How does
CCFT work?

One small program on your Mac runs a team of AI models, like a computer runs programs. It gives them work, keeps them apart, and writes down every step. Here is the whole machine, one picture at a time.

1One robot on your Mac. The models live in the cloud.

your Mac the robot (the daemon, ccftd) providers, in the cloud Claude GPT Kimi, GLM, MiniMax ... request response

No model runs on your Mac. Any model that you have a key or a plan for can join.

2The model is the CPU. The context is the RAM.

your files (on the disk) the context window (the RAM, packed again for each request) the model the CPU (forgets after each request) copies in only what this agent can see sends

The model remembers nothing. So the robot packs the context again for each request.

3One log. Every screen reads from it.

the robot the log (one encrypted file, never erased) writes every step desktop app terminal phone (later) reads

After a crash, the robot reads the log again. It never pays a model twice for the same step.

4Agents pass notes, not diaries.

the leader (big model) task board fix the login page write the tests review the change adds Claude GLM GPT each conversation stays locked

No agent reads the conversation of another agent. Only short, checked notes go across.

5Tools run in a locked box. Keys stay with the keeper.

an agent the locked box (sandbox) shell files MCP tools and skills MCP: one plug for every tool tool call keeper (key broker) internet adds the key on the way out

An agent cannot read your API keys, because it never has them.

6Talk to it. A fast talker, a big worker.

you Just talk. The ears hear when you stop. Hold a key to talk. Let go to send. ears (Deepgram, cloud) the talker (small, fast model) mouth (Deepgram, cloud) you hear the answer hands off work reports back when done the leader (big model, keeps working) only what you heard the thread

You can interrupt at any time. The work continues, and the thread keeps only the words that you heard.

7Named agents work on a schedule.

a timer (every 30 min) Scout AGENT.md MEMORY.md its own files wakes up only looks you tap to approve wants to act? asks first

When nobody watches, a named agent can only read. To act, it needs your tap.

8Plug in anything. Every plugin asks first.

the robot's expansion slots a model (a data file) tools (MCP) skills agent role goes in you approve once, tied to every file

If one file in a plugin changes, the robot asks you again.

?Predict first.

You ask one more question. What does the robot send to the model this time?

Pick an answer first.

Words on this page

the robot (daemon)
A program that runs in the background on your Mac, with no window. CCFT calls it ccftd.
model
The AI itself, for example Claude. It reads text, writes text and remembers nothing between requests.
provider
The company that runs a model on its computers, for example Anthropic.
request
One message from the robot to a provider. It asks the model for one response.
response
What comes back for one request. It arrives in small pieces.
context
Everything that the model reads in one request: the rules, your ask, the chat so far.
the log
One encrypted file where the robot writes every step. Nothing in it is ever changed or erased.
agent
One worker: a model with one job, its own conversation, and the tools that it can use.
sandbox
A locked box around a tool. The tool can only write inside its own folder.
cache
A copy of the start of your last request. Reading it again costs less.

Template for the /explain skill. Design from "How does CCFT work?" version 1 (1 Oct 2026). The example topic is CCFT, a small program that runs a team of AI models.