Coding agents

If you use Claude Code or Codex, Kyra becomes the thing that tells you a session is stuck while you are making coffee — and lets you unstick it without going back to the desk.

The agents run on your subscription, on your machine, and they do their work there. What reaches Kyra is what you ask for: an agent’s summary of what it did, its final answer (capped at about 2,000 characters), a question it stops to ask you, a short digest of recent activity, and any file you ask Kyra to read. Those summaries and questions can quote your code, because that is what the agent was working on. Kyra orchestrates and summarises — it does not copy your repository anywhere.

Set up: connect your computer.

What to say

You say	What happens
“What’s running on my machine?”	Lists your sessions — blocked, finished, and errored first
“What needs me?”	Only the ones waiting on you
“Tell the auth one to keep v1 for now”	Answers a blocked session
“Have the migration one use bcrypt instead, no need to stop”	Steers a session mid-task
“Interrupt it — do the operations folder instead”	Redirects a session
“Hand the token-refresh session to me”	Returns control to your keyboard
“Start a session to write the migration tests, branch off main”	Spawns one (asks first)
“Open a session in my editor”	Opens a Claude Code tab; you type the first message
“How much has that session cost?”	Reads back the session’s own figures

Refer to sessions however is natural — “the auth one”, “the token-refresh thing”. Kyra resolves by topic and asks when two match.

Three ways to get work done

Asked to do something, Kyra does not reflexively start a new session — a second agent on the same code is how two of them end up editing one file. It prefers, in order:

Tell the session already working on it. Cheapest, and it keeps that session’s context.
Open one in your editor, if you are at your machine. You watch the diffs land and can take the keyboard whenever you like.
Run it in the background, if you are in the car, on your phone, or your editor is not open.

Kyra always says which of the three it did. A session started somewhere you are not looking is the one case where silence would read as nothing having happened.

Driving the session you are already looking at

kyra install vscode

Then reopen any Claude Code session you had open — mediation applies at launch. Full steps →

With it on, Kyra can drive a session in your editor and not just ones it started: same conversation, same tab, whether you are typing or talking. It can send it work, redirect it mid-task, answer what it is waiting on, and read back what it has cost.

Three things worth knowing before you turn it on:

It is a local door, reachable by anything already running as you on that machine. That is not a hole Kyra opened — anything running as you can already read your files and type into a running session — but it does mean it is only as trustworthy as the code you run, which is why it is off until you turn it on.
Kyra can lower and restore a session’s permission mode, which is the point: dropping one to planning and putting it back. It cannot raise one to bypass permissions — that stays on the path that asks you on your device.
A door it cannot open loses the capability, not the session. A stale socket from a crashed run means the session runs unmediated, not broken.

Linux and WSL only. Why →

Permission modes

A session Kyra spawns runs in the mode you chose when you started kyra connect:

Mode	The session may
`read`	Read files. The default.
`write`	Also edit them.
`execute`	Also run commands.

kyra connect --permission-mode write

Kyra can change a session’s mode by voice, but only downwards freely. Raising it asks you to confirm on a later turn — never in the same breath as proposing it.

What Kyra can see

A coarse status per session: working, blocked on a question, or done. Plus what the session reports through the agent skill, a short digest of recent activity, and anything you ask Kyra to read.

Those reports are written by the agent about the work it is doing, so they can quote your code — a blocked question often has to, and so does a summary of a change. The recent-activity digest includes what you typed into the session. Both are stored so Kyra can answer “what is it doing?”, and both are sent to the model that answers you.

Not: your repository. Kyra never copies or uploads it, never reads files you have not asked for, and no agent runs anywhere but on your own machine.

The agent skill (kyra init) adds exactly two things a session can volunteer — “I’m blocked” and “I’m finished, here’s the gist”. Everything else is driven by you.

Safety

Sessions are strictly per-user. A tool only ever sees your own machine’s sessions.
A coding agent’s output is treated as untrusted data. Kyra summarises it and never executes it as instructions.
Code only ever runs inside your own agent environment. The channel decides who may drive, not what may run.
Spawning a session always confirms first.
How often Kyra asks is yours to set. Kyra has no tool for those settings and cannot widen them.
Revoke a machine from Settings → Coding agents — immediate, and does not need access to the machine.

Troubleshooting

Setup problems are in connect your computer.

Kyra names the wrong session. Say more of the topic. It resolves by topic and asks when two match.

A session finished and Kyra did not say so. It reports through the agent skill — re-run kyra init after a CLI upgrade to pick up a newer one.

Kyra sees the editor but no sessions. Mediation is separate: kyra install vscode, then reopen them.

Claude and Claude Code are trademarks of Anthropic; Codex is a trademark of OpenAI. Kyra is not affiliated with, endorsed by, or sponsored by either — the names identify the tools Kyra works with. The agents run on your own accounts and subscriptions.