Working on a new agentic editor and looking for collaborators, specially around want we want this to do.
Why a new agentic editor? My initial list of gripes with existing ones that I am familiar with or have looked into below.
Although this will, eventually, do what I want I am hoping to actually make it useful for others too; getting feedback from others on what they need which existing editors don’t do will greatly increase the chances of this been useful.
Last note before sharing list of gripes with existing editors… This will NOT be quick. I am working slowly, and methodically, on this. I am thinking on doing very small pieces at a time so I can truly understand the code that comes back from whatever I use to help me write this (likely Claude Code and OpenCode). This will be a side project with zero prospects of any sort of profit so I am spending whatever little time I can every day on this, but the emphasis is on getting it right not shipping something quickly.
Gripes with existing editors…
General notes
- Using an editor locked to a single company, and family of models, means your work stops when the provider has a problem, or the provider arbitrarily decides to lock you out [1].
- Most, perhaps all, agentic editors have git as first class citizen and anything else you have to nudge it; I want to be able to work with any version control software(VCS) of my choice.
- Editors allow to approve some tools, but then do operations like command1 && command2 which appears as not authorized even if you had approved command1 and command2.
Specific editor notes
These are issues I have either experienced or expect would have if I was to use the editor. My experience is primarily Claude Code and OpenCode; have used others (Warp, Windsurf,etc), but my daily drivers are Claude Code and OpenCode.
Claude Code
- Anthropic has moved to close subscriptions in all third party tools. This is an indication on their, strong, desire to lock people into their ecosystem so because of this I try to avoid, or minimize Claude Code only workflows.
- Editor keeps executing non approved commands even though commands with same functionality which have been approved already exists.Their Automode (approve anything they deem is not dangerous) is less secure than their old “accept edits” mode
- Single family of models - Claude
Codex CLI - have not used
- Single family of models - GPT
Github copilot
- Approvals only last current session.
- Approvals do not seem to be tied to a given folder so when we “approve all file operations during this session” it seems we may be authorizing work outside current folder too.
- It doesn’t show plans; it shows "plan saved to ~/.copilot/session-state/id/plan.md. One has to, manually go look at the plan
OpenCode
- Approvals only last current session.
PI - have not used so going on what I have read
https://pi.dev/ You Only Live Once(YOLO) default mode is too permissive. Would require a separate, isolated, development location. Having to setup and maintain an entire environment because some software lack of security, was not something I wanted to consider.
Notes
You must log in or # to comment.