Nanocoder

The deciding trait is how cheap it is to change your mind. Provider, model and mode are all flags, so trying a cheap model on a boring task and an expensive one on a hard task costs you a keystroke instead of a config rewrite. If you are cost-conscious, that habit saves more than any single feature here.

What you notice next is that it is small and young, and behaves like it. Pick it if you enjoy tuning. Pick OpenCode when you want the terminal agent with more people leaning on it.

The exposure is ordinary and real. Writes land in the working tree, commands execute where you launched them, and no containment is recorded, so the boundary is your shell. The daemon makes it worse in one specific way: a process persists per project, so a bad state does not clear when you close the terminal and you have to know it is there to clear it.

What it does right is interception. Skills carry event triggers, so an operator has a documented place to log or refuse before something runs.

1. Aiming a coding agent at modest open weights is a legitimate and difficult research position, because the binding constraint stops being reasoning and becomes whether the model emits a patch the harness can apply. 2. That failure is silent: an unapplied edit reads like a refusal.

2. No evaluation addresses it. No benchmark, no apply-rate figure, no comparison between model sizes, and the documentation is a repository page. The premise is interesting and entirely undemonstrated.

There is no cap table to worry about because there is no company. That eliminates the acquisition risk this board is full of and replaces it with volunteer attrition, which is slower, less announced and harder to plan around. Nothing is gated behind a paid tier, so there is no commercial pressure to keep shipping and no commercial reason to stop.

Moat: goodwill. Likely path: steady community maintenance, or a gradual quiet as contributors move on. Position: adopt it freely at the individual level, and never make it the only path into a workflow a team depends on.

The data-residency answer sells this better than any feature does. Point it at a model we host and the code never crosses our boundary, which turns a six-week review into a diagram. Licensing is free, so the only line item is the hardware we already own.

What is missing is administration. There is no console, no provisioning and no central log, so policy across sixty machines is a wiki page and hope. The non-interactive command does slot into a pipeline. Approved with conditions: internal models only, and one team owning the configuration.

This is the shape I keep asking for. npm install -g @nanocollective/nanocoder, point it at Ollama, and the loop closes with no account anywhere. Skills bundle commands, tools and subagents into something I can version alongside the project, and MCP servers attach without a wrapper.

Permissive licence means a fork stays legal and stays mine, and the codebase is small enough that a fork is realistic rather than theoretical. The gap is polish, which I will take over a login screen every time. Best cost story on this board.