Five engines, and through routes, almost any model.
Every AI company sells its own tool, and each one wants your work inside it. Agent Farm runs no model of its own: it drives the CLIs you already pay for, side by side, and lends Claude Code and OpenCode any other model through a route. Change the brain behind an agent and the way you work stays the same.
Claude Code, Codex, Gemini, Antigravity and OpenCode, each paid for the way you already pay. Claude Code and OpenCode can also run other models through a route: a provider's, a free one, or one at an address on your own machine. Keys stay in your OS keychain.
Switchboardengine → brain
A plan, a key or an app sign-in is each engine's own. A route is one address, one key and its models, lent to Claude Code or OpenCode when a session starts; it is never written into their files.
VII·1
Five engines
Claude Code, Codex, Gemini, Antigravity and OpenCode. What is read off disk says so; only one real test call proves an engine works.
Five agent CLIs means five installers, five sign-ins and five ways of being paid for. A file on disk can say signed in while every call fails, and a tool that only works with one of them decides your model for you.
How it works · pick a step
For a CLI it does not know, a profile runs any command line, keeps its output and reports its exit code. It is a command runner, never a sixth engine: no invented transcript, price or live view.
Settings · 2
Setting
Default
What it does
agentFarm.cliProfiles
[]
Command lines for a CLI Agent Farm does not know, with {prompt} and {args} filled in.
agentFarm.modelAliases
haiku, sonnet, opus, fable
Model names every picker offers; add a new one once. Codex's and Antigravity's lists are read from the engines themselves.
VII·1Four questions1 / 5
VII·2
Plug-in brains
Claude Code and OpenCode can run other models through a route: free or paid, from a provider or an address on your own machine. Five engines, any number of brains.
Claude Code and OpenCode are good programs, and the best model for a job is not always the one their maker sells. Pointing them somewhere else by hand means environment variables, and a variable left over in a window quietly sends a session to the wrong company.
How it works · pick a step
Removing a route first settles every saved setup that names it, and a launch that would land on a route that is gone is refused at every door.
Settings · 1
Setting
Default
What it does
agentFarm.contextWindowTokens
0 (auto)
The context ceiling. 0 = the window the engine states, else the Context of the route the session runs on, else the model's.
VII·2One route1 / 5
VII·3
Free-model radar
Set it up once as a schedule: every morning an agent checks which models are free today and adds new ones as routes, over the bridge.
A route is an address and a key you typed months ago. You want to know it still works before a night of work runs on it, without paying a turn every time a page opens.
How it works · pick a step
VII·4Free to know1 / 5
VII·5
Secret vault
Keys live in your OS keychain, open only to the agents you choose. A key's value never crosses the bridge.
Keys pasted into settings files end up in git, in screenshots and in an agent's context. Every model behind a route, every media service and every bot needs one.
How it works · pick a step
If something that looks like a key turns up in plain text in a config file, the health check tells you to move it into the vault.
Settings · 1
Setting
Default
What it does
agentFarm.mediaAuthSecret
FAL_KEY
The name of the key media steps use.
VII·5Into the keychain1 / 5
VII·6
Many sign-ins
Several sign-ins of Claude Code or Codex side by side, without signing out. An account is a config folder, not a copied credential.
Two or three plans on one machine meant signing out and in all day, the right browser profile, a fresh login, and sometimes a night of work on the wrong plan.
How it works · pick a step
An account can be switched off without being removed, and signing out clears its credentials but keeps its folder. Which account a session gets can also be left to a rule: see Account policies in Never stops.
VII·6Capture a sign-in1 / 5
VII·7
Switch the model
Change the model, effort or permissions of a running session from the Console. It takes effect from the next turn, and the conversation stays.
What it solves
Halfway through a conversation you know whether it needs the big model or the cheap one. Starting again to change it loses the context you built.
How it works · pick a step
Each turn's line ends with the model and effort that answered it, so a switch made meanwhile cannot rewrite the turn you are watching. An agent can switch too, with console.setModel.
VII·7Mid-conversation1 / 5
VII·8
Eleven doors
Start a session from the Console, a row, a right-click, a saved setup, the command palette, the bridge or Telegram: all eleven doors carry the same engine, model, account and permissions.
What it solves
A session can be started from many places, and a setting that one of them forgets is a session on the wrong model, the wrong account or with the wrong permissions, which you notice afterwards.
How it works · pick a step
From Telegram, /new developer starts a real Console session on your desktop with that agent's last setup, under the chat's daily budget.
Settings · 1
Setting
Default
What it does
agentFarm.launchDefaults
{}
The engine, model, effort and permissions every session starts with. An empty field adds no flag.
VII·8Eleven places1 / 5
VII·9
One agent, every engine
An agent's rules, duties and MCP servers are written out in the files Codex, Antigravity and OpenCode read, and kept in step in both directions.
What it solves
An agent's rules are written for one engine: Claude Code reads CLAUDE.md, the others read AGENTS.md and their own folders. Without a copy they start as strangers, and a copy kept by hand goes stale the day you change the original.
How it works · pick a step
The sync also says what Antigravity will load from that folder: which files, how big, any text that would arrive twice and any file past its 24,000-byte cut.
VII·9One instruction file1 / 5
VII·10
Any machine
Tasks, missions and their history travel between computers in git: one machine's paths are saved as tokens and open on the other machine's folders.
A new computer, or a second one, should not mean setting the farm up again. Files full of one machine's paths break on every other.
How it works · pick a step
One agent travels as an .agentpack, and another machine's sessions can be read here as they happen.
VII·10Saved as tokens1 / 5
Questions
Does Agent Farm have a model of its own?
No. It runs no model itself: it drives the agent CLIs you have installed, on your own plans and keys, and reads back what they wrote. What you pay those companies is unchanged.
Do I need all five engines?
No. One is enough. The ones you do not have stay in the pickers, greyed, with the one thing it would take to use them, and a Claude Code session and a Codex session can run side by side whenever you want both.
Can I use a model that runs on my own computer?
Yes. Give it a route at an http address on this machine, such as 127.0.0.1, and Claude Code or OpenCode will run it. OpenCode also has free models of its own that need no sign-in.
Why can only Claude Code and OpenCode use routes?
Because only those two can be pointed at another provider. Codex, Gemini and Antigravity each talk to their own maker's models, so for them there is nothing to route; you pick their model, effort and permissions instead.
What does a session on a plan cost on the page?
An engine on a plan reports tokens, not money, so Agent Farm shows tokens and says why there is no price. It never prints $0.00 for work that was paid for by a plan.
Can another engine carry on work one engine started?
Yes, by reading it: any agent can read any session, whichever engine wrote it, in the same five roles, and carry on. A resume itself always reopens on the engine the conversation was had with.