Guards stop what an agent should not do, and nothing about your work comes to us.
Your work stays on your machine. The free version never contacts us; a paid licence checks its key once a day, and that is the whole message. Your engines talk to their own providers, and anything else Agent Farm reaches is a service you switch on. On the machine, guards stop an agent before a command runs, not after: one for the shell, one for disk searches, one for permissions, one for keys and one for the browser.
Never in it: your code · prompts and answers · transcripts · file, folder and project names · account names · costs and tokens · which features you use, or how often.
Your engines talk to their own providers, as they always do. Anything else Agent Farm reaches is a service you switch on yourself: a voice for the briefing, a media model in Studio, Jev for Gatekeeper, Telegram.
Built like infrastructure.
Every number stated on this page comes from a real screen of the app or from its source. The moving parts are demos, with demo data.
35,600+tests behind 1,053 releasesTechnical DNA · CHANGELOG.md
Whatever watches your agents sees all of your work: the code, the prompts, the answers and what they cost. None of that has to leave your machine for the watching to work.
How it works · pick a step
The hooks go into Claude Code's settings.json beside your own entries. The first write keeps an .agentfarm-bak backup, and Remove Live Hooks takes out only Agent Farm's entries.
Settings · 3
Setting
Default
What it does
agentFarm.liveMonitor
on
Runs the live event listener.
agentFarm.livePort
8377
The listener's port on 127.0.0.1. Installed hooks follow a change without a reload.
agentFarm.planUsageRefreshMinutes
5 min
How often an idle farm reads your plan's usage. 0 = only when you ask.
X·1Inside VS Code1 / 5
X·2
No telemetry
No analytics, no usage statistics: we don't know which features you use, how often or on what. Free never contacts us; a paid licence only checks its key.
Most tools count how they are used and send the count home. You are left to trust what is in it.
How it works · pick a step
Export Traces (OpenTelemetry JSON) writes your sessions to a file on your disk. Where that file goes is your call; Agent Farm sends it nowhere.
Settings · 4
Setting
Default
What it does
agentFarm.digestTtsModel
empty
The fal.ai voice model the briefing is read with.
agentFarm.embeddingsProvider
off
Semantic search through gemini or openai, Ollama included.
agentFarm.notifyWebhookUrl
empty
A webhook to forward notifications to.
agentFarm.kapici.dispatch
off
Asks Gatekeeper which agent takes a request. Each use of Gatekeeper has its own switch, and all start off.
X·2Free1 / 5
X·3
ShellGuard
Before a shell command runs, the shapes that garble or hide it are refused: non-ASCII, backticks, long heredocs, long quoted text. The agent is told the safe way: write the text to a file.
What it solves
Agents write long shell commands. Some arrive garbled, a character re-encoded on the way, and some run more than they show, a second command inside a backtick. Neither is visible until it has run.
How it works · pick a step
The backtick rule applies where the shell is bash: Claude Code, Git Bash. Codex's shell on Windows is PowerShell, where a backtick is the escape character, so there it is not refused.
X·3Before it runs1 / 4
X·4
BridgeGuard
An agent that goes searching the disk for a session by its id is stopped and sent to the bridge, which reads it properly.
An agent asked about an old session goes looking for it on disk: a recursive search through every CLI's folders for its id. It burns context and quota, and those folders and their formats change with each CLI release.
How it works · pick a step
Each stop is logged and leaves a mark on that session's Console card.
X·4The hunt1 / 5
X·5
Permissions as shown
The permission a card shows is the one in force; if a policy forbids it, the session is refused, never quietly downgraded. A lint finds blanket allows and dangerous patterns.
What it solves
When a card says one permission and the process runs under another, a session you set to bypass can sit for an hour on a question nobody sees, and one you thought would ask goes ahead.
How it works · pick a step
An agent can answer a session's question over the bridge, but it can never grant a tool permission. That stays your call, in the panel.
Settings · 2
Setting
Default
What it does
agentFarm.permissionLint
on
Audits the permission rules in each agent's .claude/settings(.local).json; findings go to Insights.
agentFarm.launchDefaults
{}
The model, effort and permission mode every interactive session carries. An empty field adds no flag.
X·5One label1 / 5
X·6
Keys kept out
Route keys are scrubbed from whatever a run prints, and Gatekeeper's log keeps only a hash and a length of what it read.
A key handed to a run ends up in what the run prints: a request line, an error, a config dump. Archived, that output is your key in plain text in a file.
How it works · pick a step
A secret can be marked farm only: Agent Farm still uses it, but no agent's task receives it. The Gatekeeper key starts that way. And a key found in plain text in a config file is reported, so you can move it to the vault.
Settings · 1
Setting
Default
What it does
agentFarm.mediaAuthSecret
FAL_KEY
The name of the secret that holds the media key. The setting holds the key's name, not the key.
X·6The keychain1 / 5
X·7
Browser on a leash
A browser step opens only the sites on its allow list, or fails without running, and asks you before each run unless you chose otherwise.
What it solves
An agent with a browser can follow a link anywhere. A step that runs at night has nobody watching where it goes.
How it works · pick a step
The agent's prompt carries hard rules as well: open no link off the list, never make up a sign-in, and report NO BROWSER MCP when there is no browser. A login wall on a recipe step is reported as LOGIN NEEDED, never typed into.
Settings · 2
Setting
Default
What it does
agentFarm.browserAllowHosts
[]
Hosts every browser step may open, beside the step's own list.
agentFarm.browserApproval
ask
ask: approve each run, with the URL and objective shown. auto: no question.
X·7The step1 / 5
X·8
Superpowers
Engineering disciplines on disk, such as tests first and systematic debugging, with quality gates.
What it solves
An agent left to itself patches the symptom and says it is done. Loading a whole engineering method into every session to stop that costs tokens on every turn, even when the work is a lint sweep.
How it works · pick a step
Agent default is not off: it hands the choice back to the agent's own setting, which is what every task written before the switch existed still says.
Settings · 2
Setting
Default
What it does
agentFarm.superpowersPath
empty
The disciplines' folder. Empty: _SKILLS/superpowers beside your agent folders.
agentFarm.superpowersDefaults
developer: on
On or off per agent, by exact name. An agent not named starts without it.
X·8Asleep on disk1 / 5
Questions
Does anything about my work come to you?
No. The free version never contacts us. A paid licence checks its key once a day, and that message holds the key, a hash of the machine and the version, nothing else. Your code, prompts, transcripts, names, costs and the features you use never come to us.
What does a paid licence do without a connection?
It keeps working for 30 days without reaching us. The check only asks whether the key is valid; nothing runs on our side.
Who sees my prompts, then?
The provider of the engine you run, as without Agent Farm: Anthropic for Claude Code, OpenAI for Codex, Google for Gemini and Antigravity, and the providers you give OpenCode. A route sends a run to the provider you picked for it.
Does Agent Farm sandbox my agents?
No. It maps its permission choices onto each engine's own: Claude Code's permission modes, Codex's sandbox, Gemini's approval mode, Antigravity's mode, OpenCode's plan agent. Parallel runs work in separate git worktrees. The guards on this page stop the commands that should not run.
Do the guards work on every engine?
They act in the hook an engine runs before a tool: Claude Code's, and Codex's. BridgeGuard reports for each engine whether a hook covers it, so an engine without one is never assumed to be guarded.
Can I let one command through?
Yes, on purpose and one command at a time: af:allow-shell for ShellGuard, af:allow-id for BridgeGuard. It is written into the command, so the session's transcript shows every time it was used.