Kiln
Author: Kiln AI
Description: Kiln is a local-first desktop app and open-source Python library for building, evaluating, and optimizing AI systems. It supports evals, agents, RAG, fine-tuning, synthetic data generation, dataset management, and tool integrations including MCP. Developer entrypoint: the Python package `kiln-ai` (install via `pip install kiln-ai`). The repo also contains an MIT-licensed core library and server (REST API), plus a source-available desktop app.
Stars: 4.7k
Forks: 346
License: Other (GitHub repository license field). README notes: Python Library (libs/core) MIT; Python Server (libs/server) MIT; Desktop App is source-available and free to use.
Category: Open Source
Overview
Installation
pip install kiln-aiFAQs
Can I use MCP servers with multiple clients simultaneously?
Yes. Stdio-based servers like Kiln run isolated per client, so Claude Desktop, Cursor, and Augment Code each launch separate processes without conflict. Network-based MCP servers over HTTPS allow multiple clients to connect simultaneously to a shared endpoint using OAuth or API keys for access control, enabling enterprise-scale tooling.
Can you integrate multiple MCP servers with Claude for Desktop?
Yes, Claude Desktop supports multiple MCP servers simultaneously through its JSON configuration file. Add entries under the "mcpServers" object, each with a unique key, command, and transport settings. All configured servers load globally across sessions and appear in the conversation input box. Restart Claude Desktop after modifying the config to activate new servers.
How do I implement distributed tracing for MCP tool executions?
For MCP tool executions, install OpenTelemetry SDK and configure a TracerProvider with BatchSpanProcessor and OTLP exporter before importing your MCP framework. Frameworks like FastMCP auto-generate spans for tool calls, resource operations, and server interactions without code changes. Use backends like Jaeger or Zipkin to visualize latencies and debug agent loops via waterfall views.
How can I implement least privilege access for MCP servers?
Least privilege for MCP servers requires capability-scoped permissions, OAuth2 tokens with fine-grained scopes, and runtime enforcement via policy engines like Open Policy Agent. Use mutual TLS for transport security, container security contexts to drop unnecessary Linux capabilities, and structured logging for audit trails. Test with MCP CLI probe commands and fuzzing tools, then align with SOC2 compliance through automated scans.
What are the differences between remote and STDIO MCP servers?
Remote MCP servers run on separate machines and use HTTP or SSE transports for network communication, enabling multiple clients to share one service but adding latency. STDIO servers execute locally as child processes spawned by the client, using standard input and output streams for zero-network-overhead communication, with one server instance per client session.
What metrics should I track to optimize MCP server performance?
Track request latency from tool invocation to response completion, concurrent connection counts, and stdio transport buffer utilization to identify bottlenecks specific to MCP's tool-calling pattern. Monitor error rates per tool endpoint and measure time-to-first-token for streaming responses, as MCP servers often handle multiple simultaneous tool requests from IDE agents that expect subsecond response times for code-aware operations.