Command Code raised $5M and shipped taste-1, the first coding agent that learns from every accept, reject, and edit

Command Code closed a $5M seed and shipped taste-1, a continuous-RL layer that watches every prompt, accept, reject, and edit, then auto-generates project-level skills from the signal. The agent lives in the terminal. 29,000+ developers are running it. 320 releases have shipped. The $1/mo Go plan comes with $10-40 in free credits that stretch further on open models: 4× on DeepSeek V4 Pro, 2.7× on MiniMax M3, 99% off on MiMo V2.5 Pro.

The price is the headline. The learning loop is the story.

What taste-1 actually does

Every time you accept a diff, reject one, or edit one before merging, taste-1 records the signal. The signal gets distilled into a project-level skill, a markdown file that captures a convention. "pnpm over npm." "Tabs over spaces." "Prefer vitest." "Lowercased -v for CLI version." The next session opens with the convention already in place. No rules to write. No prompts to maintain.

The team version shares the skills. npx taste push publishes a project skill. npx taste pull installs one. Skills are open files in the repo. They travel through PR review like any other code. The first team to standardize on a shared taste library moves faster than the team that does not.

The harness, not the model

Command Code's thesis is direct: "open models can't tool-call" is a harness problem, not a model problem. Stop using open models with coding agents not built for them. Their harness closes the gap. David Thyresson (PWV, RedwoodSDK contributor) shipped multiple CLIs and a full Redwood app for $2 on DeepSeek Flash through Command Code. He had to double-check he was still on the open model. The output felt like the closed one.

That is the structural claim worth sitting with. If a $1/mo terminal agent on a 4×-credit open model produces diffs a senior engineer would accept, the cost curve for software teams has just bent. Cursor is $20/mo. Copilot is $10/mo. The Go plan is $1/mo with credits that stretch it 4× to 40×. The IDE market is the wrong place to look for the next cost-quality shift in coding agents. The terminal + open models + taste learning is where it is happening.

The catch

Three things this does not solve.

The terminal-only design is a real ceiling. Most engineers will not switch from their IDE. The agents that own the IDE layer (Cursor, Copilot, Claude Code inside VS Code) keep the seat. Command Code wins the terminal-first segment and the open-model-first segment. Those overlap. They are not the same market.

The taste signal is only as good as the feedback. If you stop reviewing diffs carefully, taste-1 stops learning. The cost of attention is what powers the loop. Teams that rubber-stamp the agent's output get a fast agent that gets dumber.

Skills are shared by default across the team that pulls them. The convention sharing is the feature. It is also a coordination surface. If your team ships one style of TypeScript and your partner ships another, the merged skill library gets noisy fast. Set the convention boundary before you push the first skill.

What we are doing this week

We are installing Command Code Go on two of our internal dev machines today. The test workflow: an internal CLI we ship for client setup automation. The eval is simple. After one week of use, count the corrections we still make to the agent's output. If the correction rate drops below where Claude Code sits for the same task, we extend taste-1 to the second workflow by end of month. If it does not, we keep it as a side tool and revisit at v1.0.

Source