sage-ai-cli 1.20.17

Sage — a local-first AI coding CLI (like Claude Code, using free/open models)

Sage AI CLI

A free, local-first AI coding agent for your terminal. Like Claude Code, but using free and open models — no API key required.

pip install sage-ai-cli
sage install   # auto-pulls models, sets best coder, builds RAG, prewarms
sage run       # start coding with the full Tier 1-3 safety harness

What Sage Does

Sage is an autonomous coding agent that runs in your terminal. It:

• Plans your task and breaks it into concrete steps
• Reads your codebase (with project-aware RAG) to ground itself in real symbols
• Writes code using TDD — tests first, then implementation
• Runs tests automatically and fixes failures
• Iterates until the task is complete
• Routes to the best-fit model per request (small for trivial, strong for hard)
• Searches the web for docs when its training data is too old
• Fine-tunes local LoRA adapters on your codebase
• Refuses to run with a model too small to follow the protocol (T1)
• Stages all writes through a tmpdir so broken output never reaches your repo (T5)
• Blocks npm install after rejecting a poison package.json (T6)

All using free AI models — no API keys, no subscriptions, no cloud dependency.

Quick Start

# Install everything (Ollama, models, RAG, optional deps) — disk-aware
pip install sage-ai-cli
sage install

# Coding agent with full safety harness
sage run

# Try the new features:
sage rag index                          # build per-project semantic index
sage ext search "react hooks"           # web search (no API key)
sage ext route "design a new system"    # see which model the router picks
sage ext detect                         # see auto-detected project context
sage ext bootstrap --finetune           # kick off project-aware fine-tune

Documentation:

• docs/SAGE_FEATURES.md — full feature catalog
• docs/SAGE_COMMANDS.md — every CLI + slash command
• docs/SAGE_ARCHITECTURE.md — module map + data flow

What's New

Waves 1–5 + Tiers A–C (Capability)

Capability	Command
Auto-pick strongest installed coder	sage ext auto-pick
Project context detection	sage ext detect
Local RAG over your codebase	sage rag index / sage rag query "..."
Web search (DuckDuckGo, no key)	sage ext search "..."
Hybrid difficulty-based routing	sage ext route "..."
LoRA fine-tune with adapter cache	sage ext finetune <model>
GCS-mirrored public datasets	sage ext datasets mirror --name all
One-shot disk-aware bootstrap	sage ext bootstrap [--finetune] [--full-datasets]

Tiers 1–3 Hardening (Safety + Quality)

These prevent the "silent quality collapse" failure mode where a small fallback model produces broken output. Wired into every sage run session.

Tier	Feature	Impact
T1	Hard model-capability floor	Refuses agentic tasks below 7B params
T2	Diff-preview before batch FILE: writes	User confirms before disk hit
T3	Bounded regenerate context	Stops feedback bombs in retry loops
T4+T12	GBNF on tool turns + project-aware imports	Hallucinated imports impossible
T5	Tmpdir staging	All writes pass through isolated workspace
T6	Run guard	npm install blocked after rejected package.json
T7+T9	Project-skeleton bootstrapper	React+Vite, Node+Express, FastAPI, fullstack
T8	RAG before first turn	Project context ready on turn 1
T10	Telemetry with secret redaction	Per-session JSONL log
T11	Two-model planner/coder default	Cheap planner + strong coder, automatic
T13	Readiness self-test	30s "write hello.js" probe at session start
T14	Disk-aware install + safe pip	Picks model tier that fits free disk

Architecture

core/run_hooks.py is the single orchestrator. sage run calls:

1. on_session_start — capability check (T1), readiness probe (T13), skeleton match (T7+T9), RAG indexing (T8)
2. on_pre_turn — grammar selection (T4+T12), planner/coder pair (T11), top-K RAG retrieval (T8)
3. on_post_turn — telemetry log (T10) with secret redaction

Every FileWriteTool.write() is wrapped by content_validator and arms the run_guard on rejection. Every RUN: shell call is gated by run_guard.allow().

See docs/SAGE_ARCHITECTURE.md for the full data-flow diagram.