Qodex local-first coding agent terminal UI

A Coding Agent That Lives on Your Machine

Qodex is a local-first coding agent CLI written in Go, actively developed by COS Q Network.

It pairs a terminal UI with streaming token rendering and an OpenAI-compatible local model endpoint — running a single locally hosted Qwen Coder model by default through llama.cpp, with vLLM and SGLang as advanced backends.

Explore Qodex on GitHub

A Self-Contained Agent Runtime

Qodex agent loop, context builder, skill loader, tool registry, and local model endpoint architecture

Every qodex invocation runs through the same core: a command entry point, the agent loop, a context builder, the skill loader, and a tool registry that talk to a local OpenAI-compatible endpoint.

The default runtime is a managed llama.cpp server serving a Qwen Coder model, installed and started during qodex setup. A validated tool executor fronts every action, and all sessions, tool events, and approvals are persisted in SQLite with WAL mode.

Backend capability detection runs at startup, so SSE token streaming turns on automatically when the endpoint supports it.

Terminal-Native Workflows

Qodex terminal chat, streaming tokens, inline approvals, and command palette features

The Bubble Tea chat UI renders tokens as they stream, with inline diff previews, a spinner, an error panel, responsive sizing, @ file autocomplete, / command autocomplete, transcript search, and a searchable command palette.

Qodex ships 98 built-in tools spanning file and symbol search, LSP code intelligence, Git workflows, CMake/clang/make/.NET/Java toolchains, package managers, archives, Docker/QEMU, ADB, and system administration — plus 31 built-in skills with skill.toml metadata and a routing engine.

Approval gates front write, shell, and network tools with risk summaries, compact diffs, once/session/always scopes, and a full audit trail.

Local-First Setup and Managed Backends

Qodex setup wizard, GGUF model download, llama.cpp server, and local-first workflow

qodex setup walks through backend choice, installs the managed runtime, downloads a catalog GGUF model with resumable transfers, waits for /v1/models readiness, and writes .qodex/config.toml only after everything succeeds.

There is no cloud dependency by default — your model, your session store, and your approvals all live on your machine. On Linux and macOS the backend is managed automatically; on Windows, WSL2 is the recommended path, while native Windows can target a manually managed OpenAI-compatible endpoint.

Sessions resume with full tool-history reconstruction, and Git operations run through the normal approval and audit workflow.

Actively Developed and Hardened

Qodex active development, testing, releases, and cross-platform builds

Qodex is under active development with a growing Cobra CLI — setup, init, config, chat, run, review, doctor, skills, sessions, serve, up, down, status, models, reset, version, and completion.

MCP clients plug in over stdio and Streamable HTTP with qodex mcp doctor verification, and repository instructions are interoperable across AGENTS.md, CLAUDE.md, GEMINI.md, Cursor rules, and GitHub Copilot.

Qodex is released under the Apache License 2.0. Read the full license on GitHub.

Explore Qodex

Follow the agent loop, terminal workflows, skills, security model, and ongoing development in the public repository.

Visit Qodex on GitHub