Decide between points and cash for every trip
AI-powered travel hacking with points, miles, and award flights. Task: AI-powered travel hacking with points, miles, and award flights. Drop-in skills and MCP servers for OpenCode, Claude Code, and Codex. Ask your AI to find you a 60,000-mile business class flight to Tokyo. It'll search award availability across 27 mileage programs, compare against cash prices, check your loyalty balances, and tell you the best play. Quick Start Install via plugin (no clone) The easiest way to use the toolkit. Pick your tool: Inside Claude Code, run: Done. 42 skills, 6 MCP servers (5 free + LiteAPI which needs a key), and the travel-hacker subagent are installed. Run claude from anywhere and ask it to plan a trip. To verify the install, run /travel-hacker:getting-started inside Claude Code. It tells you which API keys are configured and points at the local setup script for the missing ones. Or in your shell: claude plugin list | grep travel-hacker should show the plugin. In your terminal: Then start Codex and the plugin appears in /plugins. Codex pulls the marketplace catalog from the repo (.agents/plugins/marketplace.json) and the plugin manifest from plugins/travel-hacking-toolkit/.codex-plugin/plugin.json, so the install is one command. Cowork itself doesn't expose /plugin slash commands, so you can't install plugins from inside a Cowork chat. Install via Claude Code first; Cowork picks it up automatically because both share ~/.claude/plugins/. In your terminal, run claude In Claude Code, run the two install commands from the Claude Code section above Open Cowork inside Claude Desktop. The plugin is already there. Claude Desktop's chat UI doesn't support the plugin format yet, so use Cowork inside the same app for the full experience. Configure API keys The 5 free MCP servers (Skiplagged, Kiwi, Trivago, Ferryhopper, Airbnb) work immediately with zero keys. Cash flight and hotel search work out of the box. To unlock award search and the rest of the toolkit, you need API keys set as environment variables in your shell. Run the local setup script: macOS / Linux / WSL / Git Bash: Windows (PowerShell): The script prompts for each key with masked input, validates them (rejects values with single quotes that would break the export, plus a per-key minimum length sanity check), writes them to the right shell rc with a backup, and never echoes the values back. The exports are single-quoted so other shell metacharacters in the key are safe. Keys never enter your terminal scrollback or any chat session. The /travel-hacker:getting-started skill inside Claude Code points at the same script. The 4 highest-value keys: Other keys (SerpAPI, RapidAPI, LiteAPI, TripAdvisor, RESROBOT, Rejseplanen, Entur) extend specific skills. The setup skill covers them too. See the full API key reference for signup links. Five Docker-based skills (Southwest, American Airlines, Chase, Amex, TicketsAtWork) handle sites that don't have public APIs, plus a shared base image. They auto-pull on first use. See the Docker Images section for credentials and configuration. Add to your shell rc (~/.zshrc, ~/.bashrc, etc.): Reload (source ~/.zshrc) or open a new terminal, then run claude. If you keep secrets in 1Password, you can resolve at launch instead: The --no-masking flag is required because op run's default output masking interferes with Claude Code's TTY detection and triggers a non-interactive prompt error. Masking only affects stdout, not the env vars themselves, so the keys still get to claude correctly. (See 1Password CLI docs for the .env template syntax.) The toolkit reads keys from the shell environment because plugin userConfig doesn't propagate to skills that shell out to scripts (verified empirically; related upstream issues anthropics/claude-code#11927, #39125). Clone path (for OpenCode users or contributors) OpenCode doesn't have a plugin format, so OpenCode users clone the repo. Contributors who want to hack on the toolkit also clone. The setup script picks your tool (OpenCode, Claude Code, Codex, or all three), creates your API key config file, installs dependencies, and optionally installs skills system-wide for OpenCode and Claude Code. On Windows the .cmd wrapper launches scripts\setup.ps1 with an ExecutionPolicy bypass so nothing needs to be unblocked first. Then launch your tool from inside the repo: claude --plugin-dir . loads the plugin (skills, MCP servers, and the travel-hacker subagent) directly from your clone. Changes you make to the toolkit are picked up immediately. Useful when you're contributing.; tools: Claude Code. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include installing Claude Code and configuring API keys for award search (e.g., Seats.aero Pro, AwardWallet Business). Free MCP servers only provide cached data for the next 60 days; live award search requires paid keys. Integration uncertainty exists regarding Docker-based skills for sites without public APIs, which require specific credentials and browser automation setup..
Tools
Published on Agent Work Atlas:
Overview
AI-powered travel hacking with points, miles, and award flights.
Workflow
AI-powered travel hacking with points, miles, and award flights. Drop-in skills and MCP servers for OpenCode, Claude Code, and Codex. Ask your AI to find you a 60,000-mile business class flight to Tokyo. It'll search award availability across 27 mileage programs, compare against cash prices, check your loyalty balances, and tell you the best play. Quick Start Install via plugin (no clone) The easiest way to use the toolkit. Pick your tool: Inside Claude Code, run: Done. 42 skills, 6 MCP servers (5 free + LiteAPI which needs a key), and the travel-hacker subagent are installed. Run claude from anywhere and ask it to plan a trip. To verify the install, run /travel-hacker:getting-started inside Claude Code. It tells you which API keys are configured and points at the local setup script for the missing ones. Or in your shell: claude plugin list | grep travel-hacker should show the plugin. In your terminal: Then start Codex and the plugin appears in /plugins. Codex pulls the marketplace catalog from the repo (.agents/plugins/marketplace.json) and the plugin manifest from plugins/travel-hacking-toolkit/.codex-plugin/plugin.json, so the install is one command. Cowork itself doesn't expose /plugin slash commands, so you can't install plugins from inside a Cowork chat. Install via Claude Code first; Cowork picks it up automatically because both share ~/.claude/plugins/. In your terminal, run claude In Claude Code, run the two install commands from the Claude Code section above Open Cowork inside Claude Desktop. The plugin is already there. Claude Desktop's chat UI doesn't support the plugin format yet, so use Cowork inside the same app for the full experience. Configure API keys The 5 free MCP servers (Skiplagged, Kiwi, Trivago, Ferryhopper, Airbnb) work immediately with zero keys. Cash flight and hotel search work out of the box. To unlock award search and the rest of the toolkit, you need API keys set as environment variables in your shell. Run the local setup script: macOS / Linux / WSL / Git Bash: Windows (PowerShell): The script prompts for each key with masked input, validates them (rejects values with single quotes that would break the export, plus a per-key minimum length sanity check), writes them to the right shell rc with a backup, and never echoes the values back. The exports are single-quoted so other shell metacharacters in the key are safe. Keys never enter your terminal scrollback or any chat session. The /travel-hacker:getting-started skill inside Claude Code points at the same script. The 4 highest-value keys: Other keys (SerpAPI, RapidAPI, LiteAPI, TripAdvisor, RESROBOT, Rejseplanen, Entur) extend specific skills. The setup skill covers them too. See the full API key reference for signup links. Five Docker-based skills (Southwest, American Airlines, Chase, Amex, TicketsAtWork) handle sites that don't have public APIs, plus a shared base image. They auto-pull on first use. See the Docker Images section for credentials and configuration. Add to your shell rc (~/.zshrc, ~/.bashrc, etc.): Reload (source ~/.zshrc) or open a new terminal, then run claude. If you keep secrets in 1Password, you can resolve at launch instead: The --no-masking flag is required because op run's default output masking interferes with Claude Code's TTY detection and triggers a non-interactive prompt error. Masking only affects stdout, not the env vars themselves, so the keys still get to claude correctly. (See 1Password CLI docs for the .env template syntax.) The toolkit reads keys from the shell environment because plugin userConfig doesn't propagate to skills that shell out to scripts (verified empirically; related upstream issues anthropics/claude-code#11927, #39125). Clone path (for OpenCode users or contributors) OpenCode doesn't have a plugin format, so OpenCode users clone the repo. Contributors who want to hack on the toolkit also clone. The setup script picks your tool (OpenCode, Claude Code, Codex, or all three), creates your API key config file, installs dependencies, and optionally installs skills system-wide for OpenCode and Claude Code. On Windows the .cmd wrapper launches scripts\setup.ps1 with an ExecutionPolicy bypass so nothing needs to be unblocked first. Then launch your tool from inside the repo: claude --plugin-dir . loads the plugin (skills, MCP servers, and the travel-hacker subagent) directly from your clone. Changes you make to the toolkit are picked up immediately. Useful when you're contributing.
Reported result
The source does not report measured business outcomes or success rates for this workflow. It describes the capability to find optimal travel plays but provides no data on actual savings or booking success. A measurable acceptance check would be to verify that the AI correctly identifies a lower-cost option (points vs. cash) for a specific test route compared to manual search results.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include installing Claude Code and configuring API keys for award search (e.g., Seats.aero Pro, AwardWallet Business). Free MCP servers only provide cached data for the next 60 days; live award search requires paid keys. Integration uncertainty exists regarding Docker-based skills for sites without public APIs, which require specific credentials and browser automation setup.
Try it
Act as a travel hacking assistant using the travel-hacking-toolkit in Claude Code. My goal is to decide between points and cash for a trip from [Origin City] to [Destination City] on [Date]. 1. Check if the travel-hacker plugin is installed. If not, guide me through the installation steps for Claude Code. 2. Verify which API keys are configured by running /travel-hacker:getting-started. If key award search APIs (like Seats.aero) are missing, list the required keys and stop for my input. 3. Once configured, search for cash flight prices using available free MCP servers (Skiplagged, Kiwi) and any configured paid APIs (Duffel, Ignav). 4. Search for award availability across mileage programs using Seats.aero if the key is present. If not, use cached data for the next 60 days and note the limitation. 5. Compare the cash price against the points cost, calculating the cents-per-point (CPP) value. 6. Present a summary table showing: Cash Price, Points Cost, CPP Value, and Recommended Action (Pay Cash or Use Points). 7. Do not attempt to book. Wait for my confirmation before proceeding to any further steps. If any API key is missing or a tool fails, report the error and suggest the fallback skill (e.g., fallback-and-resilience) without proceeding.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.