Because you would rather build than babysit
- One install on your own computer. A laptop is enough.
- Keep the agents you already like. Nothing to relearn.
- Agents pick up where the last one stopped.
- Risky actions ask you first, on your phone.
Claude Code, Codex, Antigravity, Pi and your local models message each other and work from one shared context: the same memory, the same task ledger, the same map of your code. You stop pasting background into every chat and carrying news from one agent to the next.
Tell one agent once. The next one already knows what the project is, what was decided and what is half done. Install tethercode once, then keep using Claude Code, Codex, Antigravity or Pi exactly as you do now.
The searching, summarising and research your agents hand to tethercode run through one router. Of the models allowed to see a call and able to do it, it tries the most private first, then the cheapest, and records what each call cost. Your expensive model gets the answer, not the reading.
tethercode runs on your machine. The code map, the agents' memory and the task ledger live in your own database, not ours. What may go to a cloud model and what must stay with a local one is your setting, checked on every single call.
What they read, what they changed, what it cost and who allowed it. tethercode makes every agent act as a named person with that person's access, reads every model call before it leaves, asks a human before anything risky, and records every decision. On your servers and your GPU.
Works with Claude Code, Codex, Antigravity, Pi, local models and any provider's key. It is not another AI.
The models are good. Everything around them is missing, so you fill the gap by hand.
You explain the project, the decisions and the half-finished work again. Each tool remembers only itself, if it remembers anything.
Two of them edit the same file. One undoes the other's work. Nobody knew who was doing what, and you find out when the tests fail.
An agent rewrites a function without knowing what calls it. It works here and breaks something three files away.
Every file an agent opens and every search it runs is paid for in tokens. So it reads a few files, searches once and guesses the rest, then forgets half of that when the conversation gets long. Most agent mistakes are things it never read or no longer remembers.
Including the work a cheap or local model would have done just as well. You find out on the bill.
An approval step that looks for "delete" or "restart" is defeated by the first synonym. An agent only has to find the one word you forgot, and it will find it while doing something else.
What they read, what they changed, what it cost, who allowed it. For one person that is a nuisance. In a company it is the reason the answer is still no.
tethercode is software you install once, on your own computer or your company's servers. The AI agents you already use connect to it and keep working as before. It is not another AI: your agents and models still do the thinking.
What it adds for all of them:
Everything it stores is in your own database. Free for 3 repositories, $19.99 a month for all of them.
Session.refresh has 3 callers. Two are in files Codex is working on.
Session.refresh returns. Your billing tests call it.
Each vendor's tool knows its own agent and nothing else. Everything below needs something underneath all of them. All of it is built, and all of it is in the free edition unless its page says otherwise.
Claude Code, Codex, Antigravity and Pi message each other, and every task sits in one ledger, so "who is doing what" has an answer.
How agents work together
A new session, a shortened conversation or a different agent tomorrow begins already knowing the task.
How work carries over
A map of your repository does the looking for the agent and answers with file and line. About 90% fewer Claude tokens in our own test.
How code search works
Memory that names its sources, documents kept whole, and reasoning whose guesses wait for a person.
Inside the knowledge engine
Access by data source, a trained check on every model call, and a person's approval before a risky action.
Access and confidentiality
One router for local and hosted models: the most private that may see a call first, then the cheapest, every call recorded. On machines you run.
Models and hostingtethercode comes with a console in your browser, served from your own machine. These are the screens people open first, drawn from the console's own.
Every agent, what it is working on, what it has decided and ruled out, and what is waiting for you.
What your agents say to each other, whether each message arrived, and what they agreed. You can write to them too.
Pick any function and see what has to change with it, before anyone edits a line.
Ask a question and get the answer with the records behind it, and a plain "not found" for what it does not hold.
Which model each call went to, why it was allowed to go there, and what it cost.
What it worked out and what it only suspects, each with the path that supports it. Nothing becomes a fact until you confirm it.
A research run with its report, every source it cites, and every search and page behind it, including the pages it refused.
Illustrations of the console's screens, drawn with example data.
An agent that has just opened a customer's account should not be sending a summary to a public model thirty seconds later. tethercode reads every model call before it is sent, with a small model called Kev that was trained for that one job. We gave Kev and a general-purpose model the same 1,000 samples.
Kev ours, trained for this job
Most sensitive
Confidential
Internal
Public
Jev a general-purpose hosted model
Most sensitive
Confidential
Internal
Public
Each dot is one test sample. The samples are synthetic and the general model was judged on definitions it had never seen: the method, the tables and the limits.
Deploying for a company? Talk to us. What each edition includes. Questions and answers.