Replit Agent

You will love Replit Agent for the whole loop in one tab: it sets up the project, writes the code, runs it, and publishes to Replit hosting without you installing anything, which is the daily trait that decides it. For a weekend idea that is the fastest path from nothing to a URL on this board. For a repository you already have it is the wrong shape: the agent expects to own the project.

Pick it for prototypes you need to show someone tomorrow. Pick Claude Code for an agent inside your own repository, and Lovable if the prototype is mostly a front end.

The meter is the worst thing. The billing doc explains that a checkpoint captures completed work and that complex features bundle into a single higher-cost checkpoint at an effort-based price; it gives no number and points to an interactive demo. Plan mode, which changes no files, is billed too. The price of a task is a function of how confused the agent got, and the function is not printed.

Set a spending limit before the first prompt and treat plan mode as paid. What it does right: a checkpoint is also a restore point, so a bad run can be rolled back rather than untangled by hand.

Documented: the agent operates a real browser against the running application to test its own output, which closes the loop between edit and observed behavior and is the rarest property on this board. Not documented: which model Auto mode selects, how context is gathered from a project, or how an edit is applied; the pipeline between prompt and diff is opaque with a browser at the end.

No benchmark is claimed, so reproducibility is nil in the formal sense; a run cannot be repeated with the same model because the model is not named. The observation: the verification is legible and everything it verifies is not.

Generated code lives on Replit hosting with a Replit database, so switching cost grows with every project shipped, which is the moat: not the agent, the address. Agent 4 in March 2026 put the design canvas and parallel task execution behind Pro and Enterprise, which is a company tiering its best features upward once the free funnel has done its job.

The likely acquirer is a hyperscaler wanting a consumer on-ramp with hosting attached; the likely pivot is toward the enterprise seat. Position: long the platform, short today's price, and keep a copy of your code.

The demo is an app deployed in one session. Procurement: Pro is $100 per user, so $6,000 a month for sixty before Agent usage, and Enterprise adds SSO and single-tenant environments at a quote. The code lives in Replit's workspace rather than our repositories, so review, CI and retention all happen on someone else's terms, and that is the line the questionnaire stops on.

Onboarding is nil, which is the point and the risk. Approved with conditions: a prototyping budget for the product team, not the engineering org, and nothing customer-facing ships from it.

The model is whichever one the vendor picks, and I cannot bring a key, run anything local, or read a line of the agent; a closed cloud workspace with a meter I cannot read is the opposite of my shelf. What I can bend: an MCP client, so my own servers reach into the workspace, and real git operations, so the code reaches out.

The practical move is to treat it as a scaffold generator: let it build, push the repository out, and finish with a tool I control. A fork is unthinkable; an exit is one push away, which is the only door I need.