Overview
Yagent is a local-first coding agent written in Go. It connects to a configured inference server and provides the surrounding workflow: exploring a repository, using tools, editing files, running checks, and preserving sessions for later work.
I built the agent loop, memory, orchestration, and tools directly, without an LLM framework. A terminal interface and a plain REPL share the same runtime. Local inference servers are supported alongside explicitly configured cloud endpoints, so the workflow can be used with different model backends.
Much of the project focuses on making changes inspectable and recoverable. Yagent can show diffs, request approval for edits, and preserve Git checkpoints for undo. Repository indexing uses Tree-sitter for structural navigation, helping the agent find symbols and understand relevant code before changing it. SQLite-backed sessions support resuming, searching, and exporting previous work.
Verification is part of the workflow. Tasks can declare checks that must pass before completion, and the toolset includes diagnostics, targeted tests, and runtime smoke checks. Longer tasks can use goals, checkpoints, and reusable playbooks.
Yagent is an evolving developer tool, and results depend on the chosen model, its configuration, and the quality of the task's checks. My focus is on building a useful working environment around those models: one where the user can follow the work, review the result, and recover when a change goes wrong.
Technologies: Go, SQLite, Tree-sitter, Bubble Tea, MCP
Project Tags: