Niles: coding agent teams in your terminal

Click in: arrows move, Enter opens, Tab talks to the lead; ask it for a review. Staged; nothing runs.

NILES 4 projects 3 agents 626k tok
PROJECTS
├─ ● billing 318k
│ └─ ● w412 212k
├─ ⚠ docs 96k
├─ mobile
└─ infra
CONFIG
├─ global
└─ projects
TELEMETRY
├─ today
├─ 7 days
└─ 30 days
HELP
billing │ docs │
1:lead 2:w412
❯ Webhook retries double-charge on timeout. Issue #412.
$ gh issue view 412
• One worker can fix the retry key; a reviewer from another
family checks the diff. Spawned w412 on
codex:gpt-6.1-sol:medium.
❯
↵ open · → expand · c close · ? help · tab type

SYNOPSIS

# 1. install, with one of

brew tap Rhovian/niles && brew install niles
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/Rhovian/niles/releases/latest/download/niles-installer.sh | sh
cargo install niles

# 2. first run, in your repository

cd repo && niles

DESCRIPTION

Niles puts a team of coding agents behind one conversation. You talk to a lead agent; it hands work to agents on other models and reads their reports.

Today Niles runs agents in tmux and works with Claude Code, Codex and pi; more backends and agents are coming. One logged-in agent CLI is enough: setup detects what is installed, and solo mode lets one agent do the work itself.

Free and open source under MIT; you pay for your agent CLIs.

REMOTE

╭──────────────────────────────────╮
│ --:-- ▮▮▮▯ │
│ │
│ │
├──────────────────────────────────┤
│ │
│ --:-- │
│ │
│ │
│ ┌──────────────────────────────┐ │
│ │ │ │
│ │ docs: Agent waiting for your │ │
│ │ input │ │
│ └──────────────────────────────┘ │
│ │
│ │
│ │
│ │
│ │
│ │
│ │
│ │
│ │
├──────────────────────────────────┤
│ tap a notification │
│ │
╰──────────────────────────────────╯

The same session from your phone, through the opt-in remote app on your tailnet.

Tap the notification and answer the lead. Ask it for a review there, then look at the home view above: the worker is in its tree.

Remote setup

TODO

From ROADMAP.md: directions, with no set dates or promised order.

  • Audio operator
    talk to your lead and hear updates through a voice interface.
  • More agent types
    expand beyond the current families, with candidates such as OpenCode, Cursor CLI, Grok CLI and Devin.
  • Permission policies
    configure per-agent allowed commands and hooks.
  • Configurable approvals
    choose per skill whether its agent bypasses approval prompts. Today workers bypass them and the operator-facing entry agent keeps them.
  • Configurable pipelines
    define how work moves between agents in a project, such as plan, implement and review, and reuse that flow across tasks.
  • Remote agents
    work with agents that run elsewhere, through MCP or ACP integrations. Which protocol, or whether both, is still open.

BUGS

Linux support is not confirmed yet, and there is no Windows build.

Workers launch without approval prompts, and Niles adds no sandbox. Read the threat model before you point it at a repository.

Report bugs at github.com/Rhovian/niles/issues.

(END)