Solutions
·AI Agents
Build agents that
reason, plan, and act.
Access models with best-in-class function calling and tool use -- Kimi K2.5 with Agent Swarm, Claude 4 Sonnet, DeepSeek-V3.2. Native support for MCP protocol, structured outputs, and multi-step reasoning. Plus an upcoming API, SDK, CLI, and MCP server so your agents can provision their own compute.
Agentic Capabilities
Models and tools built for agents.
Function calling
Define tools as JSON schemas. Models like Claude 4, Kimi K2.5, and DeepSeek-V3.2 reliably call your functions with structured arguments -- no prompt hacking.
MCP protocol
First-class support for the Model Context Protocol. Connect agents to external tools, databases, and APIs through a standardized interface.
Multi-step reasoning
Models that plan, execute, observe, and iterate. Chain tool calls across multiple steps without losing context or hallucinating intermediate results.
Structured outputs
Force JSON schema compliance on model outputs. Parse agent decisions, tool calls, and state transitions without brittle regex extraction.
Agent Swarm patterns
Kimi K2.5 supports native multi-agent orchestration. Spawn sub-agents for parallel tasks, merge results, and coordinate complex workflows.
Programmatic compute (coming soon)
Upcoming REST API, Python/Node SDK, CLI, and MCP server so agents can provision GPU instances, run workloads, and manage infrastructure autonomously.
Agentic Models
Models built for
tool use and planning.
These models excel at function calling, multi-step reasoning, and autonomous task completion.
Kimi K2.5Moonshot AIAgent Swarm, multi-agent orchestration
Claude 4 SonnetAnthropicTool use, long-context reasoning
DeepSeek-V3.2DeepSeekCode agents, structured outputs
Gemini 2.5 FlashGoogleFast tool calling, multimodal
GLM-5 / Qwen3Zhipu AI / AlibabaMultilingual agentsHow It Works
Three steps to an AI agent.
Define your tools
Describe your agent's capabilities as function schemas -- API calls, database queries, file operations, or any custom tool. Pass them to the model via the Inference API.
Pick an agentic model
Choose Kimi K2.5 for multi-agent swarms, Claude 4 for complex reasoning chains, or DeepSeek-V3.2 for code-heavy workflows. All via one OpenAI-compatible endpoint.
Run the agent loop
The model reasons, calls tools, observes results, and iterates. Monitor token usage and costs in real time. Per-token billing means you pay only for the reasoning your agent actually does.
Build your next agent on Runcrate.
Access agentic models with function calling, MCP support, and structured outputs. Per-token pricing, no credit card required to start.