A tiny coding agent built on the old OpenAI completions API (/v1/completions, gpt-3.5-turbo-instruct). No tool-calling API, no SDK, no dependencies — just text in, text out, and a manual JSON parser that drives a few CRUD tools and a shell.

It is intentionally a single-file agent: agent.js, ~150 lines, no comments. Think of it as a GPT-2/GPT-3 era ReAct loop.

1. The runner sends a prompt to /v1/completions.
2. The model is told to reply with one JSON object per turn describing a tool call.
3. The runner extracts the JSON, dispatches the tool, and feeds the result back as a line beginning with OBSERVATION:.
4. Repeat until the model emits {"tool":"done", ...} or MAX_STEPS is hit.

A \nOBSERVATION: stop sequence keeps the model from hallucinating its own tool results.

TASK: <your task>
{"tool":"read","args":{"path":"package.json"}}
OBSERVATION: { ... file contents ... }
{"tool":"shell","args":{"cmd":"node -v"}}
OBSERVATION: v22.4.1
{"tool":"done","args":{"answer":"Node 22 detected."}}

create and update are the same operation under the hood — the split exists only so the model can be explicit about intent.

Requires Node 18+ (uses the built-in fetch).

git clone https://github.com/ahmadawais/clai-old.git
cd clai-old
npm install   # no-op, there are no deps

export OPENAI_API_KEY=sk-...
node agent.js "list every .md file in this repo and summarise each in one line"

Or interactively:

node agent.js
task> _

Or via the bin:

npm start

shell runs whatever the model emits, with your shell, in your CWD. There is no sandbox, no allowlist, no confirmation prompt. Run it in a throwaway directory or a container. Don't point this at a repo you care about.

Because sometimes you want to remember what an agent looked like before tool-calling APIs, before SDKs, before frameworks — when the whole loop fit in one file and the model was just a text predictor.

MIT