Slice

A token-efficient, terminal-based AI coding assistant powered by the Vercel AI SDK. Run it against OpenRouter, OpenAI, Anthropic, Ollama, OMLX, or any OpenAI-compatible endpoint.

Slice keeps context small, reads only what it needs, and responds tersely — designed for developers who want a fast assistant that stays out of the way.

Features

• Block/bordered/plain input styles — Adaptive terminal background, Tab completion for commands
• Multi-turn conversation — Full history passed to the model for context continuity
• File tools — file_read, file_write, file_edit, glob, grep, list_dir, shell
• Slash commands — /model, /new, /help, /compact, /session, /export
• Context compaction — LLM-powered summarization of long conversations via /compact
• Provider-agnostic — OpenRouter (default), OpenAI, Anthropic, Ollama, OMLX, or any OpenAI-compatible endpoint

Requirements

• Node.js 18+
• An API key for your chosen provider (or a local Ollama instance — no key needed)

Setup

git clone https://github.com/nikitaclicks/slice.git
cd slice
npm install

Config is split into a public file (committed) and a local file (gitignored). Create both in the project root:

agent.config.json — committed, no secrets:

{
  "provider": "openrouter",
  "baseURL": "https://openrouter.ai/api/v1",
  "model": "minimax/minimax-m2.5:free"
}

agent.config.local.json — gitignored, API key only:

{
  "apiKey": "sk-or-..."
}

The local file is merged on top of the public one at startup. Any field can go in either file — local.json always wins.

Then run:

npm start

Profiles

Profiles let you maintain multiple named configurations — one per provider, model, or use case — and switch between them at startup.

Each profile follows the same public/local split:

agent.config.json              # default public config — committed
agent.config.local.json        # default secrets — gitignored
agent.work.config.json         # work profile public config — committed
agent.work.config.local.json   # work profile secrets — gitignored

npm start              # loads agent.config.json
npm start work         # loads agent.work.config.json
npm start local        # loads agent.local.config.json

Each profile config is a full agent.config.json — set whichever fields differ and the rest fall back to defaults. Profiles also activate matching entries in prompts.config.yaml (see Custom Prompts).

Inspect the active system prompt

npm start -- --print-system-prompt
npm start -- work --print-system-prompt

Provider Configuration

Set provider, baseURL, and model in agent.config.json (committed). Put apiKey in agent.config.local.json (gitignored).

OpenRouter

agent.config.json:

{
  "provider": "openrouter",
  "baseURL": "https://openrouter.ai/api/v1",
  "model": "minimax/minimax-m2.5:free"
}

agent.config.local.json:

{ "apiKey": "sk-or-..." }

OpenAI

agent.config.json:

{
  "provider": "openai",
  "model": "gpt-4o"
}

agent.config.local.json:

{ "apiKey": "sk-..." }

Anthropic

agent.config.json:

{
  "provider": "anthropic",
  "model": "claude-sonnet-4-6"
}

agent.config.local.json:

{ "apiKey": "sk-ant-..." }

Any OpenAI-compatible endpoint (local or proxy)

{
  "provider": "openai",
  "baseURL": "http://localhost:1234/v1",
  "apiKey": "your-key",
  "model": "your-model-name"
}

Ollama (no API key needed)

{
  "provider": "ollama",
  "model": "llama3.2"
}

Custom Prompts

System prompts are configured in prompts.config.yaml (committed to source control — no secrets here). Prompts are resolved in order from most to least specific:

1. Profile — matched by the profile name used to start Slice (e.g. npm start work)
2. Provider + model — matched by both provider and exact model ID
3. Provider default — matched by provider only
4. Global default — fallback for everything

default: |
  You are Slice, a coding assistant...

  Current working directory: {cwd}

profiles:
  work: |
    You are Slice for the work profile.
    Prefer concise answers and bash-first workflows...

providers:
  anthropic:
    default: |
      Prompt for all Anthropic models...
    models:
      claude-opus-4-7: |
        Prompt for Opus specifically...
  openrouter:
    default: ""
    models: {}

Use {cwd} anywhere in a prompt — it's replaced with the current working directory at runtime.

A systemPrompt field in agent.config.json takes precedence over everything in prompts.config.yaml, useful for per-profile secrets or absolute overrides.

Slash Commands

Command	Description
/new	Start a fresh conversation (clears history)
/compact	Summarize older messages to reduce context size
/session	Show current model, message count, and token usage
/export [file]	Save conversation as a Markdown file
/help	List all commands
exit	Quit

Type / and press Tab to see available commands.

Configuration Reference

Field	Default	Description
provider	openrouter	AI provider: openrouter, openai, anthropic, ollama, omlx
apiKey	—	API key for the selected provider
baseURL	—	Override the provider's default API endpoint
model	nvidia/nemotron-3-super-120b-a12b:free	Model ID (format depends on provider)
maxSteps	20	Max tool-use steps per turn
timeout	none	Agent run timeout in milliseconds. Omit (default) for no limit. Example: 3600000 for 1 hour
sessionDir	.sessions	Directory for session JSONL logs
showBanner	true	Show ASCII banner at startup
display.inputStyle	block	block / bordered / plain
display.toolDisplay	grouped	grouped / emoji / minimal / hidden

Tools Available to the Agent

Tool	Description
file_read	Read file contents with optional offset/limit
file_write	Write or create files
file_edit	Search-and-replace edits
glob	Find files by glob pattern
grep	Search file contents by regex
list_dir	List directory contents
shell	Run shell commands

License

MIT