On-device agent harness for DeepSeek Chat
WebView + native tools + sandboxed workspace.
No ads · No trackers · No analytics
Download APK · Features · Tools · Architecture · Security
D-Harness wraps the DeepSeek chat site in an Android app and gives the model real tools: files, shell, HTTP, GitHub, device features. The model asks for a tool in a special code fence, the app runs it on your phone, and the result goes back into the chat.
- Shell-first agent: multi-step work in one pipeline (needs Settings > exec_shell)
- Sandboxed workspace: list, read, write, edit, search files in an app-private folder
- GitHub and HTTP: the token is stored encrypted; fetch, push and PR helpers
- Skills: 13 short guides the model loads on demand (
skill.list/skill.read) - Device tools: clipboard, battery, network, sensors, notifications, location (when permitted)
- Background agent: a foreground service keeps a tool chain alive while the app is in the background
- Material You settings: light, dark or system theme
- Android 8.0 (API 26) or newer
- A DeepSeek account (the app loads chat.deepseek.com)
Download the APK from the latest release and open it. Allow installs from your browser or file manager when Android asks.
Releases are signed with one long-lived key, so updates install over the previous version. If an older build was signed with a different key, uninstall it first.
The harness executes tools only from a fenced code block tagged invoke:
```invoke
<function_calls>
<invoke name="shell">
<parameter name="command">ls -la</parameter>
</invoke>
</function_calls>
```
- One or more
<invoke>elements per<function_calls>; arguments are<parameter name="...">tags - After the closing fence the model stops and waits for the real result
- Tool markup outside an
invokefence is shown but never run.dsmlandtool_callfences are still accepted for compatibility
An invalid call returns ok: false with error.code and error.hint, so the model can correct itself.
| Skill | Load before |
|---|---|
protocol |
writing any tool call |
shell, exec |
running commands |
workspace |
files in the sandbox |
github, http |
network and GitHub work |
device |
phone state and features |
memory, tasks |
notes and checklists |
research |
web research |
time, safety, skill |
clocks, safe use, how skills work |
Only these names exist. See TOOLS.md for the tool reference.
| Setting | Effect |
|---|---|
| exec_shell | Allows the shell tool. Off by default |
| Auto inject | Injects the tool shim into each page |
| GitHub PAT | Stored in encrypted key storage; used by github.* |
| Explore workspace | Browse, share, rename, delete the agent's files |
- No analytics SDK, ads or trackers. Workspace files and tool results stay on the device unless you use a network tool
- Secrets live in Android Keystore-backed encrypted storage and are excluded from backups
- Camera, microphone and location are granted only to deepseek.com pages
- The shell and exec allowlists reduce mistakes; they are not a security boundary. Only enable
exec_shellif you trust the model's output and the pages you browse. Details and reporting: SECURITY.md
./gradlew :app:testDebugUnitTest # unit tests
./gradlew assembleDebug # debug APK
./gradlew assembleRelease # R8-shrunk release (signed when RELEASE_* env vars are set)CI runs the JS syntax check, unit tests and the release build on every pull request. Pushes to main also sign and publish a GitHub release. See CONTRIBUTING.md.
| Project | What |
|---|---|
| Fern | Live system monitor (Material You) |
| Latch | Privacy-first local automation |
MIT. See LICENSE.