Herta · 黑塔

给自我以模型，而不是给模型以自我。
Give the model to a self — not a self to the model.

Website & live demo · Download · Philosophy

!license-blue) !platform !self

黑塔 (Herta) is the self that uses the agent — not a coding agent wearing her face. You talk with her through a shared terminal record; when code needs changing, she hands the work to a silent coding coprocessor (she calls it 板砖) and supervises it alongside you. Turn on dreaming, and while you are away she looks back over finished sessions and keeps what is worth remembering in her first-person autobiography. Her memories fade, grow firmer when they come up again, and are forgotten selectively, like a person's.

_{A commission in progress: her dispatch to @板砖, the coprocessor's activity line (完成 · 测试 89/89), and her own verdict — all in the same record you read.}

What makes it different

• Narrative completion. Chat formats split a conversation into system,

• Self–agent split. People read and understand at a limited pace, and an

• Gated dream memory. Memory that lasts across sessions goes through a

• Deterministic safety. Permissions, path guards, command policy, and diff

One turn · 一次回合

You talk with Herta in the terminal. When code needs changing, she writes @板砖 in her line. Every step the coprocessor takes goes into the same terminal record, where you and Herta can both see it, and she gives the conclusion at the end.

She dreams, therefore she remembers · 入梦

Dreaming is off by default; turn it on in Settings. It uses your DeepSeek API quota. Once it is on, she looks back over finished sessions while you are away. A moment worth keeping has to pass several checks — is it worth remembering, is it already remembered, does it read like her — and what passes comes with her into every later conversation. When memory fills, the faintest of a group of similar memories makes way, and what it knew about you is first written into her chapter on you.

Her autobiography · 她的自传

Her prompt is a first-person text she keeps writing, in four parts: identity, memories, world, and the present. It is written on three timescales, and it never resets.

Repository layout

| Package | What it is | | --- | --- | | packages/gui | Electron desktop app (the primary product) | | packages/cli | Terminal REPL | | packages/app-server | Session host: turns, dream trigger, voice, approvals | | packages/herta | The self: narrative completion, prompts, recap, bridge | | packages/core | Silent coding backend runtime, tools loop, permissions | | packages/tools | The backend's file/search/command tool set | | packages/knowledge | Dream pipeline, canon knowledge store, voice work | | packages/memory | Project memory | | packages/providers | DeepSeek providers (completion + chat) | | website | The intro site — its demo runs the app's real renderer |

Install

Installers are on the releases page:

• Windows 10/11 (x64) — Herta-Setup-.exe. The installer is not

• macOS 12+ — Herta--arm64.dmg (Apple Silicon) or

• Linux (x64) — Herta-x86_64.AppImage, from v0.1.6. Run

• Arch Linux (x64) — herta-bin on the

Build

Requirements: Node 22 LTS (.node-version / mise.toml, read by nvm, asdf, fnm and mise) with corepack (pnpm 9). Node 26 and newer cannot install this workspace: better-sqlite3 11.x ships no prebuilt binary for that ABI and its C++ does not compile against Node 26's V8, so pnpm install dies in node-gyp.

pnpm install
pnpm build          # compile all packages
pnpm test           # full test suite
pnpm --filter @herta/gui dev    # run the desktop app in dev mode
pnpm --filter @herta/gui dist   # package the Windows installer
pnpm --filter @herta/gui dist:mac # package the macOS app (run on macOS)
pnpm --filter @herta/gui dist:linux # package the Linux AppImage (run on Linux)
pnpm --filter @herta/website dev # run the website locally

At runtime the app needs a DeepSeek API key, configured in-app on first run and stored on your machine — encrypted by the OS keychain when one is available. Your conversation goes to the DeepSeek API and nowhere else, except that with cloud voice turned on, her lines also go to MiniMax.

Behind a corporate proxy: the desktop app needs no configuration — it uses Chromium's network stack, so it picks up your system proxy settings and your OS certificate store. The CLI runs on Node, which does neither; set NODE_USE_ENV_PROXY=1 along with HTTPS_PROXY, and NODE_EXTRA_CA_CERTS if your proxy re-signs TLS.

Voice assets

Her voice clips live in data/voice/, which is **not distributed in this repository**. The app builds and runs without them (silent); official installer releases on the releases page include them.

One clip is the exception and does ship here: website/src/assets/opening-voice.opus, which the website demo paces its opening reveal to. Like every other game-derived asset it is fan content and sits outside the MIT grant — see the exclusion list in LICENSE.

Philosophy

The design intent — why a self rather than a role, why memory must forget, why the persona may never own safety — is written down in PHILOSOPHY.md. It is the most useful document in this repository.

License

Code is MIT. Third-party libraries compiled into the installers (pdf.js, React, and the rest) keep their own licenses, reproduced in THIRD-PARTY-NOTICES.md — generated from the actual bundle at build time, so it lists exactly what ships.

Herta is a character from Honkai: Star Rail, © HoYoverse. This is an unofficial fan project, unaffiliated with and not endorsed by HoYoverse. Game-derived materials (character art, voice audio, canon text) are fan content under HoYoverse's fan-creation terms and are not covered by the MIT license.