Understanding EasyCC
The ideas behind the app — agents, tools, permissions, skills, version history. You don't need this to use EasyCC; read what's interesting.
Short plain-English explanations of how EasyCC works under the hood. Read what sparks your curiosity. For safety in particular, see the separate Safety page.
Agents and specialists
An agent is one folder EasyCC manages — usually one project. "Marketing", "Bakery business plan", "Med school notes". Each agent has its own files, chats, version history, settings, and installed skills.
A specialist is a focused helper inside an agent — "accessibility-reviewer", "financial-analyst", "critic". Specialists don't have their own files or chats. They're roles Claude takes on within the agent's chats.
When to make which:
| Want | Use |
|---|---|
| A new kind of work with its own files | A new agent |
| Apply a recurring lens to existing work | A specialist |
| Different settings — model, permission mode, skill set | A new agent |
| A "second opinion" review voice | A specialist |
On disk: agents are work folders you pick on the filesystem; specialists are markdown files at <agent>/.claude/agents/<name>.md.
How Claude uses tools
Claude operates in two modes during a chat: talking (generating text in the response) and doing (using a tool to read or change something). Usually both, in sequence.
The always-available tools:
- Read, Glob, Grep — view files and search across them.
- Edit, Write — change or create files.
- Bash — run shell commands.
- WebFetch, WebSearch — pull from the web.
Plus whatever your enabled MCP servers expose — GitHub, Linear, Gmail, and so on.
In the chat you see a tool card for each call: tool name, summary, status indicator. Click to expand for full content (file body, diff, command output). Different visual treatments tell you what kind of tool:
- Read tools are quiet — reading is harmless and never asks permission.
- Write tools are prominent and trigger a permission card (unless you're in Auto-accept or Plan mode).
- Commands look like terminal blocks.
- Web fetches show the URL.
- MCP calls show the connector's icon.
Why tools at all? Claude doesn't know what's in your files. It has to read them through the Read tool to answer questions. Same for searching, web fetches, and everything else.
If a tool fails, the card turns red with the error. Claude usually retries with a tweak, or you can suggest one.
Permissions, in plain English
The permission model: reads are free; anything that changes something asks first. That way you catch mistakes before they land.
What asks permission:
- Edit, create, or delete a file
- Run a shell command
- Fetch a URL
- Call an MCP server (first time)
What doesn't:
- Read, list, search
- Auto-save, version history
- Send a chat message
When EasyCC asks, you get three options:
| Choice | Effect |
|---|---|
| Allow once | This action, now. Future similar actions ask again. |
| Always allow | Don't ask for this pattern in this agent. Creates a rule under Settings → Permissions. |
| Deny | Block the action; Claude continues with what it can. |
Patterns range from exact (notes/2026-04-19.md) to broad (*). More specific rules beat less specific ones. You can edit auto-generated rules anytime.
Modes control how often the asks happen:
- Ask (default) — every change asks.
- Auto-accept edits — file changes go through; commands, fetches, and bulk deletes still ask.
- Plan mode — Claude lays out a plan and waits for one approval; then executes the whole thing.
Your IT admin may have pre-pushed rules (small lock icons in Settings → Permissions). If they're making your work harder, talk to IT.
What skills are
A skill is a saved prompt — pre-written instructions Claude follows when you click a button or type a slash command. Skills save typing, keep outputs consistent, and make functionality discoverable in the Skills panel.
Three scopes:
| Scope | Where | Visible in |
|---|---|---|
| EasyCC bundled | Inside the app binary | Library tab |
| Per-agent | <agent>/.claude/skills/<name>/SKILL.md | This agent only |
| User-scope | ~/.claude/skills/<name>/SKILL.md | Every agent |
A skill file is markdown with YAML frontmatter — name, description, optional category, tags, allowed-tools, and an optional model override (a fast-task skill might pin Haiku regardless of your default).
Run a skill either by clicking it in the Skills panel or typing /<command> in chat. Sharing is just sending someone the SKILL.md file; they drop it into their own .claude/skills/.
For day-to-day how-to, see Skills and the starter library.
Auto-save and checkpoints
EasyCC's version-history model deliberately avoids Git concepts. Two layers handle different cases:
- Auto-save records a version every time you stop typing for 2 seconds (per file). Content-addressed and deduplicated, so identical saves cost no storage.
- Checkpoints name a snapshot of your whole agent at one moment, for risky changes or labelled known-good states.
For day-to-day use see Version history. The mental model: auto-save handles "undo a few minutes"; checkpoints handle "go back to before I started this refactor".
Compared to Word or Google Docs: per-file rather than whole-document, on disk rather than in the cloud, and supporting multi-file rollbacks via checkpoints. Typical storage is 5–20 MB per agent per month.
Ctrl+S / Cmd+S forces an immediate save. You don't have to — auto-save handles it — but the option's there.
How is this guide?