin network as the first model, a GPT-6 Luna checker on SERV Reasoning, and Claude Opus 5.5 on SERV as the premium model.
How do I add Cheap First to Claude Code?
Run this command in your terminal: claude mcp add --transport http cheap-first https://cheap-first.vercel.app/mcp
What are the best alternatives to Cheap First?
The closest alternatives to Cheap First are Agentic Platform, ask-kit, Ask real people: human feedback, user testing, surveys, APEX AI, LLM, PDF, QR, JSON & Media MCP and WritersLogic agent-infra.
Where this comes from
- Description: written from the project's README, the project's website and its MCP Registry entry.
- Status and tools: checked automatically by AIPoweredBot on 21 November 2026.
- Install steps: from the official MCP Registry entry.