Explainer · 3 November 2026 · 6 min read

What is a coding harness?

A coding harness is the software that turns a language model into an agent: it decides what the model sees, what it is allowed to do, and when the task is finished. It is also the part that decides your bill.

The three layers people mix up

When someone says “the AI wrote my code”, three separate things are involved:

• The model. A network that predicts the next tokens. It has no memory, no file system and no ability to run anything by itself.
• The provider. The API that serves the model and bills per token. DeepInfra, OpenAI and Anthropic are providers.
• The harness. The program on your machine that builds prompts, calls tools, edits files, runs tests, asks for permission and loops until done. opencode, Cursor, Claude Code and Codex CLI are harnesses.

The model is the engine. The harness is the car — steering, brakes, dashboard and fuel gauge. Swap the engine and the car still drives; swap the car and the experience changes completely, even with the same engine.

What a harness actually does

• Context: chooses which files, diffs, errors and instructions go into the prompt.
• Tools: exposes read, write, search, shell and test commands to the model.
• Loop: decides when to retry, when to stop, and how to recover from a failed command.
• Permissions: decides what runs automatically and what needs approval.
• Accounting: decides how many model calls a task takes — which is what you pay for.

That last point is why two harnesses on the same model can differ by 3–5× in cost. A harness that re-sends the whole repository every turn burns tokens. One that retrieves narrowly and caches context does not.

Common harnesses in 2026

Why the harness matters more in India

Most harnesses assume a US payment method. A subscription means an international card, a dollar conversion, and a charge that moves with the exchange rate. If your card fails, the harness stops — not because the model is unavailable, but because billing is.

The second India-specific issue is distance. A model served from Virginia adds latency to every step of a loop that may run thirty times. Serving from India or Singapore is not a marketing detail; it is how the agent feels.

The third is the model menu. Harnesses that let you pick any model also let you pick a expensive one by accident. A harness that hides the menu can route cheap work to cheap models and keep the bill predictable.

Where DoneCorner fits

DoneCorner is not a harness. It is the layer between your harness and the model: an OpenAI-compatible endpoint, a key, and routing to open-weight models under one name, dc-llm.

• Works inside opencode and Cursor — the harness stays yours.
• ₹200 a month, first month free, UPI, no dollar card.
• Served from India and Singapore.
• No model picker: you ask for dc-llm and we choose the model.

If you already use opencode, the setup is one command: npx dcorner setup.

A checklist before you choose a harness

1. Can you pay for it? UPI or an international card — check before you invest a week.
2. Does it lock the model? A harness that only works with one provider is a single point of failure.
3. Where is inference served? Distance shows up as latency in every loop step.
4. What does it do with your code? Ask whether prompts are retained or used for training.
5. What does a normal task cost? Count model calls, not messages. Twenty steps is twenty billable calls.

Questions people ask

Next: what AI coding agents cost in India.