Frequently asked questions

Moji turns your AI agents into a swarm that finishes the job. Each agent gets one step, and Moji checks its work before the next step starts. When a step fails, Moji redoes that step and keeps the rest of the work. Every job stays inside the budget you set, and Moji records what ran, what it cost and what passed.

Moji Systems is the home of Moji, control infrastructure for AI agents, at mojisystems.com. Moji runs a team’s Claude Code and Codex agents with a check on every step and a hard budget on every job. Frequency AI Ltd, a UK company, builds Moji.

An agent swarm is a group of AI agents that work through a job step by step, each step with its own agent and its own check. Moji runs the swarm, prices each step and records who did what.

Yes. Open any job to read in plain words what its agent is doing and why, with a link to its branch, changes and pull request on GitHub.

Moji runs swarms that you define in your repository, and it draws them as a live map. You change a swarm by editing its file or by asking for what you want.

Claude Code and Codex do the work inside each step. Moji adds a price, a check and a recovery path around every step, and one record across all your agents.

Yes. In October 2026, researchers using AI models found PHP2Shell, two critical flaws in PHP that had gone unfixed for years. Your own agents can look for flaws in your code the same way. With Moji, a swarm of Claude Code or Codex agents hunts for flaws, proves each one with a test, writes the fix and runs your tests, and a person approves each fix before it merges.

Any site that serves web requests with PHP below 8.5.12, 8.4.27, 8.3.36 or 8.2.35 is at risk, on PHP-FPM, Apache mod_php, php-cgi, FrankenPHP or LiteSpeed. Check the version your web server runs with phpversion() or your hosting panel, because php -v on the command line can differ. Upgrade to a fixed version. If you run PHP 8.1 or older, apply the PHP team’s source patches or take the site offline until it is patched.

Yes. A security swarm can attack your own systems on a schedule, confirm each finding and send each fix back through the same test before a person approves it.

You name the systems a security swarm may test. Moji gives its agents no GitHub or cloud keys, each agent can change only its own job’s folder, and Moji keeps an append-only log of who accessed what.

Moji prices each step from past jobs and adds the prices into a budget for the whole job. You commit the budget, Moji counts each step’s reported usage against it, and new work waits at the cap.

Give the swarm a budget in Moji. Moji prices each step before the work starts, counts each step’s reported usage against the budget while the swarm runs and starts no new work once the budget is spent. Your agents keep calling their model providers under your own accounts.

Moji prices each step of the swarm from measured jobs. Those step prices add up to the job budget, which Moji then meters live while the job runs.

Your team writes its quality checks into the swarm. They can be rules on a step’s result, a score from a separate reviewing agent, files that must exist, your own CI tests or a named person’s approval. Moji runs them on every step’s output. Each swarm’s report shows how often work passes first time, how often a person had to step in, how long jobs take and how often they stayed on budget.

Moji retries the failed step in place up to the number of attempts you allow, or sends the work to the repair step your swarm defines. Completed work keeps its state, and a step whose outside effect is uncertain waits for a named person to confirm what happened.

Agents loop, misread each other’s work and pass wrong results to the next step, and each error compounds. Berkeley’s MAST study reviewed 1,642 traces and found that multi-agent systems failed 41% to 87% of their tasks across seven open-source frameworks. Moji checks every step as it completes and sends failed work back before the next step starts.

When you ship a new version of a swarm, Moji prices it. New jobs use it once you commit it, and work already moving finishes on the version it started with.

Every swarm starts with a person reviewing its work. The swarm’s report shows how often work passes first time, how often a person steps in, how long jobs take and whether they stay on budget. Once the swarm meets the bar you set, you can switch on Autopilot, and Moji records every approval it makes for you.

Yes. With Autopilot on, a swarm keeps taking new work as it arrives, and Moji records every automatic approval. The swarm stops taking new work when it reaches its spending cap.

Moji runs on your own infrastructure: one machine, or inside your private network behind your own proxy. Your workers always run on your own machines. Moji can also host its central service for your team by invitation, with your workers and code on your own machines.

Moji gives every job its own Git branch and its own folder, and an agent can change only its own job’s folder. Workers run on your machines, from one job to 64 at once, and a test, a second agent or a person checks each step before its work moves on.

We do not have a SOC 2 report yet. Moji runs inside your own security boundary, so your code and data stay in systems your review already covers, and we answer your security questions during the pilot.

Your code, documents and agent logins stay on your machines. Moji stores each job’s state, results and costs, and your agents call their model providers under your own accounts.

A gateway or model router handles one model call at a time. Moji handles the whole job from the first step to the last, and keeps its budget, its checks and its record together while it runs.

An agent platform hosts agents you build on it, and permissions what an agent may use and spend before it acts. Moji runs the agents you already use on your own machines and controls the work while it moves. Moji checks every step, fixes failures in place and keeps the job within its budget.

Moji fits teams that run repeated multi-step agent work in production, where a wrong result matters, a failed step should not restart the whole job, and inference is a material cost.

Leave your work email on the early access page, or email [email protected] with one piece of agent work you want to hand off. We set the success criteria up front, connect one swarm and begin with a person reviewing each job.

Pricing is set per customer and agreed before anything starts. Every figure on this site is illustrative.