tunaFlow is still in beta and has some rough edges here and there.
I filtered as much as I could, but it's not easy to make it production-level haha
It seems like it will be useful primarily for people running 2 or more agents (including local LLMs).
Nothing groundbreaking — I just layered a workflow slightly on top of a multi-agent session.
For over half a year, I've been using Claude Code / Codex / Gemini CLI in parallel, and 3 months after getting a MacBook and setting it up (I definitely installed VS Code), suddenly I realized that I barely spent any time looking at the VS Code code editor screen while it was open. I was basically only using the editor as a "document viewer," and I was actually working continuously with agents in the terminal. I thought there would be quite a few people like this in the age of agentic coding. (And since Opus 4.6, AI is way better than me at coding-related tasks. There's nothing to fault. Abstraction? I just need to set the direction right. It does the work cleanly without any extra fluff.)
And running multiple agents in the terminal is really annoying. Opening 3 windows in iTerm2 / tmux, copy-pasting, tracking which agent I told to do what, fine-tuning the prompt slightly differently for each engine… copy-pasting. I was trying to do all this on one screen, but I think I ended up going in a different direction.
I originally made an app called tunaPi to lazily command terminal agents via messenger, but then I thought about making a dedicated messenger instead of a universal one (tunaDish), then wondered if I should also make a desktop app (tunaChat), and eventually caught the bug and made tunaFlow. I'm not sure if it will be useful to many people... but while making tunaFlow, I also made seCall, and I'm building and using several other tools that only I use. haha (I'm planning to use it for freelance development starting in May too.)
(Below is written by Claude 😄)
What Had to Be Different
"There are already plenty of simple CLI wrappers" was the starting point, and I set four things as differentiators.
1. Engine Parity
Claude / Codex / Gemini / OpenCode are all called with the same ContextPack standard.
There's no engine-specific glue code in the feature layer.
2. Branch-adopt
When you need to "try a different direction" in the middle of a conversation, you branch like a git branch.
Branched conversations live in a separate DB space called shadow conversation, and are
completely isolated from the main conversation. When the experiment is done, only the summary goes to the main.
It's natural to use like a git branch, so you become bolder about code modifications.
3. Roundtable
When you ask an agent to "review this code," it mostly says "looks good."
Whether it's something it wrote or something someone else wrote. Human team code review works
because the reviewer is a different person from the author. Apply the same structure to agents.
Run discussions with different engines, different personas, and blind mode if needed (without seeing
the previous participant's response). Two modes: Sequential / Deliberative (parallel + integrator).
4. CLI-first
I use Claude Pro / ChatGPT Plus / Gemini paid subscriptions as-is. tunaFlow doesn't
directly call the API, so there's no additional billing. Managing subprocess is a headache instead,
but from the subscriber's perspective, not having to pay twice is huge.
Stack
Tauri 2 · React 18 · TypeScript · Rust · SQLite. Local embeddings with ONNX.
Something Odd About It
I didn't type a single line of this app's code myself.
I only made architecture decisions, reviewed PRs, and gave direction — Claude Code did all the implementation.
It's a recursive experiment of "AI tools built by AI," and it's also a case where I became the
experimental subject myself for "how far can humans remain merely as direction-setters."
Status
Beta · macOS only (Windows/Linux to follow)
i18n English/Korean nearly complete (at session 41, A2-G merge complete)
PTY in-app terminal being reworked — using external terminal in parallel
Issues / feedback welcome
tunaFlow referenced ideas and code from several open-source projects. Thanks to each maintainer.
Bundled Sidecars (Deployed with the app)
rawq (MIT) — Code search sidecar. tunaFlow includes a local patched build bundled.
code-review-graph (MIT) — CRG sidecar (Full track). Graph-based code review analysis.
context-hub (MIT) — Context sharing sidecar. Auto-installed on first run.
Design / Architecture Influences
abtop (MIT) — Runtime observability / diagnostics for AI coding agents. Influenced the Trace panel and status bar design.
hermes-agent (MIT) — memory / toolset / iteration-budget patterns.
larksuite-cli (MIT) — CLI action layering / shared-rule / async-contract patterns.
chops (MIT) — ContextPack code-slice injection idea.
codex (Apache 2.0) — Reference implementation of CLI agent protocol.
xterm.js (MIT) — PTY panel terminal rendering.
react-markdown (MIT) — Chat markdown rendering.
D2Coding (OFL 1.1) — Bundled monospace font.
Tauri (MIT / Apache 2.0) — Desktop shell framework.
