Pinloop treats the coding agent as the user interface

Running it bare prints instructions for an agent

The unusual decision is that the primary user of this interface is not the person at the keyboard.

Run the command with no arguments and it prints instructions written for a coding agent. There is a second command that gets your agent to walk you through a more structured onboarding flow, and a third that hands it the full usage instructions.

So there are three distinct modes: bare invocation for a quick instruction sheet, an onboarding flow, and a reference document. The guide is the only one that works without an account, which matters because it is also the one you might want before deciding to sign up.

The reasoning is visible in how the project is introduced: it is a job board built for your coding agent, run entirely from a terminal, and the demonstration is a sentence you paste into Claude Code, Codex, Cursor or something similar rather than a list of steps to follow yourself.

Everything except the guide talks to the project's servers, so this is a thin client over a hosted service rather than a local search tool.

Install is one sentence your agent can paste

The suggested first action is to hand your agent a single instruction:

Run npm install -g pinloop, then run pinloop welcome and follow the instructions.

Framed that way, the agent does the install and then reads the onboarding output itself, which is the whole premise of the tool in one line.

The requirement is Node 22 or newer, and the install is a global package install:

npm install -g pinloop

That is the entire setup surface. There is no configuration file, no index to build, no database to provision and no API key to paste.

The absence of local setup is worth noticing, because it is a direct consequence of the service-backed design. Since postings, preference files and judging all live on the provider's side, there is nothing to configure locally beyond the binary itself.

Which also means the tool is only as available as the hosted service is. The repository itself is small: a source directory, a build configuration, a package manifest, a licence and a few assets.

Login happens in a browser and there is no password

Account creation is a single command, and the authentication model is unusual enough to be worth describing exactly.

Login goes through a browser, and there is no password. That is a deliberate simplification rather than a missing feature, and it has the practical effect that there is no credential for you to store, reuse or leak.

It also means the login flow cannot be completed from a headless machine without a browser, which is a real constraint for anyone running this in a container or over a remote session.

The account is not optional. Everything except the guide needs one, so the three commands have different availability depending on whether you have signed in.

Two plan levels exist, and the boundary between them is not simply volume.

The free plan lets you search with words and filters to count how many postings match before pulling any of them, which is the one genuinely free and genuinely useful capability: it tells you whether a search is worth continuing before spending anything. It then allows pulling five postings per day, judging 150 postings per month through LLM calls, and 50 semantic searches per month.

Upgrading is its own command rather than a flag on the others.

Judging is metered, and your own agent changes the arithmetic

The most interesting line in the pricing is a parenthetical.

The free allowance is 150 postings judged per month through the service's LLM calls, with a note that you get infinite postings if you use your own agent instead.

That parenthesis is the whole cost model in miniature. The hosted judging is metered because the provider pays for the model calls; running the same judgement with an agent you already pay for removes that cost, and the limit stops applying.

This reframes what the paid tier is buying. On the higher plan you get 1500 postings judged per month rather than 150, and 1500 postings pulled per month rather than five per day.

So the tier difference is roughly a tenfold increase on both pulling and judging, plus the automation described below and unlimited semantic search instead of fifty a month.

The semantic search limit is what shapes interactive use. Fifty a month is comfortable for occasional exploration; it is the number that would run out first if you were using this as a daily tool rather than a periodic sweep.

Routines keep scanning after the laptop is closed

One capability only exists on the higher plan and is not about volume at all.

Routines let you set up scans that run and save postings automatically on the provider's servers, either on a fixed schedule or triggered when new postings arrive.

The significant detail is the last clause: this continues even when your laptop is closed. Everything else about the tool is interactive, agent-in-the-loop, run from a terminal you are sitting at. Routines are the one feature that runs without you.

That is the strongest argument for the higher plan and the clearest line between the two tiers, because it changes the shape of the workflow rather than its speed.

For a job search, which is largely a matter of noticing something before other applicants, a background scan that saves matches without you being present is qualitatively different from running a query when you remember to.

It also means postings accumulate on the service rather than on your machine, which is worth remembering when you later want to delete what you have uploaded.

Two of the four dependencies exist to animate progress

The package manifest is small enough to read as a description of the tool's priorities.

Four runtime dependencies are declared. One is a command line argument parser. One adds colour to terminal output. The other two are about showing progress: one updates a single line in place rather than scrolling, and one provides spinner characters.

So a quarter of the dependency budget goes to telling the user that something is still happening. For a tool whose operations are network calls against a large index, and whose selling point is an agent waiting on results, that is a defensible allocation rather than decoration.

Development dependencies are just the Node type definitions and the TypeScript compiler, and the build script is a single compiler invocation with no bundler, no transpiler configuration and no code generation step.

The published files are the compiled output, the sources, the assets and a demo recording, and the binary points into the compiled directory.

The licence is MIT and explicitly covers the command line only, which is the correct scope but worth noticing: the service, the postings and the judging are not covered by it. Privacy and terms pages are linked alongside the project site and community.

Editorial conclusion

Pinloop suits someone who wants an agent to hold the whole search loop rather than copy job descriptions into a chat window by hand, since the postings, the preference files and the judging are all reachable from one terminal. It suits you less if you only want a list of jobs to read, because the paid tier is sold on judgments and automation rather than on browsing. Before uploading a resume, read the privacy and terms pages and decide whether you want that file held on someone else's servers, and note that the MIT licence covers the command line tool only, not the service behind it.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. pinloop-ai/pinloop-cli on GitHub
4. Project website
5. README