Open-code tools

Open-source agentic coding harnesses and where they fit

Published

June 1, 2026

This course uses two harnesses: Claude Code and VS Code + GitHub Copilot. Both are commercial products tied to a provider. There is a third category worth knowing about: open-source coding agents — “open-code” tools — that you run yourself and point at whatever model you like.

What “harness” means here

A harness is the program that wraps a language model and lets it actually do things: read your files, run code, edit files, call tools, and loop until a task is done. The model is the engine; the harness is the car. The harness matters as much as the model — a great model in a clumsy harness is frustrating, and a good harness makes a mid-tier model genuinely useful.

Open-code tools vs the two we use

	Claude Code / Copilot	Open-code tools
Source	Closed, vendor-built	Open-source, community or vendor-released
Model	Tied to a provider (Anthropic / your Copilot model)	Bring your own — frontier API or local open-weights
Setup	Sign in, go	Install + configure + supply a key/endpoint
Why pick it	Polished, supported, low-friction	Flexibility, transparency, run fully local

Examples in this space include open-source CLI agents and editor extensions that accept any model endpoint (including a local Ollama server). The landscape moves fast, so treat specific names as examples, not endorsements.

Why it matters for this course

The point is conceptual transfer, not a third tool to learn:

Everything you learn in Claude Code — give the agent context, let it run and verify, inspect the diff, iterate — transfers directly to any open-code harness.
Open-code tools are the bridge between the open-weights models page and real work: they’re how you put a local, private, reproducible model into an agentic workflow.
If you ever need a fully local, private agentic setup (sensitive data, air-gapped, no per-token cost), an open-code tool pointed at a local model is the way to get it.

Note

We do not require you to install an open-code tool. It’s a stretch/awareness topic: know the category exists, understand it’s the same workflow with a swappable engine, and reach for it when privacy or cost rules out the commercial harnesses.

--- title: "Open-code tools" subtitle: "Open-source agentic coding harnesses and where they fit" date: "2026-06-01" --- This course uses two harnesses: **Claude Code** and **VS Code + GitHub Copilot**. Both are commercial products tied to a provider. There is a third category worth knowing about: **open-source coding agents** — "open-code" tools — that you run yourself and point at whatever model you like. ## What "harness" means here A **harness** is the program that wraps a language model and lets it actually *do* things: read your files, run code, edit files, call tools, and loop until a task is done. The model is the engine; the harness is the car. The harness matters as much as the model — a great model in a clumsy harness is frustrating, and a good harness makes a mid-tier model genuinely useful. ## Open-code tools vs the two we use | | Claude Code / Copilot | Open-code tools | |---|---|---| | Source | Closed, vendor-built | Open-source, community or vendor-released | | Model | Tied to a provider (Anthropic / your Copilot model) | **Bring your own** — frontier API *or* local open-weights | | Setup | Sign in, go | Install + configure + supply a key/endpoint | | Why pick it | Polished, supported, low-friction | Flexibility, transparency, run fully local | Examples in this space include open-source CLI agents and editor extensions that accept any model endpoint (including a local [Ollama](open-weights-models.qmd) server). The landscape moves fast, so treat specific names as examples, not endorsements. ## Why it matters for this course The point is **conceptual transfer**, not a third tool to learn: - Everything you learn in Claude Code — give the agent context, let it run and verify, inspect the diff, iterate — transfers directly to any open-code harness. - Open-code tools are the bridge between the [open-weights models](open-weights-models.qmd) page and real work: they're how you put a local, private, reproducible model into an agentic workflow. - If you ever need a fully **local, private** agentic setup (sensitive data, air-gapped, no per-token cost), an open-code tool pointed at a local model is the way to get it. ::: {.callout-note} We do **not** require you to install an open-code tool. It's a stretch/awareness topic: know the category exists, understand it's the same workflow with a swappable engine, and reach for it when privacy or cost rules out the commercial harnesses. ::: ## See also - [Open-weights & local models](open-weights-models.qmd) — the engines these tools run. - [Installing AI CLI Tools](install-cli.qmd) — setup for the harnesses we do use. - [Which AI model](which-ai.qmd)