Explicit permission
Scope requests to the task, surface consequences, and separate read access from writes and execution.
Personal computing · In development
AEDA
A local-first assistant focused on clear permissions, visible tool activity, adaptable model backends, and a safe path from suggestion to action.
AEDA · Approval flow
Approve code change?
Suggested after reading current workspace context
Action summary
Add XML docs and keep method privateSelected method only · targeted edit guard enabled
Review requiredDiff excerpt
+/// <summary> Validates the selected workspace
+private bool ValidateWorkspace(...)
−public bool ValidateWorkspace(...)
The direction
Assistance, not another attempt to replace everything.
Powerful CLIs, editors, coding agents, and automation systems already exist. AEDA should help ordinary users access them safely rather than pretending one small app can outperform every specialised tool.
The product focus is the Pill and Assist experience: context, permission, review, handoff, and a clear interface around tools that may otherwise be intimidating or opaque.
Product pillars
The assistant should show its work.
The core experience is designed around supervised action rather than silent automation.
Reviewable changes
Use proposals, dry runs, diffs, validation, and rollback before applying consequential edits.
Adaptable backends
Scale local models to available CPU and GPU hardware, with cloud fallback where appropriate.
Current engineering
Context and tools without invisible power.
AEDA gathers useful context from the active workspace, selected text, clipboard, windows, screenshots, and connected tools—but those inputs should remain visible and controllable.
- Permission broker for workspace, process, and tool actions.
- Selected-context proposals with targeted edit guards.
- Dry-run, apply, validation, and rollback states.
- Local model routing and hardware-aware configuration.
- Wrappers and integrations for mature coding and automation tools.
Roadmap
Make the safe path the easy path.
Development prioritises understandable assistance before expanding into more autonomous behavior.
Pill and Assist
Fast invocation, context selection, visible tool requests, and supervised action.
Backend scaling
Adapt model selection to low-end CPU systems, mid-range GPUs, and high-end local hardware.
Tool ecosystem
Integrate reliable external systems while preserving one understandable permission model.
Contact
Science, software, or something between them?
Reach out about biotechnology, laboratory tools, local-first AI, research software, or a useful collaboration.
