say what needs to happen once. use the same language you would use with a person.
an agent that runs your conversations.
mimicry lives inside a real telegram client. it handles the work in the background. nothing reaches another person until you approve it.
- ingress
- excluded chats never enter the system
- brains
- reasoning and memory stay on your machine
- egress
- every outgoing message waits for your tap
give it the outcome.
mimicry finds the people, opens only the context it needs, and carries the task between steps.
the work can move. the final message cannot. every send waits at a boundary you control.
the work stays visible.
0.4 makes the invisible parts inspectable: what enters the model, who owns the running job, and exactly where action stops.
one agent. selected context. one explicit boundary between thinking and acting.
see the budget
preview what enters the model and how much room remains before the request starts.
leave. come back.
jobs stay owned by the local server while the client changes accounts or views.
you remain the edge
the agent can prepare the work. the outgoing path remains blocked until you approve it.
trust is not the feature.
verification is. excluded chats are never collected. the brains run on your machine. prompts go directly to the AI provider you choose, so that provider sees what you send it. mimicry's cloud holds keys and billing. it never receives your chats.
- ingress
- read the open filter
- brains
- watch the provider-direct path
- egress
- verify the approval gate
follow the build.
updates, implementation notes, and the release channel are part of the product surface—not a footer you have to hunt for.