MCP server framework
Open your API to AI agents without giving it away.
AI agents reach products through MCP servers. Yours has to respect the same plans as your API: a free customer shouldn't get premium data through the agent door. NeuAira Conduit adds tiers, usage metering, limits and safe logging to any MCP server, so you can sell access instead of just serving it.
Developer preview A Python library today; a hosted gateway is on the roadmap.
from mcp_platform_core import ToolDefinition
tool = ToolDefinition(
name="get_forecast_premium",
input_model=ForecastInput,
min_tier="premium", # who may call it
cost_units=3, # what a call costs
cache_ttl_ms=300_000, # cache hits bill at zero
handler=get_forecast,
)Anyone can write an MCP server in twenty lines. Running one as a product is the hard part.
Who is calling? Are they on a plan that includes this tool? Are they over their limit? What did the call cost? The upstream is down: retry, or stop? Conduit answers all of that once, in a fixed, tested order, so each tool is just a plain function that calls your API.
What every call goes through
- IdentifyAPI key to customer record
- Check tierfree, premium or enterprise
- Cachereturn a recent answer at no cost
- Rate limitper customer, per minute
- Call your APIwith timeouts, retries and a circuit breaker
- Recordusage event, metrics and a redacted log line
Tiers bound to customers
Each API key maps to a customer and a tier. Each tool declares the lowest tier allowed to use it.
Usage you can bill
Every call emits an event with the customer, tool, cost, cache hit and duration, ready for your billing system.
Fair caching
Per-tool caching cuts load on your API, and cached answers are recorded at zero cost.
Resilient by default
Timeouts, bounded retries and circuit breakers stop a slow upstream from taking the server down with it.
No secrets in logs
Structured JSON logs with keys and tokens redacted at every depth, plus Prometheus or OpenTelemetry metrics.
Runs anywhere
Local (stdio) and HTTP transports, health and readiness checks, graceful shutdown and a Docker setup.
Proven on four different APIs
We've built working MCP servers on Conduit for weather, finance, GitHub and eBay data. Each uses a different way of signing in, from no key at all to OAuth.
How we work with you
- We build your MCP server on Conduit, wired to your existing plans and billing
- You run it in your own infrastructure; your API keys stay with you
- A yearly subscription covers updates, new MCP features and support
Questions about Conduit
How is this different from FastMCP?
Conduit is built on FastMCP. FastMCP helps you expose tools; Conduit adds what you need to run them as a paid product: customer tiers, cost per call, billing events and zero-cost cache hits.
Is it a gateway?
Not yet. Today Conduit is the framework you build MCP servers with. A gateway that sits in front of existing MCP servers is on our roadmap.
Who is it for?
Companies that already sell API access in plans, such as data, market-data and developer-tool providers, and want an MCP server that honours those plans.
What does it cost?
Usually a fixed build fee, then a yearly subscription. We'll quote after a short call about your API.
Make your API agent-ready.
Tell us about your API and plans. We'll show you what an MCP server with your tiers looks like.