For Claude and Claude Code users, this repo is interesting because it pushes on the part of the experience that usually gets ignored: startup speed, local control, and model flexibility. Agentty is trying to be a pair-programming agent that lives in your terminal without the usual Python/Node heaviness, while still supporting Claude, OpenAI-compatible endpoints, local Ollama models, and even Claude Pro/Max OAuth.
bwrap on Linux and sandbox-exec on macOS, and it restricts file tools to the workspace.read, write, edit, bash, grep, glob, git, web, search_docs, search_code, and task.What strikes me is that Agentty is not trying to win by being “another agent.” It’s trying to win by being fast, local, and less annoying. That matters more than people admit. A lot of AI coding tools feel like web apps wearing a terminal costume. If the startup is really that quick, and if the binary is truly self-contained, that’s a nicer default for Claude-style workflows than a stack that makes me wait before I can ask one small question.
I also think the retrieval angle is the most interesting part, and maybe the most overpromised. The README throws a lot of serious IR and GraphRAG terminology at you. Some of it sounds excellent on paper — hybrid search, source-tagging, local-only behavior, graceful fallback — and some of it feels like the kind of thing that can become complexity theater if it’s not tuned well. I’d be curious whether the average user actually feels the difference between this pipeline and a simpler, well-tuned search + rerank setup. That said, the offline behavior and deterministic stages are genuinely appealing. For codebases where you care about reproducibility and don’t want every lookup to depend on a cloud round trip, that’s a real advantage.
The sandboxing also feels like the right instinct. Terminal agents that can run commands are only useful if they don’t make you nervous. Default confinement plus workspace limits is the kind of boring security decision I like. It won’t eliminate risk, obviously, but it lowers the amount of trust required for everyday use.
If I were using Claude Code alongside something like this, I’d probably try it in the narrower, high-frequency spots first: local codebase Q&A, doc search, quick edits, and repeated project-specific conventions. The memory and “learns your codebase” pitch is attractive, but I’d want to see how cleanly it separates useful recall from stale junk. That’s usually where these systems get messy.
The big picture is simple: Agentty is betting that developer agents should feel like a native terminal tool, not a browser workflow with extra steps. I think that’s a worthwhile direction, and the local-first design gives it a sharper identity than most of the field.