klanex MCP server

Reliable execution for agent tool calls, as an MCP server.

klanex is an async execution layer between an agent's tool-use intents and real third-party APIs: it schema-gates hallucinated payloads before anything runs, absorbs rate limits and outages with retries/backoff/circuit breakers, encrypts credentials, keeps an audit trail, and reports back via signed webhooks. With this MCP server, an agent submits a call once — klanex makes sure it lands.

Failures come back written for the model: every error carries an llm_hint telling the agent exactly what to fix (or that klanex is already handling it), so agents self-correct instead of guessing.

Connect (hosted, recommended)

klanex's MCP server is hosted — no install needed if your client supports HTTP transports:

Endpoint	Environment	Keys
`https://api.klanexai.com/mcp`	production	`klx_live_…`
`https://api.sandbox.klanexai.com/mcp`	sandbox (free to try)	`klx_test_…`

Get an API key at klanexai.com — sandbox keys are free.

Claude Code

bash

claude mcp add --transport http klanex https://api.klanexai.com/mcp \
  --header "X-API-Key: klx_live_..."

Cursor / Windsurf / VS Code (mcp.json)

json

{
  "mcpServers": {
    "klanex": {
      "url": "https://api.klanexai.com/mcp",
      "headers": { "X-API-Key": "klx_live_..." }
    }
  }
}

Authorization: Bearer klx_... works too, for clients that can only set that header.

Connect (stdio shim)

For clients that only speak stdio (e.g. Claude Desktop), this package proxies stdio to the hosted endpoint:

json

{
  "mcpServers": {
    "klanex": {
      "command": "npx",
      "args": ["-y", "klanex-mcp"],
      "env": { "KLANEX_API_KEY": "klx_live_..." }
    }
  }
}

The shim routes by key prefix — klx_test_… keys go to the sandbox automatically. Set KLANEX_MCP_URL to override.

Tools

Tool	What it does
`execute`	Submit an HTTP call for reliable async execution: optional JSON Schema gate, retries/backoff/circuit breakers, idempotency keys, human approval gates, encrypted credentials. `wait_seconds` blocks up to 55s for the terminal result.
`get_execution`	Current status, attempts, target response, or classified error for an execution.
`list_executions`	The account's executions, filterable by status/time, paginated.
`replay_execution`	Re-run a terminal execution byte-exact — outage recovery without re-prompting the LLM.
`get_usage`	Plan and current-month usage/quota.
`list_connections`	Stored credential connections (token vault) to reference via `connection_id` — so secrets never pass through model context.

Why this beats calling APIs directly

Hallucinated payloads are rejected before they execute — pass payload_schema and mismatches return synchronously with a correction hint.
Transient failures are not the agent's problem — 429s, timeouts, and 5xxs are retried with backoff behind a circuit breaker; the agent does nothing.
A network blip can never double-execute a refund — idempotency_key makes submits safe to retry.
Destructive actions can wait for a human — requires_approval pauses execution for an approve/reject decision (API, dashboard, or Slack buttons).
Credentials stay out of model context — store them once as a connection; klanex injects them at execution time.