notion-cli
TypeScript12
A command-line client for Notion with JSON output on every command, so an agent can read, write and search a workspace without a browser.
npm install -g @coastal-programs/notion-cliView on GitHubAn agent can only think about what fits in its context window. Every MCP server you connect fills part of that window before the conversation starts. A command-line tool takes none of it until the moment it runs.
Taken up by five common MCP servers (58 tools) before the conversation even starts.
Anthropic, advanced tool use(opens in a new tab)The point where Claude's ability to pick the right tool starts to degrade.
Claude docs, tool search(opens in a new tab)Got less reliable as their input grew, even on simple retrieval tasks.
Chroma, context rot study(opens in a new tab)Tokens to answer the same question about a GitHub repo: the gh CLI against GitHub's MCP server.
Scalekit, MCP vs CLI benchmark(opens in a new tab)The agent reads the name, description and full schema of every tool on the server, up front, whether this job needs them or not. Add a few servers and a big slice of the context window is gone before you type a word.
All of that is weighed on every turn. Dozens of similar-looking tools means more chances to call the wrong one, invent a parameter, or confidently report something that never happened.
A server to start, a transport to debug, a session to keep alive. When it breaks you are guessing at what the agent actually saw.
The agent already has a terminal. The tool adds nothing to the context until it is called, and when the agent needs the details it runs --help on that one command and reads a few hundred tokens.
Output pipes through jq or grep before the model reads it. Fewer tokens on every turn means quicker responses, lower cost and more room for the actual problem.
One binary, no server. When something breaks I run the same command in my own terminal and get the same output, so it gets fixed fast.
MCP still has its place: hosted tools where many people sign in with their own accounts, or apps with no terminal. For an agent working in a shell, a well-built CLI does the same job with less noise. The numbers, the sources and the trade-offs are in the full write-up.
Free to install, read and fork. Xero and Outlook follow once they are cleaned up.
A command-line client for Notion with JSON output on every command, so an agent can read, write and search a workspace without a browser.
npm install -g @coastal-programs/notion-cliView on GitHubA command-line client for Inngest events, runs and functions, built so an agent can inspect and replay background jobs from the terminal.
npm install -g @coastal-programs/inggest-cliView on GitHubAn agent cannot ask a terminal a clarifying question, so the tools have to be predictable in the ways that matter.
Every command can return a predictable { ok, data, error } envelope, so an agent parses the result instead of scraping text.
Zero for success, distinct non-zero codes for auth, validation, not-found and rate-limit errors, so a caller can branch without reading the message.
Login opens a browser once, tokens live in the operating system keychain, and no secret is ever written to a file an agent could leak.
The things an agent does repeatedly (find a page and append to it, replay a failed run) are single commands, not five calls it has to chain.
A 30 minute call. You describe the spreadsheets, the apps and the paperwork; I tell you what I would build and roughly what it would cost.