Vectora MCP Server

Vectora

Vectora is an open-source AI assistant (Apache 2.0) built for developers — local-first, self-hosted, and designed to run as a powerful sub-agent inside any MCP-compatible orchestrator (Claude Code, Claude Desktop, Paperclip, VS Code extensions).

At its core, Vectora solves the knowledge gap problem: LLMs don't know your codebase, your docs, or the latest versions of your stack. Vectora bridges that gap with RAG (Retrieval-Augmented Generation) — ingest your docs once, and every AI interaction becomes contextually aware.

---

Why Vectora?

• Orchestrator + Specialized Agents: The Orchestrator is the primary LLM agent — it answers directly for simple queries and crafts explicit task instructions for specialists (search, coder). No wasted routing hops.
• RAG-native subgraph: Every document query goes through a full retrieve → score → rerank → inject pipeline. Results flow back to the Orchestrator for synthesis.
• 15 tools across 4 categories: Web search, vector search, file system, artifacts, memory — always available across all agents.
• Cascading embeddings: Web search results are automatically queued for embedding into LanceDB (fire-and-forget), building your knowledge base as you chat.
• Sub-agent architecture: Runs as an MCP server. Claude Code delegates complex tasks to Vectora; Vectora reasons, routes, and responds.
• Persistent memory: Cross-session memory in SQLite. Vectora remembers your preferences, project context, and decisions.
• Zero infra: SQLite + LanceDB. No Docker required for local use.
• Multi-LLM: Google Gemini (free tier), Cohere (free tier), OpenAI, Anthropic, or Ollama (fully local).

---

Architecture

Orchestrator + Workers

Every message enters through a single entry point and is routed by the Orchestrator to the right specialized agent:

START
  └─► orchestrator (responds inline OR delegates with task_query)
        ├─► [respond]      → END
        ├─► [search]       → search → search_tools → process_retrieval ↻ → END
        ├─► [coder]        → coder → coder_tools ↻ → END
        └─► [rag_subgraph] → rag_subgraph → orchestrator (synthesis) → END

RAG Subgraph

When the orchestrator routes to rag, a dedicated subgraph runs the full retrieval pipeline before synthesis:

rag_retrieve (vector_search)
  └─► rag_decide (score threshold)
        ├─► rag_inject     (score ≥ 0.7 — high confidence, inject directly)
        ├─► rag_rerank     (score 0.4–0.7 — rerank with Cohere before inject)
        └─► rag_websearch  (score < 0.4 — fall back to web + auto-embed results)

Results are injected as a SystemMessage into context. The Orchestrator then synthesizes the final answer inline, without a separate agent hop.

Artifact Tool

Agents explicitly call create_artifact to persist structured documents (plans, specs, guides, architecture decisions) to ~/.vectora/artifacts/{session_id}/ as Markdown files. The tool returns structured metadata (path, title, type, session_id, timestamp) that the Orchestrator can reference in future turns.

Cascading Embeddings

After any web_search or fetch_url call, process_retrieval automatically queues the results for embedding into LanceDB — fire-and-forget, no blocking. Your vector store grows passively as you use web search.

---

Prerequisites

Cohere — Required

Vectora uses Cohere for embeddings (embed-multilingual-v3.0) and reranking (rerank-multilingual-v3.0). It offers a generous free tier with first-class LangChain integration.

Get your key: https://dashboard.cohere.com/api-keys

Tavily — Required

Vectora uses Tavily for real-time web search and URL content extraction. It offers a generous free tier optimized for AI agents.

Get your key: https://app.tavily.com/

LLM Provider — Choose One

---

Installation

Option 1: UV — Local install (recommended)

Install Vectora globally with uv:

uv tool install vectora-agent

On first run, the setup wizard will ask for your API keys and write them to ~/.vectora/.env.

vectora        # starts chat (wizard runs automatically if no keys found)

To connect Vectora as an MCP sub-agent for Claude Code or Claude Desktop, add to your .mcp.json:

{
  "mcpServers": {
    "Vectora": {
      "command": "vectora",
      "args": ["mcp-server"]
    }
  }
}

Option 2: Docker — VPS / remote MCP server

Use this when you want Vectora running on a server and accessible from multiple machines or orchestrators via SSE.

Local (no domain):

cp .env.example .env
# Edit .env with your API keys

docker compose up -d
# SSE endpoint: http://localhost:8000/sse

VPS with Traefik (HTTPS + domain):

cp .env.example .env
# Edit .env with your API keys, VECTORA_DOMAIN and ACME_EMAIL

# Create the shared Traefik network if it doesn't exist yet
docker network create traefik-public

docker compose -f docker-compose.yml -f docker-compose.traefik.yml up -d
# SSE endpoint: https://vectora.yourdomain.com/sse

To connect from Claude Code or any MCP-compatible orchestrator:

{
  "mcpServers": {
    "Vectora": {
      "url": "https://vectora.yourdomain.com/sse"
    }
  }
}

Option 3: From Source

git clone https://github.com/brunosrz/vectora.git
cd vectora

uv sync

cp .env.example .env
# Edit .env with your API keys

uv run vectora

---

CLI Reference

vectora [options]              Start chat (resume last session for this directory)
vectora mcp-server             Start MCP server (stdio)
vectora traces                 View observability traces
vectora sessions               List all saved sessions
vectora config                 Show current configuration
vectora config --set KEY=VALUE Edit a setting

Options:
  --model MODEL        Switch LLM model (provider auto-detected). Persists.
  --ollama             Force Ollama provider (for arbitrary local model names)
  --session ID         Resume a specific session by 6-digit ID
  --new                Force a new session
  --verbosity N        Verbosity level 0–5 (0=silent, 5=debug panel). Persists.
  --version            Show version

---

Chat Commands

Input shortcuts: Enter sends, Alt+Enter or Shift+Enter adds a line break.

---

Tools Reference

15 tools across 5 categories, always available to all agents:

---

Data & Persistence

All data is stored locally in ~/.vectora/:

~/.vectora/
├── .env                    # API keys (secrets — never commit)
├── settings.json           # Runtime preferences (provider, model, verbosity)
├── data/
│   ├── vectora.db          # Sessions, memories, LangGraph checkpoints (SQLite)
│   ├── embedding_queue.db  # Async embedding queue (SQLite)
│   ├── traces.db           # Internal observability spans (SQLite)
│   └── lancedb/            # Vector store for RAG (LanceDB)
├── artifacts/              # Auto-detected plans, specs, guides
│   └── {session_id}/
│       └── *.md
├── keys/                   # Reserved for future key management
└── logs/
    ├── vectora.jsonl       # Structured JSON logs
    └── session_*.md        # Exported session audit trails

Separation of concerns:

• ~/.vectora/.env — secrets (API keys). Never versioned.
• ~/.vectora/settings.json — non-secret runtime preferences (active provider, model, verbosity, last session per directory). Managed by vectora config.

---

Tech Stack

---

Configuration

API keys go in ~/.vectora/.env (created by the setup wizard) or a project-local .env:

# LLM Provider (auto-detected from available keys if not set)
LLM_PROVIDER=google-genai
GOOGLE_API_KEY=your_key_here

# Required: RAG embeddings + reranking
COHERE_API_KEY=your_key_here

# Required: Web search + URL extraction
TAVILY_API_KEY=your_key_here

# Optional: Tracing
LANGSMITH_TRACING=false
LANGSMITH_API_KEY=your_key_here
LANGSMITH_PROJECT=vectora

Runtime preferences (model, verbosity, session history) are managed in ~/.vectora/settings.json via vectora config or the /model and /debug chat commands — no need to touch .env for these.

---

License

Apache 2.0. See LICENSE.