Nº PROOF

Benchmarks

Structural intelligence before an edit — measured on three fixture repos with deterministic proxy costs.

Methodology

Two questions per repo: What is this repository? and Is this edit safe?
Without Prism — simulated naive agent: list directories, read manifests, scan files for imports.
With Prism — Core SDK calls matching MCP tools (repository_dna, repository_overview, blast_radius).
Token estimate = bytes read ÷ 4 (common LLM heuristic). Re-run locally: bun run bench:orientation

Latest run

Recorded Aug 8, 2026, 7:09 PM

79%

Tool calls saved

70 → 15

Fewer agent steps

On small fixtures, structured Prism responses can carry more bytes than skimming a few files — the win is fewer tool round-trips and complete structural answers. Token savings widen on larger repos where naive import scans scale with file count.

Fixture	Scenario	Calls saved
m012-features	What is this repository?	−79%
m012-features	Is this edit safe?	−83%
m013-mono	What is this repository?	−75%
m013-mono	Is this edit safe?	−75%
m044-backend	What is this repository?	−79%
m044-backend	Is this edit safe?	−80%

Totals: 70 naive calls → 15 Prism calls

One-click MCP install

Add Prism to your agent in one step, then ask in plain language.

Add to Cursor Add to VS Code

Cursor / Claude Desktop — `.cursor/mcp.json` or global config

{
  "mcpServers": {
    "prism": {
      "command": "npx",
      "args": [
        "-y",
        "@repo-prism/mcp-server"
      ]
    }
  }
}

Save at the project root, then Settings → MCP → enable prism (~40 tools).

Claude Code — one command

claude mcp add prism -- npx -y @repo-prism/mcp-server

Run inside your project directory. Restart Claude Code if it was already open.