Stable Release

NARE CLI

Neural Amortized Reasoning Engine. AI coding assistant that remembers solutions, compiles patterns into skills, and gets faster with every task.

Traditional assistants forget everything. NARE builds a persistent library of verified solutions — common tasks become instant, complex tasks get validated.

Install NowFeaturesGitHub
nare — session
◆ NARE reasoning agent for software engineering
NareCLI ~/projects/nare
Autopilot mode · type /help for commands, Tab to cycle modes
> /cd ./teta
Changed to ~/projects/nare/teta
> find and fix the authentication bug in auth.py and login.py
●Intent: edit
●Plan (moderate)
1.Read auth.py to understand the authentication logic and identify potential bugs
2.Read login.py to understand the login flow and identify potential bugs
3.Analyze both files together to identify bugs that span across both files
4.Fix identified bugs in auth.py (syntax, logic, or security)
5.Fix identified bugs in login.py (function calls, syntax)
6.Verify that the fixes are consistent across both files
files: auth.py, login.py
>

INSTALLATION

Get running in under 60 seconds.

01 — Install
pip install narecli

Requires Python 3.10+

02 — API Key
export ANTHROPIC_API_KEY="sk-..."

Works with Anthropic API or any compatible proxy

03 — Launch
nare

Opens interactive REPL in your project directory

04 — One-Shot
nare "fix the auth bug"

Run a single task without entering the REPL

Alternative — Docker
$ docker-compose up -d
$ docker exec -it nare-cli nare

CAPABILITIES

Production-grade reasoning infrastructure.

Semantic Memory

FAISS-powered episodic storage. Every solved task is indexed — sub-100ms similarity search across your entire history.

HNSW indexing · automatic deduplication · persistent across sessions

5-Tier Routing

Queries are classified and routed through the optimal path: DIRECT → COMPILED_SKILL → FAST → HYBRID → SLOW.

Cache hits: 0 tokens, <100ms · Full synthesis: verified + oracle-tested

Compiled Skills

Recurring patterns crystallize into executable Python modules. The system learns your codebase's idioms.

Background compilation · skill quarantine · automatic versioning

Verified Synthesis

Generate → test → critique → retry. Solutions are validated before application, with automatic repair on failure.

Oracle feedback · confidence scoring · multi-candidate ranking

Agent Loop

Autonomous multi-step execution with tool registry, budget control, and extended thinking for complex tasks.

50 iterations · 200k token budget · 1 hour timeout · loop detection

Token Optimization

Prompt caching saves 90%+ on repeated context. File read cache, reasoning cache, and efficient tool workflows.

find_function → apply_hunks saves 90% vs full-file reads

ARCHITECTURE

Five-layer stack from terminal to storage.

CLI Layer
Interactive REPLAgent RendererTheme EngineThinking Display
Agent Layer
Triage AgentCoder AgentPlanning AgentAutonomous Loop
Core Engine
Reasoning RouterVerified SynthesisEvolution EngineCritic / Oracle
Memory & Storage
FAISS IndexEpisodic MemoryCompiled SkillsReasoning Cache
Tools & Execution
File I/OBash / ShellDocker SandboxWeb Search

Routing Decision Table

RouteWhen
DIRECTGreeting or cached conversational response
COMPILED_SKILLExecute pre-compiled Python skill module
FASTExact match from episodic memory (FAISS)
HYBRIDMemory context + targeted LLM generation
SLOWFull verified synthesis with critic + oracle

REFERENCE

Commands, shortcuts, and configuration.

Commands & Shortcuts

/helpShow all available commands
/read <file>Load file into context for the current session
/agentToggle autonomous agent loop on/off
/skillsList compiled skills from memory
/metricsShow performance analytics and route distribution
/historyView conversation history
/clearClear current context and chat history
/modeSwitch between Manual / Autopilot / Agent modes
TabCycle through available modes
Ctrl+LClear terminal screen
Ctrl+DExit NARE CLI

Environment Variables

ANTHROPIC_API_KEY
Your Anthropic API key (required)
Example: sk-ant-...
ANTHROPIC_BASE_URL
Custom proxy URL
Example: https://your-proxy.com
ANTHROPIC_MODEL
Model to use
Example: claude-sonnet-4-20250514
NARE_AGENT_LOOP
Enable agent loop by default
Example: 1
NARE_LOG_LEVEL
Log verbosity
Example: INFO | DEBUG | WARNING

System Requirements

  • • Python 3.10 or higher
  • • Anthropic API key (Claude Sonnet 4 recommended)
  • • 512 MB RAM minimum (FAISS index grows with usage)
  • • Windows / macOS / Linux
  • • Optional: Docker for sandboxed code execution
  • • Optional: GPT-4o key for oracle validation