big-AGI¶

What it is¶

big-AGI is a local-first, vendor-neutral, professional AI workspace and multi-model orchestrator. Designed for power users, researchers, and engineers, it provides a high-density, low-latency web interface to query and orchestrate multiple models simultaneously. By early January 2027, big-AGI features the Beam 2 multi-model synthesis engine, stateful code-execution sandboxes, and native FastMCP 3.1 protocol support.

What problem it solves¶

It overcomes the limitations and interface friction of single-model web clients. Instead of manually copying prompts across separate browser tabs to compare outputs, big-AGI queries frontier models (Claude 5.1, GPT-5.5, Gemini 4.0 Pro) in parallel, merges their insights via customizable consensus pipelines, and executes generated code in persistent sandboxes.

Where it fits in the stack¶

AI Assistants & Knowledge / Professional AI Workspace. It acts as a local-first control panel connecting the user's browser, remote API providers (OpenAI, Anthropic, OpenRouter), and self-hosted inference servers (Ollama, LM Studio, vLLM).

Typical use cases¶

Multi-Model Synthesis (Beam 2): Querying multiple models simultaneously, applying automated reasoning synthesis, and merging the best components into a single response.
Stateful Sandbox Code Execution: Running and debugging Python or Shell scripts within persistent, isolated containers.
FastMCP 3.1 Tool Integration: Connecting browser sessions to local databases, file systems, and custom tools via FastMCP 3.1 endpoints.
Deep Technical Research: Launching multi-step web searches with live citations, diagramming (Mermaid.js), and resumable session checkpoints.

Strengths¶

Instant Response UI: Highly optimized Next.js/React frontend handling markdown, LaTeX math, and live charts without UI lag.
Beam 2 Merge Engine: Fully customizable, program-based multi-model reasoning and voting synthesis.
Persistent Code Sandboxing: Built-in container support for safe, stateful execution of code scripts across chat turns.
Zero Lock-In Provider Support: Direct connections to 20+ model providers and local endpoints with native reasoning effort controls.
Local-First Privacy: API keys and session histories are stored locally in the browser or encrypted in transit.

Limitations¶

High Information Density: The feature-rich UI can have a learning curve for casual users seeking a basic chat client.
Browser-Bound Storage: Relying on local browser storage requires periodic manual backup exports to prevent accidental cache loss.

When to use it¶

When verifying complex architectural decisions or debugging tricky code across multiple frontier models.
When seeking a self-hostable workspace with persistent execution sandboxes and tool calling built into chat.
When requiring granular control over sampling parameters, system prompts, and tool calling schemas.

When not to use it¶

For quick, lightweight conversational chats on mobile devices where simple apps suffice.
In corporate environments that strictly forbid direct browser-to-API internet connections.

Getting started¶

1. Web App Access¶

Access big-AGI directly in your web browser at app.big-agi.com and enter your API credentials.

2. Self-Hosted (Docker)¶

Deploy a persistent, self-hosted container instance on your local machine or server:

docker run -d \
  --name big-agi \
  -p 3000:3000 \
  -e NEXT_PUBLIC_SHOW_BENCHMARKS=false \
  ghcr.io/enricoros/big-agi

Open http://localhost:3000 to access your self-hosted workspace.

CLI examples¶

# Clone and run big-AGI locally in dev mode
git clone https://github.com/enricoros/big-AGI.git && cd big-AGI
npm install && npm run dev

# Update your Docker deployment to the latest build
docker pull ghcr.io/enricoros/big-agi && docker restart big-agi

# Deploy to a private Vercel project
npx vercel --prod

API examples¶

Python: Pydantic v2 Beam 2 Configuration Validator¶

import asyncio
from typing import List, Optional
from pydantic import BaseModel, Field

class ModelWeight(BaseModel):
    model_id: str = Field(..., alias="modelId", description="Target model API identifier")
    weight: float = Field(default=1.0, ge=0.0, le=1.0, description="Voting weight")

class Beam2Config(BaseModel):
    synthesis_mode: str = Field("cot-merge", alias="synthesisMode")
    candidate_models: List[ModelWeight] = Field(..., alias="candidateModels")
    max_tokens: int = Field(2048, alias="maxTokens")
    temperature: float = Field(0.3, ge=0.0, le=2.0)
    system_instruction: Optional[str] = Field(None, alias="systemInstruction")

async def test_beam_merge_validation():
    raw_payload = {
        "synthesisMode": "majority-consensus-cot",
        "candidateModels": [
            {"modelId": "anthropic/claude-5.1-sonnet", "weight": 1.0},
            {"modelId": "openai/gpt-5.5", "weight": 0.8},
            {"modelId": "google/gemini-4.0-pro", "weight": 0.7}
        ],
        "maxTokens": 4096,
        "temperature": 0.2,
        "systemInstruction": "Synthesize the response and verify race conditions."
    }

    validated_beam = Beam2Config.model_validate(raw_payload)
    print("Beam 2 Config Validated successfully.")
    print(f"Mode: {validated_beam.synthesis_mode}")
    print(f"Active Candidates: {len(validated_beam.candidate_models)}")

if __name__ == "__main__":
    asyncio.run(test_beam_merge_validation())

LobeHub — Visual multi-agent chat interface.
OpenRouter — Unified API aggregator for model routing.
FastMCP 3.1 — Open protocol for agent tools.
AnythingLLM — RAG workspace for local documents.
Claude Code — CLI software engineering agent.

Sources / references¶

Contribution Metadata¶

Last reviewed: 2027-01-07
Confidence: high