searchio

searchio

searchio is a tool for search, page reading, and product lookup. Use it from the command line, from Python, or over HTTP. Use it as a person or as an AI agent. Search and reading need no API key.

Quick start for AI agents

An AI agent can set up searchio for its user in one pass. Follow these steps in order. Stop when the verify step passes.

1. Use Docker when Docker runs. Try:

docker run -i --rm ghcr.io/ioserviceslabs/searchio-mcp

The image holds the full system: fetch ladder, Rust engine sidecar, and a Chromium for the challenge tier. No other install step. Go to step 3.

2. No Docker: install from source. Needs Python 3.10 or newer.

pip install "searchio[mcp]"curl -fsSL https://raw.githubusercontent.com/IOServicesLabs/searchio-engine/main/scripts/install.sh | sh

The pip command installs the searchio package with the MCP server from PyPI. The curl command adds the Rust engine sidecar at ~/.searchio/bin/se-serve (linux, macOS, Windows with Git Bash). The sidecar is the fast tier-2 backend. It is optional but recommended. For the latest development build use pip install "searchio[mcp] @ git+https://github.com/IOServicesLabs/searchio".

3. Add the MCP server to the client. Claude Desktop and Cursor use this config shape.

Docker:

{  "mcpServers": {    "searchio": {      "command": "docker",      "args": ["run", "-i", "--rm", "ghcr.io/ioserviceslabs/searchio-mcp"]    }  }}

Source install:

{  "mcpServers": {    "searchio": { "command": "searchio-mcp" }  }}

4. Verify. Restart the client. Ask: "Use searchio to search the web for the current time in Tokyo." Real results mean the setup works. An error about a missing command means the config points at a path that does not exist; fix the config and retry.

Note. npx @ioserviceslabs/searchio-mcp also works as the client command in step 3. The launcher starts the Docker image when Docker runs. Without Docker it starts the local install from step 2.

1. Install

You need Python 3.10 or newer.

Path A -- pip (one command):

pip install "searchio[mcp]"searchio-mcp        # MCP server on stdio, ready for an LLM client

searchio doctor checks the install. The Rust engine sidecar is optional. Install it with the curl command in the quick start, step 2.

Path B -- Docker (no Python needed):

docker run -i --rm ghcr.io/ioserviceslabs/searchio-mcp              # stdio MCPdocker run -p 8080:8080 ghcr.io/ioserviceslabs/searchio-mcp \  searchio-mcp --transport sse --host 0.0.0.0 --port 8080

No prebuilt image for your platform? Build from source:

docker build -t searchio https://github.com/IOServicesLabs/searchio.git

The image holds the full ladder: the Python tiers, the Rust engine sidecar, and a Chromium for the challenge tier. State (clearance, sessions) persists in a volume at /root/.searchio:

docker run -i --rm -v searchio-state:/root/.searchio ghcr.io/ioserviceslabs/searchio-mcp

Path C -- from source:

git clone https://github.com/IOServicesLabs/searchio.gitcd searchiopython -m venv .venv.venv\Scripts\activate          # Windows# source .venv/bin/activate     # macOS / Linuxpip install -e '.[mcp]'searchio doctor

searchio doctor shows which fetch tiers work on your machine. Tier 0 is plain HTTP. Tier 1 adds TLS impersonation. Tier 2 adds a real browser for pages that block bots. Tier 2 starts on its own the first time it is needed.

The engine tier (optional, recommended)

Tier 2 uses the Rust searchio-engine sidecar when a se-serve binary is present. It boots in milliseconds and has no Chromium. The Docker image includes it. On a host install, one command adds it:

curl -fsSL https://raw.githubusercontent.com/IOServicesLabs/searchio-engine/main/scripts/install.sh | sh

The installer drops the binary at ~/.searchio/bin/se-serve and searchio finds it on its own. You can also set SEARCHIO_ENGINE_BIN to any binary.

MCP client setup

Point any MCP client at the server:

{  "mcpServers": {    "searchio": {      "command": "searchio-mcp"    }  }}

With Docker, use "command": "docker", "args": ["run", "-i", "--rm", "searchio"].

No Python at all? The npm wrapper picks a backend for you (Docker when present, else a local install):

{  "mcpServers": {    "searchio": {      "command": "npx",      "args": ["-y", "@ioserviceslabs/searchio-mcp"]    }  }}

2. Search

searchio search "best noise cancelling headphones 2026"searchio search "rust async runtime comparison" --intent codesearchio search "artemis mission" --domains nasa.govsearchio search "heat pump retrofit" --k 20 --json

You can set an intent: web (default), news, video, shopping, academic, code, forum, reference, local. You can also put site:nasa.gov in the query.

3. Read a page

searchio read https://example.com/some/articlesearchio read https://example.com/paper.pdf --chars 20000

The command prints the readable text. It climbs walls, bot checks, and JavaScript-only pages on its own. If a page cannot be read, you get a clear reason. You do not get an empty page.

4. Products and listings

searchio items "sony wh-1000xm5"                    # one row per product across retailerssearchio items "dyson v15" --domains bestbuy.com --jsonsearchio marketplace "ford f-150" --city seattle    # Facebook Marketplace, browser-backed

5. Research with an AI swarm

This part costs money. Set one key:

set ANTHROPIC_API_KEY=...            # Windows# export ANTHROPIC_API_KEY=...       # macOS / Linuxsearchio research "What changed in the EU battery regulation this year, and who is affected?"

To use DeepSeek instead, set SEARCHIO_LLM_PROVIDER=deepseek and DEEPSEEK_API_KEY=.... The answer comes back with findings. Each finding cites pages the workers read.

6. Run it as an MCP server

searchio is also an MCP (Model Context Protocol) server. AI agents call it over MCP. The same tools that the CLI uses are available to the agent.

Install the MCP extra:

pip install -e '.[mcp]'

Start the server over stdio:

searchio mcp# orpython -m searchio.mcp

Start the server over HTTP:

searchio mcp --transport streamable-http --host 127.0.0.1 --port 8080 --path /mcp

6.1 MCP tools

The server has these tools:

The server also has one resource: searchio://health returns service health as a small JSON object.

6.2 Payment for search and research (x402)

The MCP server has a payment gate. It uses the x402 shape. When the gate is on, a billable tool returns a 402 (Payment Required) with an X-Payment challenge until the caller pays.

Set the gate with SEARCHIO_MCP_PAYMENT:

This is the shape a real verifier uses. It has no ledger. When you host the server for others, put a real x402 verifier in front. The verifier checks payment, then the tool runs. This is how searchio is hosted next to SwarmIO for search and research credits.

7. Run it as an HTTP server

searchio serve --port 8080

Errors are honest status codes: 422 for a bad request, 400 for a target the policy refuses, 451 for a page blocked by an anti-bot wall, 502 or 503 when providers failed or are all cooling down after failures. A search answer lists providers_used, providers_failed and providers_skipped (breakers open, with the cooldown left).

If several people or agents share one server, give each its own session on /read. Then cookies from one never ride along on another's requests.

8. Run it in Docker

The Dockerfile in this repo builds the full ladder: Python tiers, the Rust engine sidecar, and a Chromium for the challenge tier. See Path B in section 1 for the build and run commands. docker compose up does the same with the state volume wired up.

Inside the container, tier 2 is the engine sidecar first; the patchright Chromium takes the pages that need a real browser (the challenge tier). Cache, learned per-domain tiers, and clearance cookies live in the searchio-state volume, so they survive restarts. Keys for the research swarm go in as environment variables, the same names as in section 10.

9. Use it from Python

import asynciofrom searchio.engine import Engine
async def main():    eng = Engine()    res = await eng.search("lithium iron phosphate battery lifespan", intent="web", k=5)    for doc in res.docs:        print(doc.title, doc.url)
    doc = await eng.read("https://example.com/article")    print(doc.text[:500])
    items = await eng.find_items("sony wh-1000xm5")    for it in items:        print(it.title, it.price, it.url)
    await eng.close()
asyncio.run(main())

10. Settings

Everything is an environment variable with the SEARCHIO_ prefix, or a line in a .env file next to where you run it. The ones people change most:

11. Good to know

• Search and reading need no keys and cost nothing. Only research calls a paid model.
• The tool never sends your credentials anywhere. Marketplace and other browser-backed lookups run anonymously.
• searchio providers lists every source and whether it is configured.
• searchio doctor is the first thing to run when something looks off.