captain-search 0.1.0

Unified web, code, and webpage search for MCP agents and terminal workflows

Captain Search

CLI-first web and code search for agents, with MCP when you need it
One CLI. Multiple providers. Clean Markdown.

---

Quick Start

Captain Search should usually start as a CLI workflow, not an MCP setup flow. Install the skill or run the CLI first, prove one command works, and only then add MCP if you explicitly need a server.

From this checkout today

git clone https://github.com/mnm-matin/captain-search.git
cd captain-search
uv sync
uv run csearch skill install --scope user
uv run csearch web "openai api" --max-results 5

After the first PyPI release

uvx --from captain-search csearch skill install
uvx --from captain-search csearch web "openai api" --max-results 5

---

Supported Providers

Web Search

You only need one provider to get started. Add more for redundancy.

Provider	Free Tier	Best For	Get API Key
Parallel	$20 on first signup	AI-native web search	platform.parallel.ai
Serper	2,500/month	Google results	serper.dev
Brave	2,000/month	Independent index	brave.com/search/api
Tavily	1,000/month	AI-optimized results	app.tavily.com
Exa	$10 credit	Neural/semantic search	dashboard.exa.ai
Perplexity	$5/mo credit	AI-powered answers	perplexity.ai/settings/api

💡 Tip: Serper + Brave = 4,500 free searches/month with automatic failover

Code Search

Provider	Free Tier	Best For	Get API Key
Exa (MCP)	Unlimited*	Semantic code context	exa.ai
Morph (Warp Grep)	Paid	Repo-local agentic search	morphllm.com
grep.app	Free (no key)	Exact text matching	grep.app
DeepWiki	Free (MCP)	Repo Q&A / Architecture	deepwiki.com
Noodlbox	Temporarily disabled	Local graph analysis	noodlbox.io

Webpage Extraction

Provider	Free Tier	Best For	Get API Key
Parallel Extract	Uses the $20 signup credit	JS-heavy pages, PDFs, clean markdown	platform.parallel.ai
Jina	1M tokens	Webpage/PDF extraction	jina.ai/reader
Trafilatura (fallback)	Free	Local HTML extraction fallback	trafilatura.readthedocs.io
MarkItDown (optional fallback)	Free	Local document → Markdown (DOCX/PPTX/XLSX/...)	github.com/microsoft/markitdown

Enable MarkItDown locally with: uv sync --extra markitdown (minimal) or uv sync --extra markitdown-all (all formats)

---

Installation

Skill install

If you want an agent to keep reusing Captain Search across sessions or repos, install the Captain Search CLI skill first.

# User-level install from this checkout
uv run csearch skill install --scope user

# Project-local install for the current repo
uv run csearch skill install --scope project

# Claude Code-compatible skill directory
uv run csearch skill install --target claude

# Once published to PyPI
uvx --from captain-search csearch skill install

The installer writes captain-search-cli under .agents/skills by default and can target .claude/skills with --target claude. From a source checkout, --runtime auto bakes in uv run --directory /path/to/captain-search csearch .... From a published package, it defaults to uvx --from captain-search csearch .... Use --force to refresh an existing install.

MCP client install instructions (copy/paste)

Only use this section when the user explicitly wants MCP client setup.

These MCP client configs intentionally use uv run --directory ... because the client may launch the server from an arbitrary working directory and should not depend on your shell having the project virtualenv activated.

You are installing the Captain Search MCP server. First ask the user which API keys they want to configure: SERPER_API_KEY, BRAVE_API_KEY, TAVILY_API_KEY, PERPLEXITY_API_KEY, PARALLEL_API_KEY, EXA_API_KEY, JINA_API_KEY, MORPH_API_KEY. Then ask which client they want to set up (Cursor, Claude Code, VS Code, Windsurf/Antigravity, OpenCode). Use only the keys they choose and remind them at least one web-search key is required.

Then provide the client-specific setup:
- Cursor: give the Captain Search Cursor deeplink and tell them to set the chosen keys as environment variables, then restart Cursor.
- Claude Code: use `claude mcp add captain-search -- uv run --directory /path/to/captain-search csearch mcp` and pass only the chosen keys with `-e KEY=value`.
- VS Code: use the `code --add-mcp` command with only the chosen env keys (or leave them blank if the user wants to fill later).
- Windsurf/Antigravity: add the MCP server to ~/.codeium/windsurf/mcp_config.json with the chosen env keys.
- OpenCode: add to opencode.jsonc under `mcp` with `type: "local"`, `command: ["uv", "run", "--directory", "/path/to/captain-search", "csearch", "mcp"]`, and an `environment` object with the chosen keys.

Install in Cursor

Install in Cursor

API keys are pulled from your environment (e.g., SERPER_API_KEY). Set them in your shell or system settings, then restart Cursor.

Install in Claude Code

claude mcp add captain-search -e SERPER_API_KEY=your-key -- uv run --directory /path/to/captain-search csearch mcp

Install in VS Code

Install via terminal:

code --add-mcp '{"name":"captain-search","command":"uv","args":["run","--directory","/path/to/captain-search","csearch","mcp"],"env":{"SERPER_API_KEY":"your-key-here","BRAVE_API_KEY":"","TAVILY_API_KEY":"","PERPLEXITY_API_KEY":"","PARALLEL_API_KEY":"","EXA_API_KEY":"","MORPH_API_KEY":"","JINA_API_KEY":""}}'

Or add to your User Settings (JSON) via Ctrl+Shift+P → Preferences: Open User Settings (JSON):

{
  "mcp": {
    "servers": {
      "captain-search": {
        "command": "uv",
        "args": ["run", "--directory", "/path/to/captain-search", "csearch", "mcp"],
        "env": {
          "SERPER_API_KEY": "your-key-here",
          "BRAVE_API_KEY": "",
          "TAVILY_API_KEY": "",
          "PERPLEXITY_API_KEY": "",
          "PARALLEL_API_KEY": "",
          "EXA_API_KEY": "",
          "MORPH_API_KEY": "",
          "JINA_API_KEY": ""
        }
      }
    }
  }
}

Or add to .vscode/mcp.json in your workspace to share with your team.

Install in Windsurf / Antigravity

Add to your ~/.codeium/windsurf/mcp_config.json (shared config for Windsurf and Antigravity):

{
  "mcpServers": {
    "captain-search": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/captain-search", "csearch", "mcp"],
      "env": {
        "SERPER_API_KEY": "your-key-here",
        "BRAVE_API_KEY": "",
        "TAVILY_API_KEY": "",
        "PERPLEXITY_API_KEY": "",
        "PARALLEL_API_KEY": "",
        "EXA_API_KEY": "",
        "MORPH_API_KEY": "",
        "JINA_API_KEY": ""
      }
    }
  }
}

Install in OpenCode

Add to your opencode.jsonc (project root) under mcp:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "captain_search": {
      "type": "local",
      "command": ["uv", "run", "--directory", "/path/to/captain-search", "csearch", "mcp"],
      "enabled": true,
      "environment": {
        "SERPER_API_KEY": "your-key-here",
        "BRAVE_API_KEY": "",
        "TAVILY_API_KEY": "",
        "PERPLEXITY_API_KEY": "",
        "PARALLEL_API_KEY": "",
        "EXA_API_KEY": "",
        "MORPH_API_KEY": "",
        "JINA_API_KEY": ""
      }
    }
  }
}

Getting Started

git clone https://github.com/mnm-matin/captain-search.git
cd captain-search
uv sync
uv run csearch skill install --scope user

If you want a no-activation, one-off run from this checkout, use uvx:

uvx --from . csearch --help
uvx --from . csearch web "openai api"

That creates console scripts in the project virtualenv. If you activate it, you can run the CLI directly:

source .venv/bin/activate
captain-search --help
csearch --help

If you do not want to activate the virtualenv, you can still call the wrappers directly:

.venv/bin/captain-search --help
.venv/bin/csearch --help

uv run is just the convenience wrapper that runs those same commands inside the project environment.

If you want a persistent command on your shell PATH, install the tool:

uv tool install .

If you publish Captain Search to PyPI, end users can skip cloning entirely:

# One-off runs
uvx --from captain-search csearch --help
uvx --from captain-search csearch web "openai api"

# Persistent install
uv tool install captain-search

Then update /path/to/captain-search in the configs above to your actual path.

---

Available Tools

Tool	Description
search_web	Search with weighted selection and optional multi-provider mode
search_code	Search code across Exa, grep.app, DeepWiki, Morph, and local exact matches
fetch_webpage	Extract content from any URL (articles, PDFs, docs)

search_web provider selector:

• auto (default): weighted single-provider selection with fallback
• multi or all: parallel search across all enabled providers
• Provider name: serper, brave, tavily, perplexity, parallel, exa, exa_mcp
• Comma-separated list for multi-provider search

search_code providers:

• Exa Code Context: Semantic search (always runs)
• grep.app: Exact text matching (always runs)
• DeepWiki: Repo Q&A (requires repo filter)
• Morph Warp Grep: Repo-local agentic search (requires repo filter + MORPH_API_KEY)
• Local Exact Matches: Repo-local fixed-string search (requires repo filter)
• Noodl: Temporarily disabled

CLI Usage

captain-search and csearch expose the same CLI. The docs use csearch, and MCP server startup now requires an explicit mcp subcommand.

# Bare invocation shows help
csearch
csearch version
csearch skill install --scope user

# Explicit MCP server startup
csearch mcp
csearch mcp --transport http --port 8000

# Direct CLI commands
csearch web "openai api" --max-results 5
csearch web "openai api" --all
csearch code "search_web" --repo mnm-matin/captain-search --format json
csearch fetch https://example.com --format json

# No-activation ephemeral runs with uvx
uvx --from . csearch web "openai api"

# Once published to PyPI
uvx --from captain-search csearch web "openai api"

# No-activation form if you want uv to launch the project env for you
uv run csearch web "openai api"

# Module form also works
uv run python -m captain_search web "openai api"

Direct command exit codes are shell-friendly: 0 for successful output, 1 for a top-level command error, and 2 for invalid CLI usage.

---

Running as a Remote Server

For teams or cloud deployment:

# HTTP mode
captain-search mcp --transport http --port 8000

# SSE mode  
captain-search mcp --transport sse --port 8000

With Authentication

export MCP_AUTH_TOKEN="your-secret-token"
captain-search mcp --transport http --port 8000

Connect via:

• Header: Authorization: Bearer your-secret-token
• Query: http://host:8000/mcp?token=your-secret-token

---

Environment Variables

Variable	Required
SERPER_API_KEY	At least one provider
BRAVE_API_KEY	At least one provider
TAVILY_API_KEY	At least one provider
PERPLEXITY_API_KEY	Optional
PARALLEL_API_KEY	Optional (web search + webpage extraction)
EXA_API_KEY	Optional
MORPH_API_KEY	Optional (code search)
MORPH_BASE_URL	Optional (defaults to https://api.morphllm.com/v1)
JINA_API_KEY	Optional (works without)
MCP_AUTH_TOKEN	For remote mode
CAPTAIN_SEARCH_LOG_ENABLED	Optional (defaults to true)
CAPTAIN_SEARCH_LOG_DIR	Optional (defaults to ~/.captain-search/logs)
CAPTAIN_SEARCH_LOG_FULL_PAYLOADS	Optional (defaults to true)

Every keyed provider also supports a comma-separated *_API_KEYS environment variable for rotation, for example SERPER_API_KEYS, BRAVE_API_KEYS, TAVILY_API_KEYS, PERPLEXITY_API_KEYS, PARALLEL_API_KEYS, EXA_API_KEYS, and JINA_API_KEYS.

---

Telemetry Logging

Captain Search writes daily JSONL telemetry files to ~/.captain-search/logs/YYYY-MM-DD.jsonl by default.

Each tool call logs:

• full input arguments
• final output returned to the client
• provider attempts and per-provider results
• repo resolution and cache events for search_code
• Parallel/Jina/fallback stages for fetch_webpage

If you want to reduce what gets stored, set CAPTAIN_SEARCH_LOG_FULL_PAYLOADS=false to log previews instead of full payloads. Set CAPTAIN_SEARCH_LOG_ENABLED=false to disable logging entirely.

---

How It Works

1. Weighted Selection: Providers chosen based on free tier limits
2. Health-Aware Fallback: If one fails or is cooling down, routing skips it and tries the next
3. Multi-Provider: search_web still supports multi-provider mode internally, while the CLI exposes that breadth as csearch web --all
4. Fetch Routing: fetch_webpage uses Parallel Extract first when PARALLEL_API_KEY is configured, then Jina Reader, then local fallback
5. Doctor Command: csearch doctor shows configured providers, recent telemetry, and current cooldown state

Default weights: Serper (42%) → Brave (33%) → Tavily (17%) → Perplexity (8%)

Customize via config.yaml:

providers:
  serper:
    weight: 50
    enabled: true
  brave:
    weight: 30
    enabled: true

---

Docker

docker build -t captain-search .

docker run -p 8000:8000 \
  -e SERPER_API_KEY=your_key \
  -e MCP_AUTH_TOKEN=your_secret \
  captain-search mcp --transport http --port 8000

---

Development

git clone https://github.com/mnm-matin/captain-search.git
cd captain-search
uv pip install -e ".[dev]"
pytest

Run the opt-in live provider audit when you want real endpoint checks plus a per-task configured-vs-working summary for web search, code search, and fetch:

RUN_E2E=1 uv run --extra dev python -m pytest -q tests/test_providers.py
LIVE_PROVIDER_CACHE_TTL_SECONDS=900 RUN_E2E=1 uv run --extra dev python -m pytest -q tests/test_providers.py
LIVE_PROVIDER_FORCE_REFRESH=1 RUN_E2E=1 uv run --extra dev python -m pytest -q tests/test_providers.py

The live audit reuses fresh results from .pytest_cache until the TTL expires and prints a session-end summary showing how many active providers are working for each task.

---

License

MIT License - see LICENSE

---

_{Built with ❤️ for the AI agent ecosystem}