Claude Code

Claude Code reads the repo, runs the commands, and keeps going until the tests pass, without you feeding it paths, which is the trait that decides it: on a real codebase you hand it a failing test and come back to a diff, not a question. The catch is the bill: there is no free tier, and Max starts at $100 a month, which is a decision, not an impulse.

Pick it if you will pay for it and want the agent to do the finding as well as the fixing. Pick Aider if you want to control the bill and watch every edit land.

The Bash sandbox is off until you run /sandbox, and even on it leaves file tools and MCP servers unconstrained, so a bad rm does its damage in your real repository before you read it. Pay-as-you-go through an API key bills a subagent that spins for an hour for the hour, and it has no way to know it is spinning. Two meters, one on your disk and one on your card, and neither has a floor.

Give it a worktree and a spend alert on day one. What it does right is multi-file editing that lands as one coherent change, with hooks that make verification deterministic rather than optional.

Claude Code documents two components worth noting. 1. Context is gathered by the agent through file reads and shell commands rather than a precomputed index, which means retrieval quality is a property of the model's search behavior and moves with the model, not the tool. 2. Verification is delegated to hooks, which makes the step reproducible: the same command runs after every edit whether the agent remembers or not.

The dataset lists no benchmark, so the capability claims are asserted rather than demonstrated, and the second design choice is the one a reader can verify at home. Legible engineering, unaccompanied by numbers.

Claude Code is a channel, not a company, and that is the point. The vendor that trains the model ships the agent, which cannot be acquired out from under you, and which has no cap table separate from the one that funds the model. Max from $100 a month says the buyer is willing, and willingness at that price is pricing power a wrapper cannot copy.

The risk is not survival but priority: a channel gets the roadmap the model needs, not the one you ask for. Position: long, with the caveat that you are buying a model subscription that happens to have a CLI.

The demo is a terminal that fixes a failing test. Team standard seats are $20 per seat per month with SSO; audit logs arrive only at Enterprise, priced as seat plus usage, which is the tier a security questionnaire will require. Sixty engineers on standard seats is $14,400 a year before anyone hits a limit, and agentic workloads are exactly the ones that hit limits.

Onboarding is an npm install and a login. Approved with conditions: Enterprise tier, hooks enforced by policy so verification is not optional, and a monthly usage report by team.

I cannot read the source, and local_models is false, so it never runs on my box and never runs offline. What keeps it installed: MCP servers in a JSON mcpServers block I version, and hooks that run my own scripts around tool calls, so I can lint, log or block anything the agent tries before it happens. That is a lot of surface for a closed tool.

If Anthropic dropped it tomorrow, nothing survives; there is no fork, only the config files I wrote for it. Grudging respect for the hook design, which others copied. Zero for the license.