1. Conduid
  2. Search
  3. Cursor Autoresearch
MCP server · Search

Cursor Autoresearch

Unclaimed search
30Low

Scored 6 months ago · breakdown

About Cursor Autoresearch

Cursor Autoresearch is an MCP server published by ergenekonyigit in the Search category.

The repository has 10 stars and 1 forks, with the last commit a month ago. Recent commit activity is the strongest signal that reported issues will get attention.

Install

From source
git clone https://github.com/ergenekonyigit/cursor-autoresearch

This server has no ConduID identity, so agent calls to it are not receipted. Pin the version you install and review the source before granting it credentials.

Ask AI

✦

Ask AI about Cursor Autoresearch

Powered by Claude · Grounded in docs

I know everything about Cursor Autoresearch. Ask me about installation, configuration, usage, or troubleshooting.

Security checks

  • ✓README presentThe repository ships documentation.

5 checks have not been run yet.

README

◈ Cursor Autoresearch

Autoresearch is a single workflow for Cursor and VS Code: MCP tools drive a measurable optimization loop, results append to autoresearch.jsonl, you can add an optional browser dashboard and packaged agent skills (autoresearch-create, autoresearch-finalize).

Field Details
Open-source project name Cursor Autoresearch
GitHub repo cursor-autoresearch (easy to discover and share)
Product / UI name Autoresearch (MCP server and marketplace extension)

Port of pi-autoresearch for editors that use pi upstream in the terminal. Same idea as karpathy/autoresearch: try a change, benchmark, keep wins, revert losses, repeat — for any primary metric (test time, bundle size, build time, Lighthouse, and more).

Autoresearch dashboard screenshot

Contents


Quick start (your own project)

Do this in the repo you want to optimize — open that folder as your workspace. The MCP server must see that folder via AUTORESEARCH_CWD (usually ${workspaceFolder}).

Prerequisites: Node.js 18+ to run the published MCP via npx (same floor as @modelcontextprotocol/sdk). Use Node.js 22+ only if you are developing this repository (pnpm install / CI — see Requirements). One-shot bootstrap also needs git on your PATH.

Cursor

  1. Add the MCP server (pick one):

    • Automated: run One-shot bootstrap — clones this repo, merges ~/.cursor/mcp.json, and symlinks skills under ~/.agents/skills/.
    • Manual: merge the JSON from the MCP config snippet into your user ~/.cursor/mcp.json (or use Cursor’s MCP UI). AUTORESEARCH_CWD must be ${workspaceFolder} so tools run in the project you opened.
  2. Restart Cursor so MCP picks up the change.

  3. Install the rule and workspace skills into that project (from the project root):

    curl -fsSL https://raw.githubusercontent.com/ergenekonyigit/cursor-autoresearch/main/scripts/add-to-your-project.sh | bash -s -- --all
    

    More options (target dir, skills-only, symlinks): Add rules and skills to your project.

  4. Install the extension (recommended) for the status bar, dashboard, and results webview: Extension (Visual Studio Marketplace or .vsix).

Then add autoresearch.md, autoresearch.sh, and drive the loop with the MCP tools — see Example: faster tests and How it works.

VS Code (GitHub Copilot + MCP)

Use the servers block in Install for VS Code. Reload the window after editing mcp.json. Steps 3–4 above still apply if you want .cursor/rules and .cursor/skills in the repo; VS Code does not load ~/.agents/skills/ automatically.


What you get

  • Closed optimization loop in your repo: one primary metric (e.g. wall-clock seconds, lower is better), agent edits code, a fixed benchmark script measures outcomes, every run is recorded so sessions can resume.
  • Three MCP tools: init_experiment → run_experiment → log_experiment (see How it works).
  • Session artifacts: autoresearch.md (goal, scope, what you tried), autoresearch.sh (repeatable benchmark; stdout must emit METRIC name=value), optional autoresearch.config.json.
  • Monorepo packages: shared engine in packages/core, packages/mcp-server (stdio MCP), packages/vscode-extension (status bar, commands, local HTTP + SSE dashboard, results webview).
  • Skills under skills/ for starting and finalizing autoresearch sessions (Cursor-oriented paths documented below).

How it works

  1. init_experiment — Name the run and set metric direction (higher/lower is better).
  2. run_experiment — Runs your benchmark (commonly ./autoresearch.sh). Parses METRIC name=value from stdout.
  3. log_experiment — Appends a line to autoresearch.jsonl. Outcomes like keep can auto-commit; discard, crash, or failed checks can revert code while keeping autoresearch files.

Typical workspace layout

File / convention Role
autoresearch.md Goal, scope, “what we tried” — keep it current.
autoresearch.sh Repeatable benchmark; metric names must match init_experiment.
autoresearch.config.json Optional — e.g. maxIterations, workingDir.
Branch name Often autoresearch/<goal>-<date>.

[!NOTE] AUTORESEARCH_CWD must point at the project you optimize (where autoresearch.jsonl and autoresearch.md live), usually ${workspaceFolder} — not necessarily the clone of this repo.


Cursor rule

.cursor/rules/autoresearch-active.mdc nudges the agent to use MCP tools and keep autoresearch.md in sync when autoresearch.jsonl exists. Install it with step 3 of Quick start (your own project), or see Add rules and skills to your project for flags and advanced options.


Add rules and skills to your project

This is Quick start (your own project) step 3 for Cursor (and optional workspace files in VS Code). Run these from the root of the project you want to equip (the folder you open in the editor). The script downloads files from GitHub by default; no clone of this repository is required.

Rules + skills (default)

curl -fsSL https://raw.githubusercontent.com/ergenekonyigit/cursor-autoresearch/main/scripts/add-to-your-project.sh | bash -s -- --all

Rules only

curl -fsSL https://raw.githubusercontent.com/ergenekonyigit/cursor-autoresearch/main/scripts/add-to-your-project.sh | bash -s -- --rules

Skills only (installs under .cursor/skills/ in that project)

curl -fsSL https://raw.githubusercontent.com/ergenekonyigit/cursor-autoresearch/main/scripts/add-to-your-project.sh | bash -s -- --skills

Install into a specific directory (without changing your shell’s current directory)

TARGET="/path/to/your/project" curl -fsSL https://raw.githubusercontent.com/ergenekonyigit/cursor-autoresearch/main/scripts/add-to-your-project.sh | bash -s -- --all

Use a local clone instead of curl (copy files) — set AUTORESEARCH_ROOT to your cursor-autoresearch checkout:

AUTORESEARCH_ROOT="/path/to/cursor-autoresearch" curl -fsSL https://raw.githubusercontent.com/ergenekonyigit/cursor-autoresearch/main/scripts/add-to-your-project.sh | bash -s -- --all

Symlink from a local clone (updates when you pull the upstream repo; requires AUTORESEARCH_ROOT):

AUTORESEARCH_ROOT="/path/to/cursor-autoresearch" curl -fsSL https://raw.githubusercontent.com/ergenekonyigit/cursor-autoresearch/main/scripts/add-to-your-project.sh | bash -s -- --all --symlink

If you already cloned this repository, you can also run the script from disk:

/path/to/cursor-autoresearch/scripts/add-to-your-project.sh /path/to/your/project --all

From this repository’s root, pnpm add-to-project runs the same script (pass a target path and flags after --):

pnpm add-to-project -- /path/to/your/project --all
Variable / flag Meaning
TARGET Project directory to write into (default: current working directory).
REF Git branch or tag on GitHub (default: main).
REPO_SLUG owner/repo used for raw.githubusercontent.com (default: ergenekonyigit/cursor-autoresearch).
AUTORESEARCH_ROOT Local checkout of this repo; when set, files are read from disk (copy) instead of downloaded.
--symlink Symlink rule and skill folders into AUTORESEARCH_ROOT instead of copying (live updates on git pull).

Install for Cursor

Details for Quick start (your own project) step 1 when you use Cursor.

One-shot bootstrap

Clones to ~/.local/share/cursor-autoresearch by default, merges ~/.cursor/mcp.json, symlinks skills:

curl -fsSL https://raw.githubusercontent.com/ergenekonyigit/cursor-autoresearch/main/scripts/bootstrap.sh | bash
Variable Default Meaning
INSTALL_DIR $HOME/.local/share/cursor-autoresearch Clone path
REPO_URL https://github.com/ergenekonyigit/cursor-autoresearch.git Git remote
SKIP_MCP — Set to 1 to skip writing ~/.cursor/mcp.json
SKIP_SKILLS — Set to 1 to skip symlinks under ~/.agents/skills/

Already cloned this repo?

pnpm install:cursor   # same as ./scripts/install.sh

Restart Cursor so MCP reloads.

~/.cursor/mcp.json

{
  "mcpServers": {
    "autoresearch": {
      "command": "npx",
      "args": ["-y", "@ergenekonyigit/cursor-autoresearch-mcp-server"],
      "env": {
        "AUTORESEARCH_CWD": "${workspaceFolder}"
      }
    }
  }
}

References: Cursor MCP, Plugins / MCP.


Install for VS Code

Use this when you work in VS Code with GitHub Copilot and MCP servers. This is the MCP part of Quick start (your own project) for VS Code.

  1. If you use npx in the config below, no local clone is required. If you prefer a local path-based setup, complete Build (from a clone).
  2. Add mcp.json:
    • Workspace: .vscode/mcp.json in the folder you open, or
    • User: Command Palette → MCP: Open User Configuration.
  3. Use the servers shape (not mcpServers):
{
  "servers": {
    "autoresearch": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@ergenekonyigit/cursor-autoresearch-mcp-server"],
      "env": {
        "AUTORESEARCH_CWD": "${workspaceFolder}"
      }
    }
  }
}
  1. Reload the window (Developer: Reload Window) or restart VS Code.

Reference: MCP configuration (VS Code).

[!TIP] Bootstrap / pnpm install:cursor only updates ~/.cursor/mcp.json. For VS Code–only setups, run the bootstrap script with SKIP_MCP=1 if you want clone + build + skills without changing Cursor’s config, then add .vscode/mcp.json or user mcp.json as above.

Skills path: Files under ~/.agents/skills/ are aimed at Cursor-style loading. VS Code does not use that path automatically — use Copilot instructions, prompts, or adapt ideas from skills/.

Rules: .cursor/rules/ apply in Cursor; for VS Code, copy the intent into project or user instructions if you want similar behavior.


Skills

If you did not use bootstrap or pnpm install:cursor:

ln -sfn "$PWD/skills/autoresearch-create" ~/.agents/skills/autoresearch-create
ln -sfn "$PWD/skills/autoresearch-finalize" ~/.agents/skills/autoresearch-finalize

You can also symlink into .cursor/skills/ (e.g. autoresearch-create → ../../skills/autoresearch-create) for workspace-only discovery. Nothing in the build requires these symlinks. The Add rules and skills to your project script can populate .cursor/skills/ for a single workspace without touching ~/.agents/skills/.


Extension

Optional but recommended after Quick start (your own project) step 4: status bar controls, local dashboard, and the results webview.

Install from Visual Studio Marketplace:

Or build a .vsix and install via Extensions: Install from VSIX…

pnpm package:extension

Shortcuts: Ctrl+Alt+X (expanded status), Ctrl+Alt+Shift+X (results webview).


Example: faster tests (e.g. Vitest)

Prompt idea: Reduce how long npm run test takes — set up autoresearch and speed it up. (Use pnpm test, pnpm vitest run, etc., to match your repo.)

  1. Create a branch such as autoresearch/vitest-speed-<date>.
  2. Add autoresearch.md and autoresearch.sh, commit so the next session can resume.
  3. MCP: init_experiment (e.g. metric vitest_seconds, lower is better) → baseline run_experiment + log_experiment.
  4. Iterate: change config → run_experiment → log_experiment with keep / discard until done or maxIterations hits.

Minimal benchmark (adjust the inner command; metric names must match init_experiment):

#!/usr/bin/env bash
set -euo pipefail
SECS="$(/usr/bin/time -p npm run test 2>&1 | awk '/^real/ {print $2}')"
echo "METRIC vitest_seconds=$SECS"

If you use pnpm exec vitest run, substitute accordingly. For very fast suites, averaging or median inside the script can stabilize the metric.


Configuration

Place autoresearch.config.json next to autoresearch.jsonl — under the directory AUTORESEARCH_CWD points at (usually ${workspaceFolder}), i.e. your optimized project, not the cursor-autoresearch clone.

Key Purpose
maxIterations Cap on counted runs in the current segment; start a new segment with init_experiment after that. Omit for no cap.
workingDir Run benchmarks / resolve paths relative to this directory (absolute or relative to workspace root). The JSON file stays at workspace root; validation fails if the path is missing or not a directory.
{
  "workingDir": "/path/to/project",
  "maxIterations": 50
}

Omit the file if defaults are enough.


Technology stack

Layer Details
Runtime Node.js ≥ 22 (CI uses 22.x; see .node-version)
Package manager pnpm 10.x (pinned in root packageManager)
Language TypeScript 6.x
Core / MCP @modelcontextprotocol/sdk, workspace package @ergenekonyigit/cursor-autoresearch-core
Extension esbuild bundle, @vscode/vsce, VS Code engine ^1.105.0
Tests Vitest 4.x (pnpm test runs all packages)
Lint ESLint 10.x (pnpm lint)

npm packages

Published packages:

Run MCP server directly with npx:

npx -y @ergenekonyigit/cursor-autoresearch-mcp-server

Requirements

  • Node.js
    • 22 or newer to develop this repository (engines.node in the root package.json, .node-version, CI).
    • 18 or newer to run the published MCP server with npx (@modelcontextprotocol/sdk declares Node ≥ 18; the published autoresearch packages do not set a tighter engines field).
  • pnpm — only needed when working from a clone of this repo; enable with Corepack (corepack enable) or install manually to match the pinned version.

Build (from a clone)

git clone https://github.com/ergenekonyigit/cursor-autoresearch.git
cd cursor-autoresearch
pnpm install
pnpm build

The MCP entrypoint after build is packages/mcp-server/dist/index.js if you prefer local path-based setup.

The Install for Cursor and Install for VS Code examples use npx -y @ergenekonyigit/cursor-autoresearch-mcp-server for simpler setup.


Develop this repository

pnpm install
pnpm build
pnpm test
Command Purpose
pnpm lint ESLint across the repo
pnpm typecheck TypeScript --noEmit for all workspace packages
pnpm build Build core, mcp-server, and the VS Code extension
pnpm test All package tests
pnpm clean Remove packages/*/dist
pnpm package:extension Produce .vsix under packages/vscode-extension/

After pnpm clean, run pnpm build before pnpm test so workspace resolution and Vitest aliases stay consistent.

Run MCP locally (after pnpm build):

AUTORESEARCH_CWD=/path/to/your/project node packages/mcp-server/dist/index.js

Contributor setup, docs deployment, local MCP, and PR expectations: CONTRIBUTING.md. Releases: RELEASING.md. Security reporting: SECURITY.md. Change history: CHANGELOG.md. License: MIT — see LICENSE.

README copy captured when this server was indexed; it may lag the repository. The original is authoritative.

Questions

About Cursor Autoresearch

How do I install Cursor Autoresearch?

No package has been published for this server, so install it from source: clone the repository and follow its README.

Is Cursor Autoresearch safe to use with an AI agent?

Its trust score is 30 out of 100 (low). It passes 1 of the 1 static security checks run so far. It has no ConduID identity yet, so agent calls to it are not receipted.

Is Cursor Autoresearch still maintained?

Yes — the last commit was a month ago. The repository has 10 stars and 1 forks.