Skip to content
MCP server

llm-sidecar

Local MCP server giving AI clients keyless web search, cited answers and claim checking

Runs locally
Install

llm-sidecar is an open-source MCP server for agent tooling. It runs locally on your machine and is started by your AI client.

Pricing
Open source
Platforms
CLI, API, macOS
Runs
Locally
Version
0.9.3
Listing updated

Similar MCP servers

Install

Run in a terminal

claude mcp add llm-sidecar -- uvx llm-sidecar

Add to claude_desktop_config.json

{
  "mcpServers": {
    "llm-sidecar": {
      "command": "uvx",
      "args": [
        "llm-sidecar"
      ]
    }
  }
}
Install in CursorOpens Cursor, which asks you to confirm. Or add it by hand:

Add to ~/.cursor/mcp.json

{
  "mcpServers": {
    "llm-sidecar": {
      "command": "uvx",
      "args": [
        "llm-sidecar"
      ]
    }
  }
}
Install in VS CodeOpens VS Code, which asks you to confirm. Or add it by hand:

Add to .vscode/mcp.json

{
  "servers": {
    "llm-sidecar": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "llm-sidecar"
      ]
    }
  }
}

Add to ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "llm-sidecar": {
      "command": "uvx",
      "args": [
        "llm-sidecar"
      ]
    }
  }
}

Add to your client's MCP settings (most use this shape)

{
  "mcpServers": {
    "llm-sidecar": {
      "command": "uvx",
      "args": [
        "llm-sidecar"
      ]
    }
  }
}

Using several servers? Build one config file with the MCP config generator.

About llm-sidecar

The README describes llm-sidecar as a local process that sits next to the tools on your machine and handles grounded question answering, web search, page reading and fact-checking, while choosing which model to run without being told. Requests are routed across Ollama for local models and OpenRouter for free or paid cloud models, and the project says a model is only used after a short probe confirms it actually answers.

According to the README, answers are built from pages fetched during the request and come with source links. When the sources disagree or do not settle a question, the tool is designed to say so instead of guessing. Claims can be graded as supported, contradicted or unverified, and whole documents can be checked claim by claim.

The project is aimed at developers who want one shared layer for grounded AI work across several local tools, and it ships as a Python package with an HTTP daemon, a CLI and an MCP server. The README marks the project as early (0.6.0) and warns that responses can take tens of seconds.

Tools this server provides (17)

Taken from the project documentation; the running server may differ.

search_web
Keyless web search.
answer_question
Answers a question from fetched sources with citations.
read_url
Returns the main text of a web page.
verify_claims
Grades claims against live evidence.
extract_claims
Pulls claims out of text.
fact_check_document
Extracts and verifies every claim in a document.
summarise
Bounded summarisation of text.
classify
Assigns text to provided labels.
extract_fields
Extracts named fields from text.
delegate
Offloads bulk work to a cheap model.
list_image_profiles
Lists image generation profiles.
generate_image
Generates an image locally on Apple Silicon.
edit_with_references
Edits an image using reference images.
upscale_image
Upscales an image.
usage_report
Reports model usage and spend.
sidecar_status
Reports daemon status.
whats_new
Lists changes since a given version.

What it does

Keyless web search
Searches the web through DuckDuckGo by default, with an optional self-hosted SearXNG instance.
Cited answers
Answers questions from fetched pages and lists the sources used, flagging when evidence is not conclusive.
Claim verification
Grades claims as supported, contradicted or unverified against live evidence, each with citations.
Routed model selection
Chooses between Ollama models and OpenRouter models by tier and budget, probing candidates before use.
Page reading
Fetches a URL and returns its main text with navigation and header content removed.
Usage ledger
Records model usage and cost so spending can be reviewed over time.

Good for

  • Give a coding assistant a way to look up current facts with source links
  • Check the factual claims in a draft document before publishing
  • Run bulk summarisation or classification on cheaper local or free models
  • Keep search and model access on one machine instead of per-tool API keys

Pricing

Open source

MIT licensed and free to self-host; optional OpenRouter usage is billed by that provider.

llm-sidecar FAQ

Does llm-sidecar need an API key?

The README says it works with no API key if Ollama is installed; an OpenRouter key is optional for free cloud models and paid routing.

How do I install llm-sidecar?

The README recommends installing with pipx, then running pipx install llm-sidecar; it also lists a run-from-source option.

Which MCP tools are exposed?

The README lists 17 tools, including search_web, answer_question, read_url, verify_claims, extract_claims and fact_check_document.

Is llm-sidecar fast?

The README says a grounded answer involves a search, page fetches and a model call, so it can take tens of seconds, longer on free models.

Is llm-sidecar free?

Yes. llm-sidecar is open source, so you can use and self-host it for free.

How do I add llm-sidecar to Claude Code?

Run this command in your terminal: claude mcp add llm-sidecar -- uvx llm-sidecar

Where this comes from

  • Description: written from the project's README and its MCP Registry entry.
  • Install steps: from the official MCP Registry entry.

How AIPoweredBot checks servers