After organizing minor issues, we plan to release it as open source through GitHub soon. While the documentation is verbose, the core concept is "Harness + Workflow". These are not just my own amateur ideas, but rather an integration of mature features, algorithms, and ideas from several well-known open source projects. For alpha testing, I created seCall (introduced below) from a blank repository in half a day and fixed initial minor bugs (those curious about seCall's quality can clone it and have it analyzed by agents at the level of Opus or Claude—code Codex), and the gemento mentioned in the video — I'm aiming to develop Gemini e4b into something practical while validating my hypotheses . Of course, while building it, I plan to test tunaFlow's core features as well.
tunapi — A backend app that allows chatting with terminal agents through multiple general-purpose messengers (Telegram, Mattermost, Slack, Discord). I created it by 'forking' a Telegram-exclusive transporter called takopi
tunadish — A desktop app positioned as a plugin, serving as the transporter upper layer of tunapi. Since we had a backend, let's build a messenger too! It was an ambitious project, but it's still in alpha version.
seCall — A project that parses conversations with terminal agents, stores them in Obsidian and a database, then makes them available as an MCP plugin to terminal agents so you can search old conversations and use them in current context. It was created to test tunaFlow.
And now introducing tunaFlow — Agents work better when they're comfortable! You can create apps or services with just ideas and tokens (but not a click). You need to discuss a lot with the architect and review the plan carefully for good quality results. In reality, small MVPs can be made quickly. When you take user domain knowledge + agents wrapped in harness and deploy them in workflows with validated results, you get genuinely high-quality deliverables)
▎ "Of the agent, By the agent, For the agent - just feed idea and tokens"
tunaFlow is an AI agent orchestration client. It's not "a chat app convenient for users," but rather focused on enabling agents to work under optimal conditions. Users decide domain knowledge and direction, agents deliberate and review those decisions together, and execute with user approval.
When using Claude Code, Gemini CLI, and Codex individually, recurring problems emerge:
- Having to explain the same context from scratch every time
- Agents can't remember what they learned from previous conversations
- It's cumbersome to pass one agent's work results to another agent
- Code review has to be done manually by people
tunaFlow solves these problems structurally.
Key Features
1. Multi-Engine Integration:
Run Claude, Gemini, Codex, OpenCode, and Ollama in a single app. Context is maintained even when switching engines mid-conversation. You can manage engine + model + persona as presets through Agent Profiles.
2. Roundtable — Agents debate each other
Multiple agents discuss a single topic sequentially or deliberatively. Claude designs, Gemini rebuts, Codex suggests alternatives. Each participant's identity is injected into the prompt to avoid role confusion.
3. Workflow Automation — 3-Role Pipeline
Architect → Developer → Reviewer roles auto-cycle based on Plan.
Plan design → Approval → Implementation Branch → Developer auto-invoked → Review RT (2-agent debate) → Pass/Fail/Rework → Completion
Auto-detection by markers, and escalation to Architect redesign after 3 failures. Patterns learned from failures are automatically injected into the next Rework.
4. ContextPack — Context optimization for agents
For each request, only necessary information is assembled for the agent:
- Recent conversation + compressed long-term memory + relevant past conversations
- Plan document + deliverables + skills
- rawq code search + code-graph dependency information
- Auto-selection of Lite/Standard/Full mode prevents token waste
5. Insight — Project quality analysis
The system pre-extracts rawq (code search), code-graph (dependency analysis), failure history, and test results, then has agents analyze only the extracted data. Instead of reading the entire project, target analysis is done with 5k~20k tokens. Quick Wins even supports automatic fixes.
6. Branch & Adopt
Branch mid-conversation for independent experiments, then insert summarized results into the parent conversation.
7. Long-term Memory & Vector Search
When 12+ messages accumulate, they're automatically summarized by topic. Past conversations are searched with FTS5 + vector hybrid, and related conversations are automatically connected.
---
Technology Stack
Tauri 2 + React 18 + TypeScript + Rust + SQLite (WAL)
DB v29 (29 migrations), Rust 188 + Frontend 175 = 363 tests
---
Noteworthy Points
▎ 415 commits, 42k lines — A codebase entirely written by Claude Code (100%), with the user responsible only for architecture decisions and direction-setting. (Not bragging. But it's become a good world :))
