Your Terminal. 17+ Providers. 17+ Agents.
Infinite Possibilities.
The open-source, multi-agent AI coding assistant that orchestrates specialized agents
to plan, write, review, test, and publish code — all from a single terminal command.
No subscriptions. No telemetry. Your models, your way.
Watch how the multi-agent pipeline orchestrates specialized AI agents to complete complex coding tasks — all from a single terminal command.
Live Preview
Multi-Agent Orchestration
Planner → Gatherer → Writer → Reviewer → Tester — all working together autonomously.
Provider Flexibility
Switch between Groq, Gemini, OpenRouter, NVIDIA NIM, or local models mid-session.
Real-Time Progress
Watch each agent execute in real-time with streaming output and live status updates.
Context-Aware Edits
Agents understand your full codebase context before making any changes.
The Core Edge
Why Agent-Nuvira?
Not another chat bot. Not a single-prompt box. A full development team in your terminal — routing across 17+ providers, learning from every run, and stretching your free quotas further than any other AI coding assistant.
Smart Multi-Provider Token Routing
Stop paying flat premium fees. Every sub-task is routed across 17+ providers — local models (Ollama, LM Studio), free tiers (Groq, Google Gemini), and high-capacity clouds (OpenAI, Anthropic, Mistral, NVIDIA NIM, OpenRouter, Azure & more) — maximizing free-use limits and paying only when complexity demands it.
17 Specialized Agent Swarm
No generic single-prompt boxes. Your goal becomes a DAG of tasks handled by 17 dedicated agents — Planner, Context-Gatherer, Writer, Reviewer, Runner, Tester, Debugger, Security Auditor, Git/GitLab, Package, PR Review, Issue Triage and more — working in parallel.
Learning Router That Gets Better
A Thompson-sampling bandit learns per provider × complexity bucket from real outcomes (cost-adjusted rewards), with hard constraints, regex routing rules, uncertainty escalation on cold starts, and promotion gates that only keep router changes that measurably improve quality.
Local FAISS Context Indexing
Blazing-fast, private, semantic code search backed by a native FAISS vector store (pure-JS fallback) that strictly respects your .gitignore. Large contexts shrink to the top-k relevant chunks — saving tokens so free quotas stretch further.
First-Class MCP Integration
Seamlessly connect your codebase to external enterprise tools — Jira, Slack, PostgreSQL, GitHub Issues, file systems — using standard Model Context Protocol servers with SSE transport.
Real-Time Team Collaboration
Share context, synchronized vector indices, custom agents, and review pipelines across your engineering team via Git-synced config and memory — everyone ships from the same learned knowledge.
Quota Ledger + Self-Healing Failover
A central quota ledger tracks tokens per provider × model with calendar-aware resets, parks exhausted providers until free tier returns, and auto-fails-over mid-session when a key expires or rate-limit hits — never a stuck session, never a quota error thrown at you.
No Server. No Telemetry. No Subscriptions.
Everything runs locally on your machine. Free (MIT), BYO API keys, direct-to-provider connections, fully offline-capable — no intermediary server, no data leaving your machine without your control.
Features
More Than a Chat Bot. A Full Development Team in Your Terminal.
Agent-Nuvira isn't just another AI assistant — it's an autonomous pipeline of specialized AI agents that work together to ship code.
Multi-Agent Pipeline
17+ specialized agents — Planner, Writer, Reviewer, Tester, Debugger, GitLab Agent, PR Review Agent, and more — orchestrated to complete complex goals autonomously.
17+ AI Providers
Connect to Groq, NVIDIA NIM, Google Gemini, OpenRouter (200+ models), AWS Bedrock, Azure OpenAI, DeepSeek, or run locally with Ollama — all through one unified CLI. 5 built-in, 12+ configurable. Or let the Auto router pick the best one per task.
Fully Offline-Capable
Run entirely offline with local models via Ollama, HuggingFace, or GGML. No internet connection required.
Privacy-First
Your code connects directly to your chosen AI provider. No intermediary server, no telemetry, no data leaving your machine without your control.
Plugin System
Extend the CLI with custom AI providers and agent types. A programmatic SDK lets you build and register new capabilities.
Self-Learning System
Adaptive model routing, trajectory scoring, and pattern extraction that improve over time. The system learns which models perform best for which tasks.
Git & PR Integration
Built-in GitLab REST API client and GitHub PR Review Agent. Create MRs, review PRs, post inline comments, run security scans on changed files — all from the terminal.
Real-Time Team Collaboration
Share config, memory, and review pipelines across your team via Git-synced collaboration — plus the PR Review Agent checks changed files for security issues, style violations, and code quality, posting results directly on the PR.
Auto Dependency Install
Runner detects 11 project manifest types (npm, pip, bundler, cargo, go, composer, dart), installs missing dependencies on failed commands, and bootstrap-installs missing package managers — via brew, apt, winget, or rustup.
Auto Model Routing
Pick Auto and the agent picks the right model for the right task — a fast small model for quick edits, a frontier model for deep reasoning, local for private work. Complexity-aware scoring across 5 dimensions: reasoning, speed, cost, privacy, reliability.
Real Provider Pricing
Routing decisions use actual per-1K-token list pricing from each provider — free tiers count as $0 — so cost optimization reflects real-world spend, not static guesses.
Benchmark-Driven Learning
Decisions get sharper with use — real benchmark quality scores and per-agent best-model stats from your own runs are blended into every routing decision.
Central Quota Ledger
Tracks tokens/requests per provider × model with calendar-aware reset windows. Exhausted providers are parked until their free tier resets (auto re-enable, no timers), and Auto routing avoids them before a call is made. A free/local-first gate keeps paid models for complex work only — with a cost summary showing exactly how much you saved.
Checkpoint & Resume
Crash, quota kill, or token expiry mid-pipeline no longer restarts the whole plan. --checkpoint saves a resume-able snapshot after every task batch; --resume continues from the first pending step — skipping completed steps and the planner entirely.
Vector Retrieval
Large contexts are chunked, embedded locally (bge-small-en-v1.5, zero new deps) and reduced to the top-k semantically-relevant chunks before the LLM call — saving tokens so free quotas stretch further. Small contexts pass through untouched; any failure fails over to full context.
FAISS-Backed Vector Store
Similarity search runs on a pluggable backend: real FAISS native bindings when built, a pure-JS IVF ANN otherwise, with an exact-cosine fallback that always works. Auto-selected per machine — buff memory backend --check shows which tier you're on. Same JSON format, zero data migration.
Self-Healing Failover
When a provider dies mid-session (expired key, exhausted quota, deprecated model), Auto mode automatically hops to the next-best provider — and an optional confirmation prompt keeps you in control of every swap. Every failover is recorded in a live timeline.
How It Works
The Multi-Agent Pipeline
A goal goes in. Code comes out. Here's what happens in between — including GitLab integration and automated PR review from v1.32.0.
$ agent-nuvira execute <goal>
🧠Planner
Analyzes goal, creates task plan
📁Gatherer
Scans codebase for context
✍️Writer
Implements code changes
👀Reviewer
Validates for bugs & style
🧪Tester
Runs tests in sandbox
✅Verify
Security scan + quality checks
🚀Ship
Commits & publishes
Parallel execution where possible
Per-agent model configuration
Automatic error recovery
Dry-run preview mode
Providers
Choose Your AI Engine
Connect to any provider — or run locally. Switch between them per command or mid-session.
Groq
Blazing-fast LPU inference for open-source models. Free tier available with generous rate limits.
Agent-Nuvira's multi-agent pipeline is available wherever you code.
VS Code Extension — Plus the Full CLI Agent
This isn't just an editor plugin — it's the complete agent-nuvira development agent that also runs from your terminal. Inline suggestions, agent chat panel, diff viewer, and all 17 agents work in VS Code and as a full CLI: agent-nuvira execute "" runs the entire multi-agent pipeline from any shell, CI job, or editor.
Build custom agents with the @agent-nuvira/sdk package. Extend the pipeline with your own specialized agent roles.
$ npm install @agent-nuvira/sdk
Web Dashboard
Monitor provider health, visualize agent execution DAGs, browse conversation history, and track cost — all from a local web UI.
$ agent-nuvira dashboard
Git & PR
GitHub & GitLab PR Automation
New in v1.32.0 — two dedicated agents for Git platform integration: PR review automation and GitLab project management, all from your terminal.
🦊
GitLab Agent
New in v1.32.0
Full GitLab REST API v4 integration for project management directly from the terminal.
Create, list, comment on, and merge merge requests
Create and comment on issues
Check pipeline status
Discover and browse projects
Supports gitlab.com and self-hosted instances
$export GITLAB_TOKEN='glpat_your_token'
👁️
PR Review Agent
New in v1.32.0
Automated PR review with inline comments, security scanning, and quality checks via GitHub API.
List open PRs across repositories
Review specific PR (#42) or all open PRs
Post inline review comments on changed lines
Run VerifyModule security + quality scans
Post summary comment with pass/fail/blockers
$export GITHUB_TOKEN='ghp_your_token'
Use them in task plans or directly via the Planner:
$ agent-nuvira execute "review open PRs in my-org/my-repo"
Architecture
Phase 1 → 11 Migration Complete
From monolithic agent switches to 11 completed phases — the complete evolution of Agent-Nuvira's execution engine. Phase 11 adds TS Compiler API-aware structural editing with parser-level accuracy.
🔧11/11Phases Complete
🧩9/9Modules Extracted
📡4Output Formats
🧪2,880+Unit Tests
Architecture & Diagrams
17 Agents. FAISS-Backed Memory. One Engine.
A dependency-graph execution engine orchestrates a registry of 17 specialized agents, and a pluggable FAISS vector store keeps semantic memory fast, private, and token-efficient. Full Mermaid diagrams live in the repo.
17-Agent Registry
Every task in the pipeline is handled by a registered agent. The orchestrator builds a dependency graph from this registry — each with its own role:
faiss-native uses real FAISS (@faiss-node/native, FLAT_IP + L2-normalized → cosine); faiss-ivf is a pure-JS IVF-flat ANN; json is exact flat cosine. Check your active tier with:
$ agent-nuvira memory backend --check
Full Architecture Docs
Deep-dive design docs with Mermaid diagrams — module specs, extensibility registry, safe execution layer, data flow, and observability bus. They render natively on GitHub.
The registry keeps the pipeline extensible — plugins and the SDK register new agents without touching the engine. The FAISS tiers keep retrieval fast and private (local embeddings, .gitignore-aware), and token-efficient top-k context means free quotas stretch further. Fully documented, fully open.
Why Agent-Nuvira?
Built Different. Truly Open.
See how we stack up against the competition.
Feature
Agent-Nuvira
GitHub Copilot
Cursor
Claude Code
Multi-Agent Pipeline
✓ 17 agent roles
✗ Single agent
✗ Single agent
✗ Single agent
AI Provider Choice
✓ 17+ providers + plugins
✗ OpenAI only
✗ Vendor-locked
✗ Anthropic only
Offline Capable
✓ Full offline
✗ Cloud-only
✗ Cloud-only
✗ Cloud-only
Privacy / No Telemetry
✓ Direct-to-provider
✗ Routes via MS
~ Limited
~ Limited
Pricing
✓ Free (MIT) + API keys
✗ $10-39/mo
✗ $20/mo
✗ $20/mo
Plugin System
✓ Full SDK
✗ None
✗ Limited
✗ None
Self-Learning
✓ Adaptive routing
✗ None
✗ None
✗ None
Terminal-Native
✓ No IDE lock-in
~ VS Code only
~ Cursor only
~ Terminal-only
Ready to Supercharge Your Development?
Join developers who are already using Agent-Nuvira to ship code faster, with more control and zero vendor lock-in.
