AIDA — AI Data Assistant

Status: 1.0.0. In daily use at APS beamlines 12-ID/9-ID. All ten phases of the original plan are delivered, the test suite is 2,449 tests green on macOS, Windows and Linux, and the config formats and CLI commands are stable: a breaking change to either now means a new major version. Every change is dated in CHANGELOG.md. Bug reports and rough edges are very much welcome — open an issue.

AIDA is a local scientific agent workbench: a simple, reliable desktop GUI (and CLI) for using AI agents in scientific work — conversation with local or cloud LLMs, correct use of domain MCP servers (pyIrena, bait_mcp, ...), correct display of rich tool results (especially PNG plots), reading and producing documents, and controlled access to your own data folders.

It is built for pyIrena and USAXS-instrument users and is deliberately not a general-purpose AI platform — see planning/DESIGN.md §1 for the rationale.

What it does

• Chat with any model you can reach — Ollama / LM Studio / any OpenAI-compatible endpoint, OpenAI itself, Claude direct, or Claude through the ANL Argo proxy. Named profiles, switchable mid-conversation.
• Use MCP servers properly — a PNG a tool returns is decoded and shown as an image, not flattened to text. Servers are grouped so a small local model isn't drowned in 100+ tool schemas.
• Work in your folders — named workspaces bundle source folders, a target folder, a provider profile, an MCP group, and skills. The agent reads, writes, searches, and (optionally) runs scripts there under a safety model you configure.
• Produce documents — Markdown in Obsidian layout (images in a sidecar folder, linked relatively) or DOCX, written into your target folder.
• Search your own documentation — optional local RAG over folders you choose, with local or cloud embeddings.
• Keep the documents you attach — an attached paper is copied into the conversation's own folder with the text that was extracted from it, and deleted again with the conversation. Ask about a figure and the agent fetches that one by label rather than being handed every image in the file. Optional Mistral OCR makes those labels reliable on multi-column journal PDFs — off by default, per workspace, and it asks before a document leaves your machine.
• Keep conversations sorted — label them by person on a shared beamline machine, or by project on a laptop, and filter the list by it. Labels are organization, not security.
• Stay out of black boxes — token counts and cost estimates per session, a tool-call log, a raw MCP result inspector, and aida doctor.

Install

From PyPI (recommended):

pip install "aida-workbench[gui,docs]"
aida doctor
aida-gui

The PyPI distribution name is aida-workbench; the import package and the console scripts are aida / aida-gui. Extras: gui (PySide6 desktop app), docs (PDF/DOCX/XLSX/PPTX reading, image handling), ocr (the optional Mistral OCR backend for figure extraction).

Already using pyIrena? Install both — in either order, in one environment or two — and wire up its MCP tools with a single command:

pip install "aida-workbench[gui,docs]" "pyirena[all]"
aida mcp add-pyirena

See docs/pyirena.md for the compatibility details and the GUI equivalent.

From a git checkout, for development:

git clone https://github.com/jilavsky/aida.git
cd aida
conda env create -f environment.yml   # or: pip install -e ".[dev,gui,docs]"
conda activate aida
aida doctor

environment.yml already includes the gui, docs and ocr extras. If aida doctor or aida-gui says PySide6 isn't installed on a headless Linux machine even though the install above reported no errors, see GUI fails to import on headless Linux — it's usually a missing system Qt library, not a missing Python package.

First run

1. aida doctor — confirms Python, config files, keychain, and folders.
2. Launch aida-gui. On a fresh install it offers to set up a provider profile and a first workspace; you can also do both from the Configure menu (Providers…, Workspaces…) or from the CLI (aida workspace new).
3. Pick a workspace and profile in the toolbar and start typing.

docs/installation.md → docs/providers-and-secrets.md → docs/workspaces.md is the full path from nothing to a working session.

Documentation

Task-oriented setup and configuration guides — providers, workspaces, the safety model, MCP servers, scripting, RAG, a GUI tour — live in docs/. Fully-commented example config files are in examples/config/.

Status and roadmap

All ten phases are implemented: config and diagnostics, the provider layer and agent loop, MCP with typed artifacts, persistence and workspaces, the PySide6 GUI, documents and the safety model, MCP management UI, RAG, coding/scripting, headless automation (aida run, stored workflows, the in-app scheduler — see docs/workflows.md), and moving a setup between machines (docs/moving-and-sharing.md).

What is still open — manual acceptance checks that need a real beamline, a conda-forge decision, full backup/restore — is in TODO.md. PLAN.md holds the reasoning behind each item, and its §2 the things considered and deliberately not committed. What has been delivered, and why, is in planning/COMPLETED.md; the per-phase checklists are archived in planning/completed/.

Released and unreleased changes, version by version, are in CHANGELOG.md.

Requirements

Python >= 3.11. Tested on macOS, Windows, and Linux.

License

MIT — see LICENSE.