Twin Me Up 0.2.1

Fixed

• twin-me-up now runs commands as well, so twin-me-up doctor works exactly like twinmeup doctor. Without a command it still starts the MCP server, so agent hosts, uvx twin-me-up and the desktop extension launch it as before.

Upgrade: uv tool upgrade twin-me-up.

Twin Me Up 0.2.0

Agents can now ask you for permission and wait for your answer, safely, even through a crash.

Added

• Durable approvals. When twin_authorise says ask, an agent requests approval with twin_request_approval and waits on twin_approval_status. You decide in your own terminal with twinmeup approvals approve or decline, behind your passphrase. An approval grants exactly that action, once and for an hour unless you say otherwise, and every decision becomes learning material for your twin.
• Safe retries. Every write tool takes an optional request_id. A retry returns the first result instead of recording twice.
• Pi Durable extension (experimental). twinmeup extension export pi-durable gives durable agents the twin's tools, your brief, and enforcement that can ask you and wait through a restart. Tested live against Pi Durable 1.2.0.
• Claude Desktop bundle. Each release now attaches a one-click .mcpb desktop extension.
• Limited grants with twinmeup grant add --uses N, and pending approvals in twinmeup status.

Changed

• Store schema 2. Older stores migrate when opened.
• A new README section: Where it can still fail.

Install or upgrade: uv tool install --upgrade twin-me-up, or run uvx twin-me-up.

Full details in CHANGELOG.md.

Twin Me Up 0.1.0

A personal work twin for every AI agent. It learns how you decide, and agents act only within the authority you grant.

Highlights

• One twin, every agent. An MCP server, a native pi extension, a CLI bridge and function-calling schemas. Thirteen agent systems and frameworks documented, ten verified live, including Hermes, pi, Claude Code, Gemini CLI, Cursor, OpenCode, Microsoft Agent Framework, OpenAI Agents SDK and LangChain.
• Learning you ratify. Agents propose lessons from your decisions and corrections. Only you can confirm them, and your passphrase guards every human-only operation.
• Authority you grant. Scoped, optionally expiring grants. Anything else returns ask. Pre-action hooks for Hermes, Claude Code and pi enforce it outside the model and fail closed.
• Evidence, not vibes. Freeze held-out decisions and test the twin against a generic assistant and a rules-only baseline.
• Yours. Local SQLite with owner-only files, a tamper-evident audit log, portable export and erasure. No telemetry.

Install

uv tool install twin-me-up    # or: pipx install twin-me-up
twinmeup init
twinmeup connect --list

Or paste the one-prompt install into any agent. The research behind the design explains why the twin never infers authority.

See the changelog for everything in this release.