Karma Yoga Get the CLI

Technical details

How Karma Yoga works

You give a manager a goal. It divides the work among AI agents, checks their progress and brings their answers back to you. Karma Yoga keeps that team together on your own computer.

One manager, several conversations

An agent is a running copy of an AI tool, working on a conversation of its own. The manager is the same kind of agent, given the job of coordinating the others. The other agents take jobs such as making a change, checking it or writing an explanation. The manager can add agents when the work needs them, or put one aside and bring it back later.

One manager, a crew of agents You give the manager a goal. The manager divides the work among agents. An arrow returns from the agents to the manager to show their reports coming back. You set the goalThe manager divides the workA builder makes changesA reviewer checks themA writer explains themReports back
You talk to the manager. The agents have separate jobs and conversations.

A team and its conversations form a session. The manager stays visible beside the other agents, so you can follow their work or talk to one directly. Messages between agents are explicit and recorded. What an agent writes only in its own window is not automatically sent to the others.

Local processes, remote thinking

Each agent runs as an ordinary process on your computer, using Claude Code, Codex or Pi. Karma Yoga’s background program is its parent and keeps its working window open. The tool reads files, runs commands and saves the conversation locally. When it needs the model to reason or write, it sends a network request to the provider’s servers and receives an answer.

The agents run through your own Claude Code, OpenAI Codex or Pi installation, signed in to your own accounts. Codex is OpenAI’s coding agent and can use your ChatGPT account. Karma Yoga brings no AI model or account of its own. The work uses your existing subscriptions or paid usage.

The tool is local, the model is remote On your computer, Karma Yoga keeps the agent tool running. The tool reads files, runs commands and saves its conversation. A request goes to the model at your provider, and an answer returns to the tool. Your computer Karma Yoga keeps the tool running The agent toolFiles, commands and conversationLocal CPU, memory and disk Request Answer The model at your provider
The model does not run in each agent’s local process. The tool connects its answers to work on your computer.

Sharing one computer

You do not need to load a separate AI model into memory for every agent. The local tools still use CPU and memory, and commands they launch can use much more. Several agents compiling projects or running tests at once can put real pressure on the machine.

Karma Yoga limits how many agents can be active at once and reports machine load, memory, disk space and resources used by each agent. The manager can reduce parallel work when the computer is busy. The manager is instructed to give agents editing the same project separate working copies, so their unfinished changes do not collide in one folder. Those copies still need to be reviewed and combined.

An idle agent can be parked. That stops its process while keeping the saved conversation, and a new message wakes it again. Automatic parking is optional and checks that the agent is safely idle first. The manager stays running.

Conversations and memory

An agent’s working memory is its conversation, including instructions, answers and results from its tools. The model can see only a finite amount at once, called its context window. Its size is measured in tokens, small pieces of text. As the conversation grows, the tool condenses older parts to make room. That keeps work moving, but a summary can lose detail.

The conversation stays with the tool An agent uses your signed in AI tool. That tool saves its conversation. Karma Yoga remembers which conversation to return to. An agent uses your AI toolThe tool saves the conversationKarma Yoga rememberswhich conversation to return to
The tools keep the conversations. Karma Yoga keeps track of which one belongs to each agent.

The saved conversation on disk and the context sent with a model request are different things. Each tool keeps its own saved conversations. Karma Yoga keeps references to them and the team’s settings, and copies the conversations only when you make a backup. Those references let it resume an agent through its tool, provided the saved conversation still exists. During setup, Karma Yoga extends how long Claude Code keeps old conversations.

Context is not the same as allowance

A full context window means the tool needs to condense what the model sees. An exhausted account allowance means the provider will not supply more model answers. Some subscriptions have limits over periods such as five hours and a week. Metered accounts charge for usage instead. Agents sharing an account draw from the same allowance.

If a provider stops supplying answers at a limit, the agent cannot continue its model work until the allowance resets or capacity becomes available. Its saved conversation remains, so it can continue from there.

Karma Yoga’s usage view forecasts whether an account’s remaining allowance will last until its reset. Reserve means it is on pace to last. Deficit means it is on pace to run out first. The manager can run fewer agents or move work to another available account. For metered accounts, Karma Yoga reports spending and can compare it with a limit you set. These are pacing signals, not a spending cap; missing usage data is shown as unknown.

A message is handed over, then checked

When the manager or another agent sends a message, Karma Yoga pastes it as one block into the receiving program’s working window and submits it. Other messages and typing wait their turn, so they cannot be mixed into the middle of it.

From a message to a confirmation The manager sends a message. Karma Yoga hands it over whole and checks whether the agent tool accepted it. The sender receives confirmation or a reason it could not be confirmed. The manager sends a messageKarma Yoga hands it over wholeDid the agent’s tool accept it?Confirmation, or a reason it failed
The sender hears whether the message reached the agent’s tool.

It then checks for evidence that the tool accepted the message, such as the message appearing in the saved conversation. The sender learns whether it was accepted. If Karma Yoga cannot confirm that, it reports a problem rather than silently claiming success.

Karma Yoga uses the tool’s status, screen and saved conversation to distinguish an agent that is working, waiting on a question or idle. A busy agent can receive a message before it is ready to act on it. Accepting a message does not mean it has finished the work.

You can close the window

A background program keeps every agent’s working window running whether or not you are looking, like a switchboard holding several calls. Closing the window you use to watch them does not close the agents. You can return to the same team later.

Karma Yoga includes its own version of this program so it can check that messages really arrived, guard against an agent being closed by accident, and bring every working window back after a restart.

Three ways to leave your work Closing the window leaves agents running. A sleeping computer pauses their work. Stopping a team keeps its conversations for when you start it again. Close the windowAgents keep workingPut the computer to sleepWork pauses until it wakesStop the teamStart it again on saved conversations
Karma Yoga runs on your computer. It can keep working only while that computer is awake.

If closing the laptop lid puts the computer to sleep, work pauses until it wakes.

A restart does not mean starting over

Karma Yoga saves a record of which teams should be running and which conversation belongs to each agent. After a computer restart, it reads that record and brings those teams back on their saved conversations. A team you deliberately stopped stays stopped until you start it.

Returning after a restart or an update After the computer restarts, Karma Yoga reads its saved record and resumes the teams that should be running. An update leaves agents running when it can, or restarts them on their saved conversations. After a computer restartRead the saved recordResume the teams that were runningAfter a Karma Yoga updateKeep agents running when possibleIf a restart is needed,resume saved conversationsInterrupted agents are told to continue
Saved conversations let an agent return to its work. An interrupted step may need to be done again.

An update keeps agents running if the new version can work alongside them. If it needs to restart them, they return to their saved conversations. When you update from a terminal, Karma Yoga asks before interrupting busy agents.

The step an agent was doing may be interrupted, along with any background work it started. Karma Yoga tells the agent to continue and check what needs doing again. If a tool process exits unexpectedly, Karma Yoga tries to bring it back on its saved conversation. If an agent cannot be restored, the manager hears which one failed and why.

What leaves your computer

Your AI tools send their work to their providers under your accounts and those providers’ terms, as they do when you use the tools directly. Running Karma Yoga locally does not mean the AI itself runs locally. An agent can also use the network to carry out its work. An optional plugin can contact the service it connects to.

Karma Yoga sends usage statistics and error reports by default. You can turn them off. Update checks still run. Feedback is sent only with your agreement and can include your words and recent error logs. Setup and tool installation also contact the relevant providers.

The agents have your access

Agents run as you, with your permissions. They can read files you can read, use the access keys available to them and act through your accounts. Teams are not isolated from one another. Give a team the trust you would give a person using your computer.

Each AI tool has its own rules about which actions it allows. Karma Yoga can approve requests to let work continue without you watching. Pi has no approval step or isolated workspace, and other tools can also run with their safeguards turned off. Karma Yoga does not provide a separate security barrier.

Instructions tell agents to treat quoted material from files and websites as information, rather than orders. That reduces the risk of following a hidden instruction. It cannot remove the risk. Choose folders, tools and connected services you trust before starting a team.