Trpc Agent Go

What trpc-agent-go does

trpc-agent-go provides a Go-native runtime for building agent systems rather than a single-purpose chatbot or standalone MCP endpoint. It combines LLM agents, tool calling, graph-based workflows, session state, memory, knowledge retrieval, evaluation, and telemetry. MCP is one part of its tool ecosystem: an agent can use MCP tools alongside ordinary Go function tools and other supported capabilities.

The framework is aimed at applications that need service-friendly execution. Streaming runs, context cancellation, concurrent workflows, persistent state, and OpenTelemetry instrumentation are part of the surrounding runtime. Supported application patterns include customer support, data analysis, DevOps automation, business process workflows, and research or document question-answering systems.

How it works

Agents run through a runner that receives a user, session, and message context. Multiple agents can be arranged in chains, parallel execution, or cycle-based workflows. GraphAgent adds typed workflow nodes and conditional routing for more explicit orchestration.

Tools may be ordinary functions converted into callable agent tools, MCP tools created from a server connection, web search, code execution, or custom services. Memory can be attached at the runner level or exposed through tools, allowing agents to retain context across sessions. The framework also includes knowledge retrieval, artifacts, agent skills, and integrations for A2A and AG-UI applications.

Agent Skills use folders containing SKILL.md files. A filesystem repository can load skills from local roots or from an HTTP(S) archive, with local caching controlled by SKILLS_CACHE_DIR. A separate evolution service can review completed sessions, apply quality gates, and publish reusable skills for later turns.

Setup and configuration

Use the Go module trpc.group/trpc-go/trpc-agent-go as the application framework. A typical application creates a model-backed agent, registers tools, constructs a runner, and invokes the runner with user and session identifiers. MCP support is illustrated through an MCP server connection passed to the MCP tool adapter.

Memory services can be created in memory and supplied to the runner. Langfuse examples add tracing attributes to runs, while OpenTelemetry provides the broader tracing and metrics integration. For skill repositories, set SKILLS_CACHE_DIR when the default cache location should be replaced.

When skills are connected through an LLM agent with code execution enabled, the response processor can be disabled so fenced Markdown code in assistant output is not automatically executed while skill_run is available.

Tools and capabilities

• Function tools created from Go functions
• MCP tools connected through an MCP server connection
• Web search, code execution, and custom service tools
• Chain, parallel, cycle-based, and graph workflows
• Sessions, memory, artifacts, and knowledge retrieval
• Reusable SKILL.md workflows with local or archived repositories
• Session review and gated skill publication through agent self-evolution
• Evaluation runs with configurable repetition and result status
• OpenTelemetry tracing and metrics, with Langfuse examples
• A2A, AG-UI, and MCP protocol integration

Limitations and notes

The project is a Go framework for embedding agent capabilities in an application. Its MCP role is tool integration within that framework, so developers still need to create the surrounding agent, runner, and connection setup. Skill execution requires particular care when code execution is enabled, especially if assistant-generated fenced code should not run automatically.