Docs
Deep dives into how Claudin saves tokens, keeps sessions fast, and keeps you in the loop.
Overview
This section covers the features that run underneath every Claudin session, plus how to extend and automate it. Most of the performance features exist for one reason: LLM context is expensive, and a chatty terminal is the fastest way to burn through it.
This section is organized in four groups: performance features that cut the tokens a session spends on noise, agents & workflows for delegating and orchestrating work, ways to extend Claudin with skills, plugins, and MCP servers, and automation for running things without you watching. Notifications, at the end, is about staying in the loop when a long session needs you.
Performance
Every one of these ships on by default and requires no setup. Together they're the difference between a session that runs out of context at hour one and one that's still going at hour three.
Strips progress bars, download lines, and passing test noise from 60+ CLI tools before it reaches the model — errors and warnings always pass through untouched.
24% cheaperA clip-frontier invariant and per-provider cache profile keep prompt-cache hits alive across history management, instead of silently breaking the cached prefix.
~25k tokens/readLarge files come back as a structural skeleton — every function and class signature with its line range — then unfold one symbol on demand instead of paying for the whole file.
~20k tokens/queryMulti-page web research runs in an isolated subagent context on the cheapest model, so the results come back — not the browsing history.
Agents & workflows
Delegate work to a subagent with its own context, tools, and model — built in (Explore, Plan, WebResearcher...) or defined as a markdown file in .claudin/agents/. Chain several of them into a phase state machine with agent workflows.
Fork with full context, run in the background, or isolate in a git worktree — plus custom agents that mix models across providers.
local or GitHub-triggeredA phase state machine that runs agents automatically — from /workflows, or triggered by a labeled GitHub issue with claudin workflow watch.
Extensibility
Add capabilities Claudin doesn't ship with: on-demand instructions, shareable bundles of agents and hooks, or a live connection to an external tool.
SKILL.md files invoked with /name or automatically by the model — plus the full list of bundled skills.
Shareable bundles of agents, skills, hooks, and MCP servers, namespaced so they never collide with your own.
project, user, or local scopeConnect external tools and data as MCP servers — .mcp.json, OAuth, and the claudin mcp CLI.
Automation
Run Claudin without someone watching: deterministic hooks on lifecycle events, plus scheduled prompts, stopping conditions, headless runs, and deep links.
Shell commands, prompts, or HTTP calls that fire deterministically — not left to the model's discretion.
cron · goals · headless · deep linksScheduled prompts with /loop, stopping conditions with /goal, -p for scripts and CI, and claude-cli:// links.
Notifications
Not every feature in this section is about tokens. When a session needs your attention — a permission prompt, a finished long-running task — Claudin can tell you outside the terminal.
iTerm2 · Kitty · Ghostty · OS toastsAuto-detects the best channel for your terminal and sends a "needs attention" alert the moment Claudin is blocked on you.
Next steps
New to Claudin? Start with install, then wire up a provider:
- Install — get running in under a minute.
- All providers — Anthropic, OpenAI, Gemini, Ollama, and 200+ more.
- Changelog — what shipped, release by release.