Skip to main content

🔮 OBLIVION AI

The Open-Source Terminal AI Coding Agent — Voice-Native, Self-Hosted, Zero Lock-In

A privacy-first, zero-cost alternative to Cursor, GitHub Copilot, and Claude Code — running entirely in your terminal.

PyPI version Python 3.11+ License: MIT Downloads Built with Textual

"Code is conversation. Make it natural." — M.E.E.R.A.

Quick Start • Why Oblivion • Architecture • Comparison • MCP Integration • Roadmap


🧭 What is Oblivion?

Oblivion AI is a terminal-native, fully open-source AI coding agent built around M.E.E.R.A. (Multi-modal Engineering and Reasoning Assistant) — an autonomous ReAct-style engineering partner that lives in your CLI, understands your codebase semantically, edits multiple files atomically, and talks back with a real voice if you want it to.

It was built on the premise that the best coding agent shouldn't require a subscription, a cloud account, or your code leaving your machine to be genuinely useful. Oblivion runs against free and local LLM backends by default, indexes your repository with a three-layer hybrid search engine, and exposes its own code-intelligence tools to other editors over MCP — so it plays well with the tools you already use instead of competing with them.


⚡ Quick Start

# Basic text-only install
pip install oblivion-agent

# With local Voice STT + Neural TTS support (recommended)
pip install "oblivion-agent[voice]"

# With premium ElevenLabs voice support
pip install "oblivion-agent[all]"
# Launch the Terminal UI
oblivion

# Launch the MCP server (for Claude Desktop / Cursor / Zed)
oblivion mcp

# Open the MCP Inspector in your browser
oblivion inspect

No API keys required to get started — Oblivion's default fallback chain runs on free-tier and local models out of the box.


🌟 Why Oblivion

🤖 M.E.E.R.A. Core A ReAct-loop reasoning engine, not a single-shot autocomplete — plans, calls tools, observes results, and iterates.
⚡ 13-Model Fallback Chain Auto-cascades across Ollama Cloud, Gemini, Groq, OpenRouter and more with smart rate-limit cooldowns, so you rarely hit a dead end.
🎙️ Real Voice I/O Wake-word activation ("Hey Jarvis"), silence-based auto-stop, and 8 neural voice personas via Edge-TTS / ElevenLabs.
🔍 3-Layer Hybrid Search Exact symbol lookup (SQLite FTS5) + full-text search + vector semantic embeddings (ChromaDB) with AST-aware chunking.
🔌 Built-in MCP Server Exposes 10 read-only code-intelligence tools directly to Claude Desktop, Cursor, and Zed — Oblivion augments your other tools too.
🛡️ 3-Tier Safety System Read / Mutate / Destructive action tiers, with unified diff previews before anything touches disk.
🧰 38 Agent Tools Full Git suite, test runners, atomic multi-file batch_edit, web search, PyPI/npm lookups, background server management.
📚 14 Domain Knowledge Packs Auto-detected, context-aware packs (React, Next.js, Django, Security, and more), capped at 2 per request to stay lean.
💻 Runs on Modest Hardware Developed and validated end-to-end on a 13GB RAM laptop with no GPU.
📟 Bonus: Intel 8085 Emulator A full microprocessor simulator and assembler, tucked inside the same TUI.

🆚 Oblivion vs. The Rest

Capability Oblivion AI Cursor GitHub Copilot Claude Code
Open source ✅ MIT ❌ ❌ ❌
Cost to run Free (local/free-tier LLMs) Subscription Subscription Usage-based
Runs fully in terminal ✅ ❌ (IDE fork) ❌ (IDE plugin) ✅
Multi-model fallback chain ✅ 13 backends ❌ single provider ❌ single provider ❌ single provider
Local-model support ✅ Ollama ❌ ❌ ❌
Native voice I/O ✅ wake-word + TTS/STT ❌ ❌ ❌
Hybrid symbol + semantic search ✅ 3-layer (FTS5 + ChromaDB) Semantic only Limited Semantic only
Exposes its own MCP server ✅ 10 tools ❌ ❌ ❌
Tiered write-safety with diff previews ✅ 3-tier Partial ❌ ✅
Data leaves your machine Optional (local mode available) Yes Yes Yes

Oblivion doesn't aim to replace every workflow — it aims to be the agent that works the same way whether you have a $0 budget or an enterprise API key, and hands its own code intelligence to whichever other tool you're already using.


🏗️ Architecture

3-Layer Hybrid Search Engine

                    ┌─────────────────────────────────────────┐
                    │               USER QUERY                 │
                    │   "Where is parse_llm_output defined?"    │
                    └────────────────────┬──────────────────────┘
                                          │
                   ┌──────────────────────┼──────────────────────┐
                   │                      │                      │
                   ▼                      ▼                      ▼
        ┌───────────────────┐  ┌────────────────────┐  ┌───────────────────┐
        │  LAYER 1: SYMBOL  │  │  LAYER 2: FTS5 TEXT│  │  LAYER 3: VECTOR   │
        │   SQLite Table    │  │    SQLite FTS5      │  │  ChromaDB + MiniLM │
        │  Exact Names Only │  │  Full-Text Match     │  │  Concept Matcher    │
        └─────────┬─────────┘  └─────────┬───────────┘  └─────────┬──────────┘
                  │                       │                        │
                  └───────────────────────┼────────────────────────┘
                                          │
                                          ▼
                   ┌──────────────────────────────────────────┐
                   │            MERGE & RANK ENGINE             │
                   │  1. Deduplicates exact line ranges         │
                   │  2. Symbol matches ALWAYS win rank #1      │
                   │  3. Merges vector context for fallback     │
                   └──────────────────────────────────────────┘

Reasoning & Execution Loop

 User Prompt ──▶ M.E.E.R.A. (ReAct Loop) ──▶ Tool Selection (38 tools)
                       │                              │
                       │                              ▼
                       │                    ┌─────────────────────┐
                       │                    │  3-Tier Safety Gate  │
                       │                    │  Read / Mutate /     │
                       │                    │  Destructive          │
                       │                    └──────────┬──────────┘
                       │                               │
                       ▼                               ▼
              Model Fallback Chain            Diff Preview + Execute
        (Ollama Cloud → Gemini → Groq → …)     (atomic batch_edit)

Tech stack: Python 3.11+, Textual TUI framework with cyberpunk styling, SQLite (FTS5), ChromaDB, faster-whisper (STT), Edge-TTS / ElevenLabs (TTS), Model Context Protocol (MCP).

Scale (v3.1.0, Aug 2026): ~45 Python files · ~12,000+ lines of code · 4,000+ downloads on PyPI · 38 tools (22 read-tier, 13 mutate-tier, 1 destructive, 1 special) · 6-model smart-cooldown fallback for the default configuration, expandable to 13 supported backends.


🔌 MCP Server Integration

Oblivion isn't just a standalone agent — it also serves its code-intelligence layer to other editors over the Model Context Protocol. Add this to your claude_desktop_config.json to give Claude Desktop, Cursor, or Zed direct access to Oblivion's 3-layer search and symbol index:

{
  "mcpServers": {
    "oblivion": {
      "command": "oblivion",
      "args": ["mcp"],
      "env": {
        "WORKSPACE_DIR": "/absolute/path/to/your/project"
      }
    }
  }
}

This exposes 10 read-only tools — symbol lookup, semantic search, file structure inspection, and more — so your existing editor gets Oblivion's understanding of the codebase without needing to run the full agent inside it.


🎙️ Voice-First Workflow

  • Wake-word activation: say "Hey Jarvis" to start a session hands-free.
  • Silence-based auto-stop: a 3-second pause automatically ends your utterance — no push-to-talk needed.
  • 8 neural voice personas via Edge-TTS (free) or ElevenLabs (premium), so M.E.E.R.A. can sound however you want it to.
  • Built on faster-whisper for local speech-to-text, keeping voice interaction private by default.

🛡️ Safety Model

Every action Oblivion can take falls into one of three tiers, and nothing in the Mutate or Destructive tiers runs without a unified diff preview first:

Tier Examples Confirmation
Read (22 tools) File search, symbol lookup, git log, test discovery None needed
Mutate (13 tools) Multi-file batch_edit, git commit, dependency install Diff preview shown
Destructive (1 tool) File/branch deletion Explicit confirmation required

🗺️ Roadmap

  • Jarvis-style voice personality: barge-in interruption, proactive unprompted commentary, emotion-aware tone adjustment
  • Continued expansion of the free-model fallback chain (Kimi K2, GLM-5.2, Qwen3-Coder and beyond)
  • Refactor of the TUI dashboard into smaller, more maintainable modules
  • Deeper MCP tool surface for third-party editor integrations

📄 License

Distributed under the MIT License — free for personal, academic, and commercial use.


👨‍💻 Creator & Community

Oblivion AI is designed and built by R. Rohith (@Rohith-s-hub), a final-year B.Sc. Computer Science student from Tamil Nadu, India.

⭐ If Oblivion saves you a subscription or two, consider starring the repo. ⭐

Metadata

Release files for oblivion-agent 3.3.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for oblivion-agent 3.3.1
File Size Uploaded
oblivion_agent-3.3.1.tar.gz 1.5 MB Details

Built distribution (wheel)

Table of built distributions (wheels) for oblivion-agent 3.3.1
File Interpreter ABI Platform
oblivion_agent-3.3.1-py3-none-any.whl Python 3 none any Details

Total release size: 1.8 MB

Release files / oblivion_agent-3.3.1.tar.gz

Download URL oblivion_agent-3.3.1.tar.gz
Size 1.5 MB
Tags Source
SHA-256 checksum
How to use checksums
5d40f756608fbcf44166983d41477ba3eea40933b6f502dfe67a02dda7f0d5d3
BLAKE2b-256 checksum
How to use checksums
1590f2cb6c35f5299caf9faae69e337705979734269880ee5a6f5335b12c57fe
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.2

Release files / oblivion_agent-3.3.1-py3-none-any.whl

Download URL oblivion_agent-3.3.1-py3-none-any.whl
Size 272.1 kB
Tags Python 3
SHA-256 checksum
How to use checksums
37e50b6b24481c52abfb0aaff4ced18eb08c84ca990c4be8281b899f660d387d
BLAKE2b-256 checksum
How to use checksums
9c4abf5483e882d583cf0ecaf06e4201151dbdfa54e1ae2229c950001740d10b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.2

Release history Release notifications | RSS feed

3.4.3

2 release files

3.4.2

2 release files

3.4.1

2 release files

3.4.0

2 release files

3.3.4

2 release files

3.3.2

2 release files

This release

3.3.1 This release

2 release files

3.3.0

2 release files

3.2.2

2 release files

3.2.0

2 release files

3.1.0

2 release files

3.0.0

2 release files

2.10.8

2 release files

2.10.7

2 release files

2.10.6

2 release files

2.10.5

2 release files

2.10.3

2 release files

2.10.2

2 release files

2.10.1

2 release files

2.10.0

2 release files

2.9.0

2 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page