Short answer: AICredits runs a remote MCP server at https://api.aicredits.in/mcp. Authenticate with a read-scoped akm- management key and your assistant can look up live models, INR pricing, cost estimates, your balance, recent usage and your key list, instead of guessing from stale training data. It never makes model calls.
Disclosure: the aicredits.in links are affiliate links and I use the service. Tool names, scopes and URLs come from AICredits' MCP docs as read on 16 Oct 2026. I have not verified how every Claude client handles remote HTTP servers, and the example outputs described below are illustrative.
Ask a chatbot what the cheapest model with a 128K window costs in rupees and you'll get a confident answer built on whatever it remembers. Prices and model lists change every few weeks, so the answer is often wrong. The fix is to let the assistant call a tool that reads the live catalog. That is what this MCP server is for. If MCP is new to you, read what is the MCP protocol first, then come back.
What does the server expose?
Seven tools, all read-only:
| Tool | Scope needed | What it does |
|---|---|---|
models_list | none | List models with live INR pricing per 1M tokens; filter with search |
model_get | none | One model's pricing, context window, modalities, status |
pricing_estimate | none | Estimate a request's INR cost from token counts |
docs_search | none | Search AICredits docs, get links |
credits_balance | wallet:read | Your wallet balance in INR |
usage_summary | usage:read | Recent request count, INR spend, top models (up to 90 days) |
keys_list | keys:read | Your inference keys, metadata only, never the secret |
The first four need a valid management key but no particular scope. The other three check the scopes you granted.
How do you connect it?
First, create the key. In the dashboard, open Management API (beta), choose Create management key, and tick only wallet:read, usage:read and keys:read, or fewer. Copy the akm- value; it is shown once. It must be a management key. Your sk-live- inference key will not work here.
Claude Code (from the docs):
claude mcp add --transport http aicredits https://api.aicredits.in/mcp \
--header "Authorization: Bearer akm-YOUR-MANAGEMENT-KEY"
Cursor, in .cursor/mcp.json:
{
"mcpServers": {
"aicredits": {
"url": "https://api.aicredits.in/mcp",
"headers": { "Authorization": "Bearer akm-YOUR-MANAGEMENT-KEY" }
}
}
}
Generic JSON for any client supporting streamable HTTP:
{
"url": "https://api.aicredits.in/mcp",
"transport": "http",
"headers": { "Authorization": "Bearer akm-YOUR-MANAGEMENT-KEY" }
}
Claude Desktop. The documentation does not give a Claude Desktop snippet, so I won't invent one. Desktop's support for remote servers with custom headers has varied by version and plan. Open the app's MCP or connectors settings and check whether it accepts a remote HTTP URL with an Authorization header. If it only supports local stdio servers, a common workaround is a local bridge such as the mcp-remote npm package, which I have not tried against this server. Use Claude Code or Cursor, where the docs show a working config, if you want the least-effort path.
The transport is streamable HTTP; SSE-only legacy clients are not supported. To check the connection, ask the assistant to list its AICredits tools, or run claude mcp list in Claude Code.
What should you ask it? Example prompts
Copy these. The first four come from AICredits' docs; the rest are mine. Outputs will depend on live data, so I'm not showing numbers.
- "Which is the cheapest model on AICredits that handles 128K context?"
- "What's my AICredits balance and what did I spend this week?"
- "Estimate the cost of 50K input and 2K output tokens on claude-sonnet-4.5."
- "List my AICredits API keys and their budgets."
- "Search the AICredits docs for how the per-key budget works and link me to the page."
- "Compare three chat models for a 10-million-token monthly workload. Use pricing_estimate for each and show the result as a table in rupees."
- "Which model did I spend the most on in the last 30 days, and is there a cheaper alternative with similar capabilities? Say which parts are from tool data and which are your judgement."
Prompt 7 matters. The tool gives numbers; "similar capabilities" is the model's opinion. Ask it to separate the two. The same caution applies to the catalog ranking script: price sorting is a shortlist, not a decision.
Does the estimate match the real bill?
Treat pricing_estimate as a planning figure and verify it once. Price is built from provider USD price x live USD/INR x 1.05 forex buffer x 1.05 platform fee. Cached and long-context pricing follow provider-specific rules in the pricing docs (for example Claude cache reads at 0.1x and requests over 200K tokens at 2x input), so a plain token-count estimate can undershoot for those cases. A quick check: make one small call, read the cost on your usage page, and compare it with what the tool predicted for the same token counts. I haven't run that comparison, so I can't tell you the gap.
What if it doesn't connect?
| Symptom | Likely cause | Fix |
|---|---|---|
| 401 "Invalid management key" | You pasted an sk- inference key | Create an akm- key on the Management API page |
| Balance or usage tool refuses | Key lacks wallet:read or usage:read | Create a new key with that scope; scopes are fixed per key in the docs I read |
| Client hangs or errors on connect | Client only speaks SSE or stdio | The server needs streamable HTTP; use a client that supports it |
| Tools list is empty | Header not sent | Re-check the Authorization: Bearer akm-... line and restart the client |
| Management API disabled message | The beta API is temporarily off (503) | Retry later |
What are the security notes?
An MCP key is a credential sitting in a config file that an AI tool reads. Treat it that way.
- Scope down.
wallet:read,usage:read,keys:readis the most an assistant needs. A leaked key with only those can read, not create keys or spend credits. - Never grant write scopes (
keys:write,team:write) to a key an assistant holds. - Separate keys per client. One key for your laptop, a different one for CI, so you can revoke one.
- Keep it out of git. Use an environment variable or your client's secrets mechanism. Config files like
.cursor/mcp.jsonget committed by accident. - Remember what the data contains.
keys_listreturns metadata, such as names and budgets, never secrets. Even so, model output can end up in logs or shared chats. - Prompt injection. Tool results are data the model reads. The tools here return AICredits data, so the injection risk is low, but if you mix this server with web or email tools in one session, the prompt injection lesson and the MCP security checklist apply.
- Revoke on doubt. Management keys can be revoked from the Management API page at any time, and they expire after at most a year.
What this server cannot do
- It cannot run chat completions. Point your tool's model provider at
https://api.aicredits.in/v1with ansk-live-key for that, as in the Claude-in-India walkthrough. - It cannot top up your wallet. That happens in the billing page, as described in the UPI top-up guide.
- It is not a replacement for the usage dashboard when you need a full audit; for exports use the beta management API.
Bottom line
If you spend your day inside Claude Code or Cursor, adding this server costs one command and a read-only key, and it replaces "what does that model cost these days?" with an answer from live data. Keep scopes narrow, verify one estimate against a real bill, and keep inference on a separate sk- key. To try it, create an AICredits account, make a read-only management key and run the claude mcp add line above. For the wider picture, the AICredits review is the hub.
