- Home
- Skills
- Agents & Orchestration
- Walkie Talkie
Works with the AI Agents you already use
Walkie Talkie
Running two or three Claude Code terminals on the same repo is fast until they edit the same file, both run a migration, or one rebases while the other is mid-change.
$2.99
Walkie Talkie
Example session with this skill installed
I have another Claude Code terminal open on this repo. Check what it is doing, then claim src/auth and start on the JWT login flow.
- Read your context and instructions
- Compiled the walkie talkie
Checked in with Walkie Talkie. One other session is active in this repo:
you are bravo | project shop-api
alpha 4m "payments refactor" holds src/payments, src/db/migrations
No overlap with src/auth, so I claimed it:
$ walkie claim src/auth
claimed 1
Coordination status
| Item | Value |
|---|---|
| Your callsign | bravo |
| Other session | alpha, payments refactor |
| alpha holds | src/payments, src/db/migrations |
| You hold | src/auth |
| Conflicts | none |
The JWT flow needs a refresh_tokens table, but alpha holds the migrations folder. Instead of editing it, I asked alpha:
$ walkie send alpha "Need a refresh_tokens table (user_id, token_hash, expires_at). Can you add it to your migration batch? I'm in src/auth." --kind ask
sent
Next steps
- Build the token service and auth middleware in
src/authnow. - At my next checkpoint,
syncfor alpha's reply before writing the token storage code. - When finished: mark the task done, release
src/auth, and send a notice toallthat the auth API changed.
Connects securely to your tools. The creator never sees your data.
About this skill
Running two or three Claude Code terminals on the same repo is fast until they edit the same file, both run a migration, or one rebases while the other is mid-change. Agent sessions cannot see each other. Walkie Talkie gives them a shared local channel.
What it does
- See who else is working. Each session gets a callsign (alpha, bravo, ...) and sees the other sessions in the same repo, what they are working on and which paths they hold.
- Claim files before editing. Claiming a file or folder another session holds is refused until the sessions coordinate or the user overrides.
- Split work. A lead session posts tasks to a per-repo board with dependencies. Other sessions take the next unblocked task atomically, so no task is taken twice.
- Message and hand off. Asks, replies, handoffs and notices, capped at 400 characters.
- Warn before shared changes. Agents send a notice before rebases, dependency upgrades, migrations, dev-server restarts and file moves.
Built to be cheap on usage
| Action | Output |
|---|---|
| Check-in with nothing new | nothing (0 tokens) |
| Check-in with 2 messages | about 35 tokens |
| Join and see one other session | about 20 tokens |
| Waiting for a reply | 0 tokens while waiting |
Agents check in only at checkpoints (before editing, after a unit of work, before the final reply), never after every tool call. A session that finds no other sessions stops using the skill.
Optional automatic mode (Claude Code)
Run print-hooks and add the output to your Claude Code settings. Sessions then register on start, receive only new messages with each prompt, are asked before editing a file another session claimed, and sign off on exit. Nothing is installed automatically.
Safety
- Messages from other sessions are treated as information, not instructions. Agents will not push, delete files or share credentials because another session asked.
- Everything stays in
~/.walkie-talkie/on your machine. No network access, no environment variables, no subprocesses. Claimed files are recorded by path and never opened.
Known limitations
- Coordinates sessions on one machine only.
- Without the optional hooks, a session that never checks in is invisible to the others, and claims are advisory.
- Tested on Linux. macOS should behave the same; Windows is untested.
Coordinates sessions on one machine only. Without the optional Claude Code hooks, a session that never checks in is invisible to the others and claims are advisory. Messages are capped at 400 characters. Sessions idle for 45 minutes drop off the list. Windows is untested.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
Trust & safety
Security scan
Security scan passed 7 days ago
Requires
Permissions
File Scopes
Reads and writes only ~/.walkie-talkie/ (session list, message log, task boards, lock files). No network access, no environment variables, no subprocesses. Claimed files are stored by path only and never opened. ~/.claude/settings.json is changed only if the user asks for automatic mode and approves the hook snippet shown by print-hooks.