Spettro is a terminal-first multi-agent coding assistant written in Go.

Quick install

curl -sSfL https://raw.githubusercontent.com/aploide/spettro/main/install.sh | sh

Installs to ~/.local/bin by default (no sudo needed); set INSTALL_DIR to override. Self-updates (/update in the TUI, or the built-in update check) write in place to that same directory, so they never need sudo either.

It uses a configurable agent manifest (spettro.agents.toml + agents/*.md prompts), parallel sub-agent spawning via TOOL_CALL and an agent tool, plus specialized orchestrator/worker roles (plan, coding, ask, explore, code, git, test, review, docs).

Highlights

Configurable multi-agent system via spettro.agents.toml and agents/*.md
Parallel TOOL_CALL spawning of sub-agents
Permission policies: ask-first, restricted, yolo
Live tool traces in planning/coding runs
Fantasy-backed provider routing for OpenAI, Anthropic, and OpenAI-compatible text calls
Multi-provider model support via models.dev catalog + OpenAI-compatible endpoints
Anthropic extended-thinking levels (/thinking off|low|medium|high|x-high|max)
Conversation persistence and resume per project
Project trust prompt before first use in a folder

Build and run

Requirements:

Go 1.26.1+

git clone https://github.com/aploide/spettro
cd spettro
make build
./bin/spettro

Alternative:

go run ./cmd/spettro

First-time setup

At first launch:

Confirm folder trust.
Run /connect to add an API key (or local endpoint).
Run /models to select provider/model.
Start with plan (default agent) and switch with Shift+Tab.

Common commands

Spettro commands are entered with a leading /.

/help show help text
/exit, /quit quit Spettro
/mode, /next cycle active agent/mode
/connect connect provider or local endpoint
/models [provider:model] [api_key] open selector or set directly
/permission <ask-first|restricted|yolo> set execution policy
/permissions [ask-first|restricted|yolo] show/set permission policy
/permissions debug <on|off> toggle permission diagnostics
/budget <n|0> set request token budget (0 = unlimited)
/thinking <off|low|medium|high|x-high|max> set extended-thinking budget (Anthropic Claude Opus/Sonnet; ignored by providers without thinking)
/plan [prompt] switch to plan mode or run plan prompt
/approve execute pending approved plan through coding agent
/tasks [list|add|done|set|show] manage session tasks
/mcp <list|read|auth> manage MCP resources and auth tokens
/skill <list|install|info|uninstall|enable|disable|where> manage Agent Skills (Claude Code / OpenAI / Anthropic format)
/hooks show effective runtime hooks
/compact [focus] summarize conversation history
/compact auto <status|on|off> configure auto-compact
/compact policy show compact thresholds/counters
/clear auto-save and clear current conversation
/resume load a previous saved conversation
/init analyze the repo and create/update SPETTRO.md
/remote [:port] expose a loopback HTTP/SSE control plane (see docs/remote.md)
/remote local [:port] expose the control plane to LAN devices on 0.0.0.0 (see docs/remote.md)
/telegram setup <token> configure a Telegram bot relay so you can drive Spettro from a chat (see docs/telegram.md). Alias: /tg.
/<custom> [args] run your own saved prompts from ~/.spettro/commands/ or .spettro/commands/ (see docs/custom-commands.md)

For full commands and keybindings, see docs/commands.md.

Editor integration (ACP)

spettro --acp runs Spettro as an Agent Client Protocol
agent over stdio, so ACP-capable editors like Zed can drive it from their
native agent panel — with streamed reasoning, tool call reporting, mode
switching, and permission prompts. See docs/acp.md.

Project docs

Session Lifecycle — auto-save, resume, compact, clear
Goal Mode — autonomous /goal runs
Runtime Hooks — custom bash middleware
Clipboard and Attachments — image paste, file attach
First-Launch Onboarding
Spettro Subscription — login, plans, credits
Agent Manifest
Agent Prompts
Agent Skills
Getting started and workflow
Commands and keybindings
Custom slash commands — save reusable prompts as your own /commands
Configuration and storage
OS sandboxing
Architecture overview
Remote control plane
Agent Client Protocol (editor integration)
Telegram relay
Extended thinking levels
Troubleshooting
Documentation Index

Development

make test
make build
make build-all