Four command-line agents, and what each is good for
Last updated: 2026-09-22
Short answer: all four read files, run commands and edit code on your machine. The difference isn't what they can do — it's who makes them, what they're good at, and what they demand of your machine before they'll start.
| Tool | Origin | Good at | Needs before install |
|---|---|---|---|
| Claude Code | Anthropic, official | Long context — pulling in a large body of code or documents at once | Node; desktop build available |
| Codex | OpenAI, official | Writing and changing code | Node; desktop build available |
| Hermes | NousResearch, open source | General purpose — a bit of everything | Python 3.11+ |
| OpenClaw | Open-source personal assistant | Personal assistant with an app and a dashboard | Node 22 or newer |
One thing that's easy to miss
All four are separate programs. Each reads its own config file and keeps its own model settings. So "install a different one and see" isn't the whole job — you also have to tell it which model to use and which address to call.
How to fill that in, and what to check when it doesn't work, is in attaching a key you already have.
Choosing by your machine
If you'd rather not install a runtime, take a build that ships a desktop installer — dependencies come with it, and setup is a double click.
If you already have a development environment, all four install fine. Then look at the other condition: they require different versions, and those requirements move upward as the tools update. Reading the requirement first beats reading an error afterwards.
One trap worth naming: if you use a version manager (nvm and friends) and switch between versions, a tool that worked yesterday can refuse to start after you land on an older one. It usually tells you which version it wants — switch back to it.
Choosing by task
- It needs to read a lot of code or documents before acting — the long-context one saves you from chopping the material up first.
- It's code, writing or changing it — the code specialist is the better fit.
- You want a general personal assistant — the open-source pair lean that way, and one of them ships an interface.
You don't have to pick one
Nothing stops you running several. A common setup is two on the same machine: one for work that needs to read a lot, one for pure code.
They can share the same model credit — configure it once and point each tool at it.
Next
- Attaching a key you already have — how each tool connects, and what to check when it doesn't
- Checking that it actually finished — it says it's done; how do you confirm
- Agent platform roundup — the same choice from a different angle