InfoMesh

Fully Decentralized P2P Search Engine for LLMs
No credit card. No hosted API key required. No per-query fees.

Quick Start • Why InfoMesh • Features • What's New • Architecture • Security • Credits • Contributing • Docs

💡 Why InfoMesh?

The Problem

Every AI assistant needs real-time web access — but that access is gated behind expensive, proprietary search APIs:

This creates a paywall barrier for independent AI developers, open-source assistants, and researchers. Small projects and local LLMs simply cannot afford real-time web search.

The prices above are illustrative examples, not current provider quotations. Actual pricing, free tiers, logging policies, and limits vary by provider and deployment.

The Solution

I started building AI agents and quickly hit a wall: there was no free web search API. Every provider wanted a credit card, a billing account, or a monthly subscription — just to let an AI agent look something up on the web. That felt wrong.

So I built InfoMesh — a decentralized search engine where the community is the infrastructure:

• No central server — every participant is both a crawler and a search node.
• No per-query cost — contribute crawling, earn search credits. The more you give, the more you can search.
• No vendor lock-in — standard MCP protocol integration, works offline with your local index.
• No central query collection — local-only queries stay on your node; distributed search sends the query to participating peers. Local diagnostics may include query text.

InfoMesh does not compete with existing commercial search providers. Those companies serve human search at massive scale with ads-based monetization. InfoMesh provides minimal, sufficient search capabilities for AI assistants — for free, via MCP — democratizing real-time web access without per-query billing.

I just wanted my AI agent to search the web without reaching for my wallet. If you've felt the same way, InfoMesh is for you.

🆓 Free. Every Interface. No Exceptions.

No hosted API key or billing account is required. Operators may enable API-key authentication; crawl rates, credit accounting, and resource limits still apply. There are no rate limits tied to dollar payments. You run a node, you contribute to the network, and search is free — forever.

InfoMesh vs Other Web Search MCP Servers

Looking for a free web search MCP server? Here's how InfoMesh compares to common alternatives:

InfoMesh combines decentralized peer search, offline local search, no required hosted API key, and crawling/indexing of your own content without per-query billing.

🔐 Safe by Design — Layered Defenses for Your Node

Most search engines ask you to trust them. InfoMesh asks you to trust math.

There is no central InfoMesh query-collection server. Local-only searches stay on your machine; distributed searches send queries to peers, and network participation can publish or replicate indexed content. Local sessions, analytics, and diagnostic logs may retain information, so self-hosting is not a guarantee of zero query retention.

Contribute to the network → earn credits → search without per-query fees, subject to resource and crawl limits.

That's the entire deal. No catch.

How InfoMesh Layers Its Security

Unified Trust Score

Every peer earns a continuously updated trust score based on behavior, not identity:

Trust = 0.15 × uptime  +  0.25 × contribution  +  0.40 × audit_pass_rate  +  0.20 × summary_quality

Compliance Built In

Bottom line: InfoMesh uses cryptography, audits, and contribution incentives to reduce reliance on a central provider. Local-only search keeps queries local; peer search and publication share data with other nodes. There are no per-query fees, but operating limits and trust assumptions still apply.

🚀 Quick Start

Install & Run (Two Steps — No Git Required)

All you need is a Linux terminal (Ubuntu, Debian, etc.). No prior Python or developer experience required.

Step 1 — Install uv (Python package manager, one-time setup):

curl -LsSf https://astral.sh/uv/install.sh | sh

After this finishes, close and reopen your terminal (or run source ~/.bashrc). This ensures the uv and uvx commands are available.

Step 2 — Run InfoMesh:

uvx infomesh status

uvx automatically downloads and runs InfoMesh — no git clone, no pip install, and no manually managed virtual environment. InfoMesh requires Python 3.12+. The first run downloads dependencies; subsequent runs reuse the cached tool environment.

Try It Out

# Crawl a webpage and index it
uvx infomesh crawl https://docs.python.org/3/library/asyncio.html --depth 0

# Search locally, and peers when P2P dependencies are available
uvx infomesh search "asyncio"

# View the node dashboard (works over SSH too)
uvx infomesh dashboard --text

--depth 0 limits this example to one page; omitting it follows links according to the crawl configuration (unlimited depth by default). The base package does not include P2P. After installing the native prerequisites, use uvx --from 'infomesh[p2p]' infomesh search "asyncio" for peer search. An empty local index needs a successful crawl or import before it can return local results.

Install Permanently (Optional)

If you use InfoMesh regularly, install it as a persistent tool so you don't need the uvx prefix:

uv tool install infomesh

# Now run directly:
infomesh status
infomesh crawl https://example.com
infomesh search "example"
infomesh dashboard --text

Connect to Your AI Assistant (MCP)

Add InfoMesh as an MCP server in VS Code (Copilot), Claude Desktop, Cursor, or Windsurf — no hosted API key is required by default. Operators can enable authentication with INFOMESH_API_KEY.

Claude Desktop / Cursor / Windsurf use the following mcpServers format:

{
  "mcpServers": {
    "infomesh": {
      "command": "uvx",
      "args": ["infomesh", "mcp"]
    }
  }
}

VS Code (Copilot) uses a servers root in .vscode/mcp.json instead:

{
  "servers": {
    "infomesh": {
      "type": "stdio",
      "command": "uvx",
      "args": ["infomesh", "mcp"]
    }
  }
}

Your AI assistant can now crawl and search your local index via MCP without per-query billing. To enable peer search, prepare the P2P prerequisites and replace args with ["--from", "infomesh[p2p]", "infomesh", "mcp"]. The base examples above do not install P2P dependencies.

From Source (Contributors / Developers)

If you want to contribute code or run from source:

System Prerequisites

The P2P optional dependency (libp2p) includes C extensions (fastecdsa, coincurve, pynacl) that may require native build tools. These are needed for P2P installs such as uv tool install 'infomesh[p2p]' and for the repository's development dependencies, which also include libp2p.

Linux (Debian / Ubuntu):

sudo apt-get update && sudo apt-get install -y build-essential python3-dev libgmp-dev

macOS:

brew install gmp
# Xcode Command Line Tools must also be installed

Windows: Use WSL2 (recommended) or install Visual Studio Build Tools + GMP.

Note: The base package (uv tool install infomesh, or uv sync --no-dev from source) does not need P2P build dependencies. Plain uv sync includes the dev group by default, so this repository's normal contributor setup does need them.

Clone & Run

# Clone and install with dev dependencies
git clone https://github.com/dotnetpower/infomesh.git
cd infomesh
uv sync

# Start InfoMesh with the TUI dashboard
uv run infomesh start

# Or run headless (servers / CI)
uv run infomesh start --no-dashboard

Docker

docker build -t infomesh .
docker run -d --name infomesh \
  -p 4001:4001 -p 8080:8080 \
  -v infomesh-data:/data \
  infomesh

Verify It Works

# Search your local index and peers
uvx infomesh search "python asyncio tutorial"

# Check node status
uvx infomesh status

# Crawl a specific URL on demand
uvx infomesh crawl https://docs.python.org/3/

# Export your index as a portable snapshot
uvx infomesh index export backup.zst

Examples

Ready-to-run Python scripts are available in the examples/ directory:

# Local search
uv run python examples/basic_search.py "python tutorial"

# Crawl → index → search pipeline
uv run python examples/crawl_and_search.py https://docs.python.org/3/

# Programmatic MCP client
uv run python examples/mcp_client.py "async programming"

See examples/README.md for the full list.

✨ Features

Core Capabilities

MCP Integration — Free Web Search for AI Assistants

Most commercial search APIs charge per query or require a paid subscription. InfoMesh exposes 5 consolidated MCP tools without per-query billing — no hosted API key is required by default; operator-configured authentication and resource limits still apply:

Legacy tool names (search, search_local, network_stats, etc.) are still accepted for backward compatibility.

Configure in VS Code / Copilot / Claude Desktop / Cursor

This example is for Claude Desktop and Cursor. VS Code / Copilot requires the servers format shown in Connect to Your AI Assistant. These base-package examples use the local index; the same section explains how to enable P2P.

{
  "mcpServers": {
    "infomesh": {
      "command": "uvx",
      "args": ["infomesh", "mcp"]
    }
  }
}

Optional Add-ons

# Vector search with ChromaDB + sentence-transformers
pip install 'infomesh[vector]'

# Local LLM summarization via Ollama
pip install 'infomesh[llm]'

# JavaScript rendering for SPA pages (Playwright)
pip install 'infomesh[browser]'

# Chinese/Japanese/Korean tokenization (jieba)
pip install 'infomesh[cjk]'

# Everything
pip install 'infomesh[all]'

These commands install dependencies, not every runtime prerequisite. Vector search, local LLMs, and JS rendering must also be enabled/configured; JS rendering needs Playwright browser binaries, and local summarization needs an available model/runtime.

CLI Commands

infomesh start                  # Start node (auto-prompts starter index download)
infomesh start --background     # Headless mode (no dashboard)
infomesh start --role crawler   # DMZ crawler-only node
infomesh search "query"         # Search local index + peers
infomesh search --local "query" # Search local index only
infomesh crawl https://...      # Crawl a URL
infomesh dashboard              # Interactive TUI dashboard
infomesh doctor                 # System diagnostics (10 checks)
infomesh bench                  # Performance benchmarks
infomesh feedback stats         # Implicit search quality signals
infomesh feedback top-urls      # Highest-quality URLs by feedback
infomesh feeds import subs.opml # Import RSS/Atom feeds from OPML
infomesh index import --starter # Download community starter index
infomesh mcp                    # Start MCP server (stdio)
infomesh mcp --http --port 8081 # Start MCP server (HTTP)

🆕 What's New in v0.2.0

The v0.2.0 feature milestone recorded in CHANGELOG.md groups 100+ features across search intelligence, RAG support, security, observability, and developer experience. This historical milestone label is not the current package version: this checkout declares 0.1.15. The highlights below include optional capabilities and library utilities, not a guarantee that every feature is enabled in every runtime path:

Search Intelligence

RAG & Answer Extraction

Crawler Enhancements

Security & API

Developer Experience

See CHANGELOG.md for the complete list of changes.

🏗️ Architecture

Tech Stack

Search Flow (Target Latency: ~1 second)

🔒 Security & Trust

InfoMesh is designed with a zero-trust assumption — every peer is potentially adversarial. The system provides multiple layers of defense:

Content Integrity

Network Security

Key Management

• Ed25519 key pairs stored in ~/.infomesh/keys/
• Key rotation: infomesh keys rotate — rotates application signing keys and saves a dual-signed revocation record locally; it does not rotate the separate libp2p transport key
• Peer identity: the libp2p transport uses a peer ID derived from its own public key; the application signing-key identifier is a separate identity

Current propagation limitation: rotation writes .bin records, while the startup publisher scans .json files. Do not rely on automatic DHT revocation propagation from this command until that integration mismatch is corrected.

Unified Trust Score

Every peer has a continuously updated trust score:

Trust = 0.15 × uptime  +  0.25 × contribution  +  0.40 × audit_pass_rate  +  0.20 × summary_quality

🏢 Enterprise Readiness

InfoMesh includes features for production-oriented deployments, but the current package is classified as Alpha. Validate your environment and the runtime conformance limits before production use:

Split Deployment (DMZ / Private Network)

Enterprise environments can separate crawlers from indexers across network zones:

┌─────────── DMZ ──────────────┐       ┌──────── Private Network ────────┐
│                              │       │                                 │
│  infomesh start --role crawler ───────▶  infomesh start --role search │
│  (crawls the public web)     │  P2P  │  (indexes + serves queries)    │
│                              │ auth  │                                 │
│  infomesh start --role crawler ───────▶  infomesh start --role search │
│                              │       │                                 │
└──────────────────────────────┘       └─────────────────────────────────┘

Three node roles:

Configuration example (~/.infomesh/config.toml):

# DMZ Crawler node
[node]
role = "crawler"
listen_address = "0.0.0.0"

[network]
index_submit_peers = ["/ip4/10.0.0.1/tcp/4001/p2p/<INDEXER_1_PEER_ID>", "/ip4/10.0.0.2/tcp/4001/p2p/<INDEXER_2_PEER_ID>"]

# Private Search/Index node
[node]
role = "search"
listen_address = "10.0.0.1"

[network]
peer_acl = ["<CRAWLER_1_PEER_ID>", "<CRAWLER_2_PEER_ID>"]

Replace each placeholder with the corresponding node's actual libp2p peer ID. Index submissions require complete peer multiaddrs and authenticated v2 acknowledgments; bare IP/TCP addresses or legacy HTTP destinations are not supported.

CLI usage:

# Start as DMZ crawler
infomesh start --role crawler --seeds tech-docs

# Start as private indexer
infomesh start --role search --no-dashboard

Operational

• Resource Governor — CPU/memory thresholds, I/O priority, and bandwidth controls with 4 preset profiles (minimal, balanced, contributor, dedicated). Dynamic throttling responds to sampled system load and process RSS; this is not a hard memory cap or a guarantee against the OS OOM-killer
• Long-Run Resilience — StartupLock + PID cmdline validation prevent duplicate node processes for the same data directory; infomesh stop issues SIGTERM and waits for graceful shutdown before clearing the PID file
• Runtime Heartbeat — Long-running _serve workers write runtime_status.json every 10 s; the admin API exposes the latest heartbeat, degrade level, throttle factor, and process memory via /health?detail=1, /status, and /metrics
• Pre-flight Checks — Disk space and network connectivity verified before startup
• Load Guard — QPM (queries per minute) + concurrency limiting to prevent node overload
• WAL Mode SQLite — Safe concurrent reads during dashboard refresh without locking crawl writes
• Structured Logging — All library code uses structlog with machine-parseable output
• Docker Support — Dockerfile with volume mounts for persistent data; production readiness still requires deployment-specific validation

Configurable

• TOML Configuration (~/.infomesh/config.toml) with environment variable overrides (INFOMESH_CRAWL_MAX_CONCURRENT=20)
• Value Validation — Selected configuration values are validated or clamped with structured warnings; validation is field-specific, not a guarantee for every value
• Dashboard Settings — Common settings are editable via the TUI Settings tab; advanced fields still require configuration-file or CLI changes
• Energy-aware Scheduling — LLM-heavy tasks preferentially scheduled during configured off-peak hours (1.5× credit multiplier)

Compliance

• robots.txt strictly enforced — respects all crawl directives
• DMCA Takedown — Authorized signed notices persist exclusions and relay obligations; delivery to unreachable peers cannot be guaranteed within 24 hours
• GDPR — Distributed deletion records for personal data; right-to-be-forgotten support
• Content Attribution — AI-generated summaries labeled with content_hash + source URL
• Paywall Detection — fetch_page() detects and respects paywalled content
• Terms of Use — Clear TERMS_OF_USE.md covering crawler behavior and data handling

Scale

• Designed for thousands of nodes with Kademlia DHT routing
• 3-layer deduplication prevents index bloat (URL normalization → SHA-256 exact → SimHash near-duplicate)
• zstd-compressed snapshots for efficient index sharing between nodes
• Common Crawl data import for bootstrapping large indexes

💰 Earning Credits

Credits are the incentive mechanism that keeps the network healthy. They are tracked locally per node — no blockchain, no central ledger.

How Credits Work

Credits Earned = Σ (Weight × Quantity × TimeMultiplier)

Earning Actions

PR rows are ledger action weights, not proof of automatic GitHub rewards. A merged PR earns these credits only when the corresponding contribution is recorded in the ledger.

Time Multiplier

• Base actions: Always 1.0×
• LLM actions during off-peak hours (configurable, default 23:00–07:00): 1.5×
• Off-peak scheduling is energy-conscious — the network preferentially routes batch LLM work to nodes currently in off-peak

Search Cost

Fairness Guarantees

• Non-LLM participation: 10 credited crawls/hr cover 100 searches/hr at the normal Tier 1 rate (50 at the debt rate); this is credit arithmetic, not a throughput guarantee
• LLM contribution score capped: LLM actions contribute at most 60% of the effective contribution score used for tiers; this does not cap raw ledger earnings or balance
• Uptime rewards: 0.5 credits/hr just for keeping your node online, regardless of hardware
• Credit exhaustion does not block search: Even with zero credits, the ledger permits search under the rules below; resource safeguards can still limit requests

💳 Zero-Dollar Debt — No Credit Card, No Real Money

What happens when your credits run out? You keep searching.

InfoMesh doesn't cut you off. There's no paywall, no "please enter your credit card," no upgrade button. Instead, there's a simple, human-friendly recovery path:

Credits ran out
     │
     ▼
┌─────────────────────────────────────┐
│  🟢 Grace Period (72 hours)         │
│  Search works normally.             │
│  Balance goes negative — no penalty.│
└──────────────┬──────────────────────┘
               │ 72h passed, still negative?
               ▼
┌─────────────────────────────────────┐
│  🟡 Debt Mode                       │
│  Search continues at 2× cost.      │
│  Debt accumulates.                  │
└──────────────┬──────────────────────┘
               │ Earn credits → balance > 0
               ▼
┌─────────────────────────────────────┐
│  🟢 Back to Normal                  │
│  Debt cleared. Grace reset.         │
│  Full speed ahead.                  │
└─────────────────────────────────────┘

The key principle: Debt in InfoMesh is measured in credits, not money. You recover by contributing, not by paying. Run your node, crawl some pages, keep the network alive — and your debt disappears naturally.

No credit card. No dollars. No subscription. No "trial expired" popup. Just run your node, and you're back.

🤝 Contributing

We welcome contributions of all kinds — code, documentation, bug reports, feature ideas, and seed URL lists.

Getting Started

# Clone and install
git clone https://github.com/dotnetpower/infomesh.git
cd infomesh
uv sync --dev --locked

# Run the supported test suite
uv run pytest tests/ --ignore=tests/test_vector.py -x -q --tb=short

# Run linter + formatter
uv run ruff check infomesh/ tests/
uv run ruff format --check .

# Run type checker
uv run mypy infomesh/ --ignore-missing-imports

Ways to Contribute

Code Style

• Formatter: ruff format (black-compatible, 88 char lines)
• Linter: ruff with E, F, I, UP, B, SIM rules
• Type hints: Required on all public functions
• Docstrings: Required on all public classes and functions
• Tests: Every PR should include tests for new functionality
• No print() in library code — use structlog

Pull Request Workflow

1. Fork the repository
2. Create a feature branch: git checkout -b feat/my-feature
3. Write code + tests
4. Run uv run ruff check infomesh/ tests/, uv run ruff format --check ., uv run mypy infomesh/ --ignore-missing-imports, and uv run pytest tests/ --ignore=tests/test_vector.py -x -q --tb=short
5. Submit a PR — merged contributions have ledger reward weights of 1,000 – 100,000 credits; crediting requires recording the contribution, not merely opening or merging a PR.

See CONTRIBUTING.md for the full guide.

📖 Documentation

Detailed documentation is available in the docs/ directory:

📌 Documentation is also available in Korean (한국어).

📊 Project Stats

Working-tree snapshot on 2026-09-20. File/line counts include Python files under infomesh/ and tests/; source modules include package initializers. Test results below are from Python 3.13, excluding the optional vector suite and with the opt-in live LLM check skipped.

🗺️ Roadmap

The table records completed historical component milestones, not a guarantee that every runtime integration or deployment requirement is complete. Current focus is on community growth and production hardening; see the runtime conformance record for evidence and operating limits.

What's Next

• 🌍 Public bootstrap nodes — expand community-maintained seed nodes across multiple regions

  Current configuration: bootstrap/nodes.json contains one Azure East US peer. P2P-enabled nodes attempt bootstrap discovery automatically; this list is not a live availability guarantee, and firewalls/NAT may require configuration.

• 🎭 JS rendering — Playwright-based SPA crawling for JS-heavy sites
• 📱 Web dashboard — optional browser UI alongside the TUI
• 🔍 Semantic search fusion — BM25 + vector hybrid ranking with RRF
• 🌐 Multi-language stemming — language-specific tokenization and stemming

⚖️ Legal

• robots.txt: Strictly enforced. Sites that prohibit crawling are never crawled.
• Copyright: Full text stored as cache only; search results return snippets with source attribution.
• DMCA: Authorized signed notices persist exclusions and peer relay work. A 24-hour delivery deadline cannot be guaranteed for unreachable peers.
• GDPR: Distributed deletion records. Nodes can exclude pages with personal data.
• AI Summaries: Labeled as AI-generated, linked to source via content_hash, original URL always provided.
• Terms of Use: See TERMS_OF_USE.md for full terms.

🙏 Acknowledgements

InfoMesh stands on the shoulders of excellent open-source projects:

httpx • trafilatura • libp2p • SQLite • ChromaDB • Textual • FastAPI • mcp-python-sdk • uv • structlog • zstandard

MIT License — Copyright 2026 InfoMesh Contributors

If you find InfoMesh useful, consider ⭐ starring the repo — it helps others discover the project.