Agent River — Development skill for Claude Code


Agent River

**One local post office for you and your AI agents.**

[English](README.md) · [繁體中文](README.zh-Hant.md)

Send a letter to Codex, Claude, or an agent you already run. Bring several agents into the same conversation, let them read each other's replies, and keep the discussion to a fixed number of rounds. Agent River handles delivery and keeps the correspondence together in a local web interface.

Your agents keep their own tools, models, and login sessions. Agent River starts their local processes and records requests, replies, and delivery status on disk. Model calls still use the providers configured by those agents.

What you can do

• Write to one agent. Choose a recipient or let the post office route a request by capability. A single-recipient agent can ask a colleague for help and receive the result back.
• Start a group discussion. Select multiple recipients once. Each round waits for everyone's replies before sharing them with the group. Choose 1–3 rounds; the default is 3, with a final prompt to summarize agreement, differences, and next steps.
• Steer the conversation. Follow up with the whole group or selected agents. A new instruction starts a fresh round budget and prevents the previous discussion from adding more rounds.
• Read the conversation without the delivery noise. Your group letter appears once. Earlier messages, forwarding records, and retries are folded away; model choices and round progress use compact labels.
• Connect existing agents. Codex and Claude have built-in runners. Other agents can use a stdin/stdout adapter or poll the mailbox. Otter is one example of a separately connected agent, not a bundled dependency.
• Use Telegram when you want it. The optional dashboard supports owner commands, bounded sessions, agent launches, and approval workflows alongside the web post office.