savant-code

README

A terminal-native AI coding assistant that audits every change before it touches your repo.

Built with TypeScript/Bun, governed by the ECHO Protocol, and designed for local-first use with Ollama.

v0.0.9 — Local-first Ollama support, privacy-first BYOK defaults, sandbox engine network gating, /health command, and full launch-ready artifacts.

---

Get Started in 30 Seconds

# Install the CLI
npm install -g @savant-code/cli

# Run it. If Ollama is running, it will auto-detect and use it.
savant-code

No Ollama yet?

# macOS / Linux
curl -fsSL https://ollama.com/install.sh | sh
ollama serve

# Windows: https://ollama.com/download/windows

Then run savant-code again, or type /health inside the chat to verify the connection.

---

Overview

Savant-Code is a TypeScript monorepo that builds and ships the terminal-native AI coding assistant Savant Code and the public @savant-code/sdk. The CLI provides multi-agent orchestration, custom skills, MCP tool discovery, mode switching (EDIT / ANALYZE / SCAFFOLD), and local-first Ollama support. The SDK, agent runtime, multi-agent orchestration engine, tool layer, and LLM provider shims are shared so both surfaces ship from one codebase.

The whole project ships under ECHO Protocol v0.2.0 — the same 15-law agent discipline that governs the Savant ecosystem. Every change goes through the RED → GREEN → AUDIT → SELF-CORRECT → COMPLETE Perfection Loop FSM, with a hard 10-iteration cap and a 10% Levenshtein change-cap per pass.

---

Key Technologies

Layer	Tech	Version
Runtime	Bun	1.3.14 (engines >=1.3.11)
Language	TypeScript	5.5.4 (strict: true, noImplicitReturns: true)
TUI	OpenTUI + React 19	@opentui/core 0.2.2, react ^19.0.0
State	Zustand + Immer	zustand ^5.0.8, immer ^10.1.3
Validation	Zod	^4.2.1
LLM SDK	Vercel AI SDK	ai ^5.0.52 + @ai-sdk/anthropic 2.0.50
MCP	Model Context Protocol	@modelcontextprotocol/sdk ^1.18.2
Code parsing	tree-sitter (WASM)	@vscode/tree-sitter-wasm 0.1.4
HTTP / WS	ws, node-fetch, custom SDK client	ws ^8.18.0
Package manager	Bun workspaces (hoisted)	bunfig.toml [install] linker = "hoisted"

---

Features

CLI (@savant-code/cli)

• Multi-agent orchestration — 9 specialized agents coordinate via ECHO Protocol: Detective finds issues, Forge implements, Verifier audits, Recorder manages FIDs, Thinker reasons, Scout explores, Researcher investigates, Scribe documents.
• /init command — scaffolds .agents/types/{agent-definition,tools,util-types}.ts and a starter knowledge.md.
• Slash commands — /new, /history, /bash, /goal, /loop, /feedback, /theme:toggle, /login, /logout, /exit, plus agent-specific commands.
• @filename and @AgentName mentions — file and agent mentions with inline autocomplete.
• Bash mode — !command or /bash to run shell commands inline (with confirmation).
• Goal loop — /goal sets a goal condition; /loop <cadence> schedules recurring prompt execution (e.g., /loop 5m check build status). The loop scheduler manages cadence, run counts, and convergence detection.
• Mode switching — EDIT / ANALYZE / SCAFFOLD execution-scope modes, togglable at runtime via UI.
• Streaming & cancellation — token-by-token SSE streaming with mid-stream cancellation, retry-with-backoff, and subagent streaming for parallel work.
• Knowledge files — project-level knowledge.md plus per-user home-dir knowledge, auto-loaded into agent context.
• Skills — OpenClaw-format SKILL.md files discovered at startup, schemas sent to the LLM, available as native tools.
• MCP tools — Model Context Protocol servers discovered at startup, schemas published to the LLM API.
• Context compaction — 4-layer progressive auto-compaction: L0 (summarize old turns), L1 (compress tool results), L2 (prune stale context), L3 (aggressive reduction). Preserves critical context while reducing token usage.
• Universal copy buttons — hover-to-copy on code blocks, tool outputs, and file diffs throughout the TUI.
• Gateway providers — TokenRouter, NVIDIA NIM, OpenCode Go, and Cloudflare Workers AI via @savant-code/llm-providers.
• Theming — light/dark toggle (/theme:toggle), Neon Slate aesthetic.

SDK (@savant-code/sdk)

• SavantCodeClient class — single entry point for running agents from any Node.js / Bun / browser app.
• Streaming events — handleEvent callback receives RunState updates, tool calls, file diffs, and final output.
• Custom agents — pass agentDefinitions: AgentDefinition[] to override defaults.
• Custom tools — pass customToolDefinitions to extend the tool registry.
• Cancellation — AbortSignal propagates through subagent streams.

Agent Runtime (@savant-code/agent-runtime)

• LLM-agnostic — calls any provider registered with @savant-code/llm-providers (OpenAI-compatible chat, Anthropic, etc.).
• Multi-step loop — model decides tool → tool executes → result fed back → repeat until end_turn or budget exhausted.
• Tool registry — built-in (read_files, write_file, run_terminal_command, code_search, web_search, spawn_agents_inline, …) + custom + MCP.
• Cost aggregation — per-call token counts and USD cost estimates surfaced in RunState.

ECHO Protocol Integration

• 9 specialized agents — Orchestrator, Detective, Forge, Verifier, Recorder, Thinker, Scout, Researcher, Scribe
• FID-Bound Execution — Code is never written until the FID converges
• Perfection Loop FSM — RED → GREEN → AUDIT → SELF-CORRECT → COMPLETE
• Separation of Duties — The agent that writes code cannot verify it
• 15 Laws — 4 immutable process + 11 extended code laws

---

Repo Map

Workspace	Package	Purpose
agents/	@savant-code/agents	Public agent definitions shipped with the CLI
cli/	@savant-code/cli	CLI source — UI, commands, state, hooks, OpenTUI/React components
common/	@savant-code/common	Shared types, tool definitions, utilities
evals/	@savant-code/evals	ECHO-native benchmark v2 runner + legacy eval fixtures
packages/agent-runtime/	@savant-code/agent-runtime	Agent loop, tool executor, LLM API integration
packages/code-map/	@savant-code/code-map	tree-sitter code indexing, language detection
packages/database/	@savant-code/database	Database abstraction layer
packages/llm-providers/	@savant-code/llm-providers	Public LLM provider shims
sdk/	@savant-code/sdk	Public SDK — SavantCodeClient, types, build + verify scripts
scripts/tmux/	@savant-code/tmux	tmux CLI helpers used in interactive test runs

---

Quick Start

1. Clone and Install

git clone https://github.com/savant0x/savant-code.git
cd savant-code
bun install

2. Run the CLI (development)

# Run the CLI in dev mode
bun run dev

# Or run with a specific permission mode
bun run dev -- --permission-mode safe

3. Build for Release

# Build the SDK
bun run build:sdk

# Build the CLI binary
bun run build:binary

4. Use the SDK

import { SavantCodeClient } from '@savant-code/sdk'

const client = new SavantCodeClient({
  apiKey: process.env.SAVANT_CODE_API_KEY,
  cwd: '/path/to/your/project',
  onError: (err) => console.error('Savant-Code error:', err.message),
})

const result = await client.run({
  agent: 'savant',
  prompt: 'Add error handling to all API endpoints',
  handleEvent: (event) => console.log('Progress', event),
})

5. End-User Install

# npm
npm install -g @savant-code/cli

6. Use with local Ollama (zero API key)

If you have Ollama installed and running, Savant Code auto-detects it on first launch and routes inference to your local daemon — no API key, no account, no prompts.

# Start Ollama in the background, then run the CLI
ollama serve
savant-code

Run /health inside the chat to verify the Ollama connection, available local models, and current permission mode.

---

CLI Commands

Command	What it does
bun run dev	Launch CLI in dev mode

| bun run build:sdk | Build the SDK for npm publish | | bun run build:binary | Build the CLI binary | | bun run ci | CI gate — build SDK and release artifacts | | bun test | Run test suite | | bun x tsc --noEmit | Type check | | bun x eslint . --max-warnings 0 | Lint |

---

ECHO Protocol

This project ships with ECHO Protocol v0.2.0 — the single bootstrap file for agent behavior.

Core Principles

• FID-Bound Execution — Code is never written until the FID converges
• Perfection Loop — RED → GREEN → AUDIT → SELF-CORRECT → COMPLETE
• Separation of Duties — The agent that writes code cannot verify it
• No Deferrals — Every approved work item must be completed

15 Laws

4 immutable process laws (Read 0-EOF, Present Before Act, Verify Before Proceed, Call-Graph Reachability) + 11 extended code laws. strict: true in TypeScript.

Key Files

File	Purpose
ECHO.md	The 15 Laws + Perfection Loop FSM + FID lifecycle
ARCHITECTURE.md	Agent roster and tool restrictions
protocol.config.yaml	Build commands, quality bar, paths
dev/fids/	Feature Implementation Documents
dev/session-summaries/	Session audit trail
dev/LEARNINGS.md	Cross-session lessons

---

Configuration

What	Where	Format
ECHO Protocol runtime config	protocol.config.yaml	YAML — language, commands, quality limits
TypeScript base config	tsconfig.json	JSON — strict: true
ESLint config	eslint.config.js	Flat config
Bun config	bunfig.toml	TOML — linker: "hoisted"

---

Validation

# Build
bun run build:sdk && bun run ci

# Test
bun test

# Type check
bun x tsc --noEmit

# Lint
bun x eslint . --max-warnings 0

# Format
bun x prettier --write .

---

Documentation

• ECHO.md — The 15 Laws + Perfection Loop FSM
• ARCHITECTURE.md — Agent roster and tool restrictions
• protocol.config.yaml — Build commands, quality bar
• CHANGELOG.md — Release history
• docs/launch/landing/index.html — Public landing page
• dev/LEARNINGS.md — Cross-session lessons
• dev/session-summaries/ — Session audit trail

---

License

Apache-2.0 — see LICENSE for full text.

---

Savant-Code is the public TypeScript monorepo for the Savant-Code agent framework.

Savant • 2026