AgentTasker MCP

Search the web, read websites, and run MCP tools in parallel.

Run ten searches or five calls to another MCP tool in one batch. Your agent picks concurrency up to your limit (default: 10). Background batches, search caching, rate limiting, and compact results are built in.

Python 3.10+. Zero third-party runtime dependencies. Local stdio MCP transport.

Quick start

1. Install uv and Git. uv includes uvx, which installs the server on first use. No clone needed.

2. Add AgentTasker to your coding app. Pick one option below.

Claude Code

claude mcp add --scope user --transport stdio agent-tasker -- uvx --from git+https://github.com/S3bRR/agent-tasker-mcp.git agent-tasker-mcp-server

Codex CLI / IDE extension

codex mcp add agent-tasker -- uvx --from git+https://github.com/S3bRR/agent-tasker-mcp.git agent-tasker-mcp-server

Cursor / Claude Desktop / other JSON-based clients

Create your MCP config with this JSON, or merge the entry into an existing file. Cursor uses ~/.cursor/mcp.json (global) or .cursor/mcp.json (project). Claude Desktop: Settings → Developer → Edit Config.

{
  "mcpServers": {
    "agent-tasker": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/S3bRR/agent-tasker-mcp.git", "agent-tasker-mcp-server"]
    }
  }
}

For Cursor, also add "type": "stdio" inside the agent-tasker entry.

VS Code / Copilot: copy-ready config. OpenCode: copy-ready config.

3. Restart your app and approve/enable AgentTasker if prompted. Ask:

Use AgentTasker to read https://example.com and tell me its title.

This needs no API key or search setup. You should see five tools: execute, execute_batch, get_batch, cancel_batch, and list_remote_tools.

GUI app cannot find uvx? Set command to its full path (command -v uvx on macOS/Linux, where.exe uvx on Windows). First launch needs network access; increase your client's startup timeout if necessary.

Related MCP server: local-mcp

Enable search or other MCP tools

Fetch webpages: works immediately, without an API key. Try:

Use AgentTasker to fetch these five URLs in parallel and summarize them. Use a background batch if it might take a while.

Web search: append --search-provider brave to the server arguments and set BRAVE_SEARCH_API_KEY in its environment. Or use --search-provider searxng --searxng-url http://localhost:8080 with a running SearXNG instance that enables JSON. No provider file needed. Copy-ready search setup →

Difficult websites: Markdown is preferred, with local HTML extraction otherwise. Opt in to --reader-fallback jina for hosted rendering fallback on public pages; it shares URLs with Jina and is off by default. Website reading, privacy, and curl examples →

Use AgentTasker to run ten different searches about Python asyncio with concurrency 10, in the background. Poll for results and give me a concise summary.

Other MCP tools: append --mcp-config /absolute/path/to/servers.json. Start from this example, then ask:

Use AgentTasker to call the fetch server's fetch tool for five URLs with concurrency 5.

AgentTasker starts its own configured stdio connections; it cannot reuse your harness's existing MCP sessions. Remote tools are not cached or retried because they may have side effects. Provider quotas still apply; search presets default to one request/second.

Prefer a local install?

On macOS/Linux, with Python 3.10+:

git clone https://github.com/S3bRR/agent-tasker-mcp.git
cd agent-tasker-mcp
./setup.sh --client cursor

Replace cursor with claude, codex, vscode, opencode, or generic. Paste the printed config into your app's config file, then restart the app. The installer never overwrites your settings. Keep the checkout/virtual environment in place.

# Include search or other MCP servers in the generated config:
./setup.sh --client codex --search-provider brave --mcp-config /absolute/path/to/servers.json

# Regenerate config later, without reinstalling:
.venv/bin/agent-tasker-mcp-server --print-config vscode

Windows installation and full harness setup →

Reference

• Batch examples · Repeated MCP calls

• Tools, limits, caching, cancellation, and migration

• MIT license

Only connect servers you trust. Cancelling a batch skips queued work; already-running calls may finish. Background jobs and caches are process-local and disappear when the server restarts.