PluginBench
MCP Server
Active
MIT

Prism MCP Server

io.github.Shailesh200/prism

Local-first repository intelligence for AI agents: understand codebase DNA, health, blast radius, and impact analysis without uploading data.

What is the Prism MCP server?

Prism is a local-first software intelligence engine that provides AI agents with deep repository analysis. It exposes 28 read-only tools via MCP to help agents understand codebase structure, assess change impact, and answer questions about repository health and dependencies—all without sending code to external servers.

Prism analyzes your repository locally to give AI agents like Claude and Cursor intelligent context about your codebase. It maps dependencies, calculates blast radius for changes, identifies health issues, detects import cycles, and explains code ownership. Agents can ask natural-language questions ("What breaks if I edit this file?" or "Is this repo healthy?") and Prism handles the analysis, eliminating the need for agents to manually explore directories or guess at impact.

How to install Prism

Copy-paste configuration for popular MCP clients.

transport: stdio
Config generated by PluginBench — verify against the source before use.
Environment / auth
  • PRISM_WORKSPACE

    Optional override when the MCP client starts from a directory outside the project.

  • CURSOR_API_KEY
    secret

    Optional override to spawn local Cursor SDK workers. Prefer prism init / first start_job (Cursor.auth.login). Do not put this in mcp.json for normal use.

  • PRISM_DISPATCH_ROLE

    Set to worker to skip Intelligence tools and hide start_job / start_my_day / init. Dispatch job agents do not attach Prism MCP.

~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "prism": {
      "command": "npx",
      "args": [
        "-y",
        "@repo-prism/mcp-server",
        "workspace"
      ],
      "env": {
        "PRISM_WORKSPACE": "<YOUR_PRISM_WORKSPACE>",
        "CURSOR_API_KEY": "<YOUR_CURSOR_API_KEY>",
        "PRISM_DISPATCH_ROLE": "<YOUR_PRISM_DISPATCH_ROLE>"
      }
    }
  }
}

Tools & capabilities

Tools this server exposes to the agent.

  • repository_dna — Identify languages, frameworks, domains, and technology stack of the repository
  • repository_health — Calculate health score and identify health factors
  • repository_map — Generate a repository map with configurable zoom levels
  • blast_radius — Assess the impact and blast radius of changes to a target file or path
  • test_impact — Find related tests affected by a change
  • safe_delete — Determine if a file or module can be safely deleted
  • dependency_graph — Analyze dependency relationships and graph structure
  • dependency_cycles — Detect import cycles in the codebase
  • find_symbol — Find symbol declarations in the repository
  • find_references — Find all references to a symbol
  • review_changes — Review and analyze changes in a diff or branch
  • explain_area — Explain what a code path or area is for and its ownership
  • explore_code — Explore usages, ownership, and similar code patterns
  • engineering_health — Generate engineering health report with entropy, drift, debt, and hotspots
  • testing_report — Analyze test suites and coverage
  • security_report — Left-shift security checklist and findings
  • backend_report — Analyze routes, data models, environment, and jobs
  • landmarks — Identify repository entrypoints and anchor files
  • list_packages — List workspace packages
  • stack_profile — Analyze stack signals and technology profile

Use cases

  • Ask an AI agent "What breaks if I change this file?" to get instant blast radius analysis without manual exploration
  • Review pull requests by asking "Is this change safe?" and get impact assessment across dependencies and tests
  • Understand unfamiliar codebases by asking "What is this repository?" to get DNA, landmarks, and health in seconds
  • Detect import cycles and structural debt with automated cycle detection and engineering health reports
  • Assess whether code can be safely deleted by analyzing all usages and dependents across the repository

Prism MCP server FAQ

What is the Prism MCP server?

Prism is a local-first repository intelligence engine that exposes 28 tools to AI agents via MCP. It analyzes your codebase locally—without uploading anything—to answer questions about structure, health, change impact, and dependencies.

Is Prism free?

Yes. Prism is open-source (MIT license) and free to use. Core analysis runs entirely on your machine with no network requests or account required.

How do I install Prism in Cursor?

Click the one-click install link in the docs, or manually create `.cursor/mcp.json` with the config: `{"mcpServers": {"prism": {"command": "npx", "args": ["-y", "@repo-prism/mcp-server"]}}}`. Then enable it in Settings → MCP.

How do I use Prism with Claude Desktop?

Add the same npx config to your Claude Desktop MCP settings file, then restart Claude. You can then ask questions in plain language and Claude will call Prism's tools automatically.

Does Prism upload my code?

No. Core analysis is completely local and makes no network requests. Optional features (GitHub, PageSpeed, etc.) are off by default and gated by consent in `~/.prism/workspaces/<key>/consent.json`. Agents cannot grant consent.

What does Prism analyze?

Prism analyzes repository structure, dependencies, code ownership, test coverage, technology stack, health metrics, import cycles, and change impact. It works with any language and framework.

README (reference)

Source of truth, from the repository.

<!-- smoke-test: dispatch token telemetry -->

Prism

Node.js Website npm @repo-prism/cli npm @repo-prism/mcp-server VS Marketplace Open VSX License: MIT

Local-first Software Intelligence Engine for humans and AI agents.

Website: https://www.prismhq.in · Docs: https://www.prismhq.in/docs · Get started: https://www.prismhq.in/docs/start/get-started

Maps, graphs, blast radius, health — on your machine. No account. Nothing uploaded for core analysis.

Prism is not an AI coding assistant. It is the intelligence layer behind an IDE extension, a CLI, and an MCP server that agents can call.


Table of contents

  1. Get started in 60 seconds
  2. Install — step by step
  3. CLI — commands
  4. MCP — tools & how agents use them
  5. Docs
  6. Privacy
  7. Build from source
  8. License

Get started in 60 seconds

Needs Node.js 22+. No need to clone this repository.

# 1) Go to your project
cd /path/to/your/project

# 2) CLI — no --workspace; uses git root from where you run
npx -y @repo-prism/cli doctor
npx -y @repo-prism/cli dna
npx -y @repo-prism/cli health

# 3) MCP once (Claude Code), then ask in plain English — no tool names:
#    claude mcp add prism -- npx -y @repo-prism/mcp-server
#    “How healthy is this repo?” / “What breaks if I change src/index.ts?”
You want…Follow
Terminal / CIA. CLI
Cursor / Claude / Codex agentsB. MCP
Visual map & dashboardsC. IDE extension
Embed in code@repo-prism/core

All surfaces call the same Core SDK. None re-implements analysis.


Install — step by step

A. CLI — @repo-prism/cli

  1. cd into your project.
  2. Run doctor:
    npx -y @repo-prism/cli doctor
    
  3. Confirm Workspace points at your repo (from git root is normal). A warn on Index cache on first run is expected.
  4. Analyse:
    npx -y @repo-prism/cli dna
    npx -y @repo-prism/cli health
    npx -y @repo-prism/cli blast src/index.ts --fail-on high
    
  5. Optional global install:
    npm install -g @repo-prism/cli
    prism doctor
    prism health --verbose   # shows workspace + index timing
    

Package: npmjs.com/package/@repo-prism/cli · Guide: CLI docs

B. MCP — @repo-prism/mcp-server

Set up once. Then ask the agent in plain language — never type tool names like repository_health. The server tells the agent which tools to call.

One-click (Cursor)

Add to Cursor · Install guide · Benchmarks

Copy-paste config (Cursor / Claude Desktop)

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

Save as .cursor/mcp.json (project) or add to Claude Desktop config. Enable in Settings → MCP (~40 tools).

Claude Code

  1. cd /path/to/your/project
  2. claude mcp add prism -- npx -y @repo-prism/mcp-server
  3. Restart Claude Code if it was already running.
  4. Ask: “What is this repository?” or “How healthy is this codebase?”

Cursor (manual)

If the one-click link above does not work:

  1. Open the project folder in Cursor.
  2. Create .cursor/mcp.json:
    {
      "mcpServers": {
        "prism": {
          "command": "npx",
          "args": ["-y", "@repo-prism/mcp-server"]
        }
      }
    }
    
  3. Settings → MCP → enable prism → wait for ~40 tools.
  4. In Agent chat, ask normally (no tool names).

Claude Desktop / Codex

Same npx / @repo-prism/mcp-server config — no --workspace. Full numbered steps for every client: MCP install.

Only add --workspace / PRISM_WORKSPACE if auto-detection picks the wrong folder.

Package: npmjs.com/package/@repo-prism/mcp-server

C. IDE extension (Prism)

  1. Install Prism (Marketplace / Open VSX id prismhq.repo-prism):
  2. Open Folder on your project.
  3. Command Palette → Prism: Open Prism.
  4. Wait for indexing; use the UI.
  5. (Cursor) Also complete B. MCP so the agent shares the same analysis — IDE install.

CLI — commands

Global flags: -w/--workspace, --json, --no-color, -q/--quiet, --verbose, -y/--yes.

Exit codes: 0 ok · 1 gated finding (--fail-on) · 2 usage error · 3 Prism failed.

Diagnostics

CommandPurpose
prism doctorEnvironment, workspace chosen, index
prism indexBuild / refresh the index

Understand

CommandPurpose
prism dnaLanguages, frameworks, domains, stack
prism healthHealth score + factors
prism map [--zoom repo|package|feature|file|symbol]Repository map
prism explain <path>What a path is for / ownership
prism explore <target>Usages, ownership, similar code
prism stackStack signals
prism featuresInferred features
prism landmarksEntrypoints & anchors
prism packagesWorkspace packages

Assess a change

CommandPurpose
prism blast <target> [--fail-on low|mid|high]Blast radius
prism review [paths…] [--base <ref>]Change review
prism safe-delete <target>Safe to delete?
prism rename <target> <newName>Rename edit sites
prism test-impact <target>Related tests

Structure

CommandPurpose
prism depsGraph size / hubs
prism cycles [--fail-on any]Import cycles
prism symbol <name>Find declaration
prism refs <name>Find references
prism route <from> <to>Dependency path

Reports

CommandPurpose
prism engineeringEntropy, drift, debt, hotspots
prism testingSuites + on-disk coverage
prism securityLeft-shift checklist
prism backendRoutes, data, env, jobs
prism bundleBundle weight (ingested stats)

CI sketch:

- uses: actions/setup-node@v4
  with:
    node-version: "26.5.0"
- run: npx -y @repo-prism/cli review --base origin/main --fail-on high
- run: npx -y @repo-prism/cli cycles --fail-on any

MCP — tools & how agents use them

28 read-only tools + 3 optional prompts (orient, before_edit, review_diff). No network. No consent APIs for agents.

You do not type tool names. After MCP is connected, ask in plain language — the server’s instructions tell the agent when to call which tool.

You sayAgent should call
“What is this repo?”repository_dna / landmarks
“Is it healthy?”repository_health
“What breaks if I edit src/…?”blast_radius, test_impact
“Can I delete this?”safe_delete
“Review my changes”review_changes
GroupTools
Orientationrepository_dna, repository_health, repository_map, repository_overview, list_packages, stack_profile, landmarks, explain_area
Graphsdependency_graph, dependency_cycles, knowledge_graph, feature_graph, list_features, find_symbol, find_references, dependency_route
Impactblast_radius, safe_delete, rename_impact, test_impact, breaking_change_hints, review_changes, explore_code
Reportsengineering_health, health_history, backend_report, testing_report, security_report

Setup steps: MCP install · Full tool list: MCP tools.


Demo — with vs without Prism

Side-by-side script: plans/notes/M-063-demo.md · measured savings: prismhq.in/benchmarks

  1. Ask “What is this repository?” — Prism returns DNA + landmarks in 2–3 calls; a naive agent walks directories.
  2. Ask “Is this edit safe?” on a file — blast_radius replaces scanning the tree for imports.

Reproduce numbers: bun run bench:orientation (after bun run build).


Docs

Site: https://www.prismhq.in/docs

TopicLink
What Prism isdocs/what-is-prism
Install (CLI / IDE / MCP)docs/start/install
Quickstartdocs/start/quickstart
Capabilitiesdocs/reference/capabilities
Task guidesdocs/guides
Usage (CLI / IDE / MCP)docs/usage
Dispatchdocs/guides/dispatch
CLI referencedocs/reference/cli-commands
MCP toolsdocs/reference/mcp-tools
Known limitationsdocs/help/known-limitations
FAQdocs/help/faq
ChangelogCHANGELOG.md

Markdown sources also live in docs/ in this repo. Local preview: bun run docs:dev


Privacy

Core analysis makes no network requests (proven by a trap test suite). Optional features (GitHub, PageSpeed, Gravatar, …) are off by default and gated per purpose in ~/.prism/workspaces/<key>/consent.json. Agents cannot grant consent.

→ PRIVACY.md · SECURITY.md


Build from source

For contributors (not required to use CLI/MCP):

git clone https://github.com/Shailesh200/prism
cd prism
nvm use          # Node 26.5.0
bun install
bun run build
bun run verify:milestone

See CONTRIBUTING.md.

Published npm packages under this org: @repo-prism (cli, mcp-server, core, and engine libraries).


License

MIT

Related MCP servers

DEDevTime logo

DevTime

Active

Local-first, read-only MCP server that verifies repository claims against code evidence without execution or cloud calls.

11
Python
Apache-2.0
View repository →
MCmcp-airlock logo

Governance proxy for MCP servers: allowlist, dry-run enforcement, human confirmation, blast-radius limits, and audit trails.

27
Python
MIT
View repository →

Autonomous AI sports ad exchange with quote-bound Stripe settlement and live delivery telemetry.

View repository →
SHshuck-convert logo

shuck-convert

Maintained

Convert between document formats — DOCX/PDF to Markdown, Markdown to DOCX.

1
Python
MIT
View repository →
SHshuck-file logo

shuck-file

Maintained

Any file in, Markdown out — read only what matters.

6
Python
MIT
View repository →

On-chain economy for AI agents: escrow, marketplace, data rights, reputation, disputes.

3
Python
View repository →