yoker-assistant

A personal assistant that communicates by email, built on yoker-as-SDK.

This is a Yoker showcase package. It demonstrates the yoker-as-SDK mode: Python owns the process and calls Yoker as a library for the reasoning half. The package is also a Yoker plugin provider (dual-mode): it exposes its own tools via __YOKER_MANIFEST__ for any Yoker consumer to load.

Full documentation is published at yoker-assistant.readthedocs.io — including a tutorial that tells the build story end-to-end, a quickstart with a worked example, and a per-page reference for architecture, security, and configuration.

Status

Initial release candidate. The loop, the agent seam, the mailbox seam, the bounded tool set, the custom md→html tool, the tests, and SECURITY.md are all in place. All quality gates green. See TODO.md for the remaining backlog (HTML styling polish, attachment handling, batch processing, make run-demo).

What It Does

The assistant is a personal assistant that communicates by email. You email it; it reasons about the email, acts on your behalf using a curated set of safe tools, and emails back. The inbox is the entire UI — no TUI, no CLI prompts, no web.

How it works

1. Python owns the email loop. It polls an IMAP inbox for UNSEEN messages, fetches each one, and hands it to the agent as a simple From/Subject/Date/body payload.
2. The agent reasons. It categorizes the email, decides what actions to take, uses its bounded tools (read, write, web search, git, etc.), and composes a reply in markdown.
3. The agent converts the reply to HTML via a custom md_to_html tool (the showcase's "create your own bounded tool" example).
4. Python sends the reply. It emails the HTML verbatim, marks the original read, and archives it.

The bootstrap flow

On first run, PERSONAL.md does not exist. The agent detects this and replies with a welcome message and questions about you (name, preferred address, project context, tone, goals). You answer by replying over email. The agent iterates with you until it has enough to write the initial PERSONAL.md — and optionally commits and pushes it via yoker:git. After bootstrap, every subsequent email is the next message in the same session.

The git demo beat

The visible "acts on behalf of the owner" moment: the agent autonomously maintains its own PERSONAL.md learned-behaviours file in version control. When you email a preference ("from now on, always CC me on deploy-related replies"), the agent updates PERSONAL.md, commits, and pushes — all via bounded git tools, not a shell.

Quick Start

make env-dev                              # install all dependencies
make test                                 # run the test suite
python -m yoker_assistant --once          # one poll iteration, then exit

--once is the demo/test mode: one poll iteration and exit. Drop --once for the long-running mode (polls every 60 seconds until SIGINT/SIGTERM).

A Yoker backend is a prerequisite — either a local Ollama install or a cloud LLM provider API key. If you do not already have one, run uv run yoker init once to write ~/.yoker.toml with a backend of your choice.

See Installation for the full setup guide, including email account configuration. Plugin registration is automatic — the package self-injects it at startup.

Architecture

This package demonstrates yoker-as-SDK: Python owns the email loop (poll, fetch, reply, archive) and calls Yoker as a library for the reasoning half. The package is also a Yoker plugin provider (dual-mode): it exposes its own md_to_html tool via __YOKER_MANIFEST__ for any Yoker consumer to load — including itself.

The sister project yoker-writing-assistant demonstrates the complementary mode — yoker-as-runtime — where Yoker is the entry point and the package runs under it as a plugin.

Both projects share a common quality bar documented in STANDARDS.md.

Documentation

Full documentation lives in docs/ and is published to ReadTheDocs:

https://yoker-assistant.readthedocs.io

The Tutorial tells the build story end-to-end — why this package exists, the two halves (Python loop vs agent reasoning), the seams (Yoker SDK seam + simple-email-gw seam), the handoff contract (payload format + four-way branch), the bounded tool set and the safety model, the persistent-session architecture, the custom md→html tool story, the dual-mode architecture, the git commit/push demo beat, recipient safety, and what is out of scope.

Supporting pages: Installation, Quickstart, Architecture, Security, Configuration, API, Changelog.

The AGENTS.md file provides the project guide for agents working on this codebase. PACKAGE.md provides AI-optimized package documentation for consumers.

License

MIT