Getting started

Install

Silex is a local MCP server distributed as a Python package. It scans the codebase on your machine; nothing leaves it. Requires Python 3.10+.

# Requires Python 3.10+
pip install silex-archaeology

No card. No account. No telemetry. The commands below install into an isolated virtual environment — the way we run it in the field.

Windows — PowerShell

PowerShell uses $env:USERPROFILE, never %USERPROFILE%. The %VAR% syntax is cmd.exe only — in PowerShell it does not expand and silently creates a literal %USERPROFILE% folder (a 400 MB venv landed inside one in the field). Use $env:USERPROFILE in PowerShell; keep %USERPROFILE% only in classic cmd.exe.

Create an isolated venv and install the wheel into it:

# PowerShell — create the venv and install
python -m venv $env:USERPROFILE\silex-mcp\venv
& $env:USERPROFILE\silex-mcp\venv\Scripts\pip install silex-archaeology

# Sanity check — the server answers outside your client
& $env:USERPROFILE\silex-mcp\venv\Scripts\silex-mcp-server.exe --help

Linux / macOS

Same recipe with the POSIX venv layout:

# bash / zsh — create the venv and install
python3 -m venv ~/silex-mcp/venv
~/silex-mcp/venv/bin/pip install silex-archaeology

# Sanity check
~/silex-mcp/venv/bin/silex-mcp-server --help

Connect your agent

Silex speaks MCP over stdio. Add it to any MCP-aware client — Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Zed. The VS Code extension is coming soon; MCP is the path today. There are two config surfaces:

Claude Desktop — claude_desktop_config.json

On Windows this file lives at %APPDATA%\Claude\claude_desktop_config.json (open $env:APPDATA\Claude from PowerShell). Five lines are enough — the command plus SILEX_TIER=free:

{
  "mcpServers": {
    "silex": {
      "command": "silex-mcp",
      "args": ["serve", "--stdio"],
      "env": { "SILEX_TIER": "free" }
    }
  }
}

Point command at the venv binary above if it isn't on your PATH. Then fully quit and reopen Claude Desktop — the tools icon lists the silex tools.

Claude Code / VS Code — .mcp.json (primary)

Windows drive-case bug. claude mcp add writes the path into ~/.claude.json as C:/… while the VS Code extension reads c:/… — the case mismatch breaks server discovery (it stalled a field run). The robust path on Windows is a .mcp.json at the project root as the primary form; keep claude mcp add for Linux/macOS.

Create .mcp.json in the root of the project you want Silex to see:

// .mcp.json — project root (Claude Code & VS Code)
{
  "mcpServers": {
    "silex": {
      "command": "silex-mcp",
      "args": ["serve", "--stdio"],
      "env": { "SILEX_TIER": "free" }
    }
  }
}

On Linux / macOS you can instead run claude mcp add silex -- silex-mcp serve --stdio.

First scan

On first use, Silex scans the workspace and builds a local knowledge base: architecture, dependencies, workflows and business rules — each claim anchored to file:line. Your agent then queries that knowledge through the MCP tools below.

MCP tools

All tools live under the silex-archaeology namespace. The full, always-current list — names, parameters and examples — is generated from the real registry on the MCP tools reference page.

ToolReturns
create_workspaceRegisters a repository and triggers the background scan.
get_workspace_statsScan status and totals — poll until completed.
get_code_communitiesHierarchical map of the codebase communities (Leiden).
get_structural_findingsBridges, hubs and orphan areas — no LLM spend.
get_rules_by_idsBusiness rules with provenance and confidence.
generate_domain_reportA shareable Markdown report of the domain.

Evidence & confidence

Every claim carries a verdict, a confidence percentage and a provenance anchor. Verdicts:

verifiedThe claim is backed by code evidence at the anchor.
flagged ⚠The evidence points to a risk — review before changing.

Confidence is shown as a rounded percentage in monospace (96%). Anchors are file:line links into your source, e.g. Order.cs:99-131.

Unknown territory

Silex never guesses. What the scan cannot prove is reported as unknown territory, so your agent knows where the map ends.

Shareable report

Each scan produces a report you can share with your team — the same evidence your agent sees, readable by humans.

← Back to product