Kyma API MCP Server

Kyma API is an LLM API gateway that routes one key to 120+ models and publishes measured uptime per model.
Every model behind one endpoint, with public rankings and a spend cap you set.
https://mcp.kymaapi.com/mcp

Install · Tools · Security and spend · Examples · Troubleshooting · Registries

Kyma API is an LLM gateway: one OpenAI-compatible endpoint (plus an Anthropic-compatible one) in front of open and frontier models, with uptime measured per model from real traffic and scheduled probes. This server gives any MCP client the live catalog, prices, rankings, uptime and your own spend as read tools, and one chat tool that is guarded by a per-connection spend cap.

• Hosted, remote, OAuth sign-in. Streamable HTTP with OAuth 2.1, PKCE and dynamic client registration. No API key is pasted into the client.
• A dedicated key per connection. Separate from your REST API keys, capped at $10 per 30 days by default (adjustable from $1 to $500 on the approval screen or at kymaapi.com/integrations), revocable at any time.
• Read tools never charge. Catalog, pricing, rankings, uptime, credits, spend, your usage and your requests are free to call and auto-approved by most clients.
• send_message is priced before it spends. The client must allow it first; max_cost_usd is enforced by deriving a max_tokens that fits the ceiling from the catalog price, so a call above it is refused before anything is sent. Every reply reports what the call cost, what has been spent and the cap.
• Measured uptime, not a promise. get_model_uptime returns the 30-day rate per model from kymaapi.com/status. If a route fails mid-request, Kyma retries the same model on another route.
• A read-only endpoint for reviewers. https://mcp.kymaapi.com/mcp/readonly accepts the same tokens but keeps only read scopes for the request: no send_message, no set_low_balance_alert, whatever the grant was approved with. Metadata: /.well-known/oauth-protected-resource/mcp/readonly.
• Beyond chat. The same Kyma key covers speech-to-text, text-to-speech, embeddings, rerank, image and video through the REST API; list_models shows all of them.

Install

claude mcp add --transport http kyma https://mcp.kymaapi.com/mcp

{ "mcpServers": { "kyma": { "url": "https://mcp.kymaapi.com/mcp" } } }

{ "servers": { "kyma": { "type": "http", "url": "https://mcp.kymaapi.com/mcp" } } }

# ~/.codex/config.toml
[mcp_servers.kyma]
url = "https://mcp.kymaapi.com/mcp"

codex mcp login kyma

{ "mcpServers": { "kyma": { "url": "https://mcp.kymaapi.com/mcp", "type": "streamableHttp" } } }

gemini extensions install https://github.com/kyma-api/kyma-mcp-plugin

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

Tools

18 tools. 16 appear on a default connection; list_keys and set_low_balance_alert appear when their optional scopes are granted on the approval screen.

Every send_message reply carries cost, spent_usd and spend_cap_usd, plus structuredContent with the model that served it and the routes tried. get_spend shows what is left before you call it.

Security and spend

• What a connection can do: read the catalog, prices, rankings, uptime, your balance and this connection's spend; send chat completions up to its cap; with optional scopes, list your keys masked and set a low-balance alert.
• What it can never do: create or delete API keys, change billing or payment methods, buy credits, register accounts, or see provider internals. Those scopes do not exist on this server.
• Spend cap: each connection gets a dedicated key with its own cap, $10 per 30 days by default, adjustable from $1 to $500. The cap is enforced on the server; when it is reached send_message returns an error until the window resets or you raise the cap at kymaapi.com/integrations.
• Tokens and keys: access tokens live 7 days, refresh tokens 90 days and rotate on use; connection keys are stored hashed (SHA-256) and can be revoked without touching your REST API keys.
• Prompt injection: model output that reaches your agent is untrusted text. Keep your client's confirmation prompt on for send_message, and do not let the agent paste keys or balances into other tools.
• Data: request logs are kept 90 days for billing reconciliation and abuse handling, then deleted. Kyma does not train on customer data. Privacy: kymaapi.com/privacy.

Skills

Three skills teach an agent how to use the tools well: pick-a-model (compare price, measured uptime and fit), check-spend-and-credits (balance, cap, remaining, reset date), send-a-test-completion (one billed call, only after an explicit Allow). They live in skills/ and install into any agent that reads skills: