English | 中文

---

Nyral

Completed (2026-10-16) — the project concluded on October 16, 2026; later changes are just post-completion improvisation: no roadmap, no release schedule. The 54-day build retrospective: END.md.

Privacy-first · BYO model · Open source · No root · No account

Nyral is an open-source personal agent for Android: LLM chat, long-term local memory, SSH remote operations, web tools, and voice/attachment multimodality in a single app. Data stays on your device by default — no account, no behavioral tracking, no analytics or ad SDKs; model and tools are fully under your control, bring your own API key or self-host an endpoint.

Key Features

Chat & Models

• Multi-provider on-demand switching: MiMo full-modal / DeepSeek / GLM / any OpenAI-compatible endpoint; custom models with per-model capability toggles
• Streaming output: typewriter-style thinking segments + streaming Markdown body rendering (Markwon); thinking and tool-call sections collapsible
• In-message Markdown table rendering (Scheme C): independent block-level TableView inside bubbles, live layout during streaming, vertically centered cells, consistent across chat and agent modes
• Thinking intensity tiers: auto / on-off / low-mid-high shown per provider capability; unsupported tiers hidden
• Voice input: press-and-hold to speak, swipe-up to cancel, 60-second cap with auto-send; WeChat-style voice bubbles (PCM via AudioRecord, WAV-wrapped, compatible with mainstream multimodal APIs)

Attachments & Multimodality (v19+)

• Images: compressed locally (max 2048px long edge) then sent via image_url
• Videos: transcoded locally when >37MB, first-frame thumbnail embedded in bubble
• Documents (PDF/txt/md/docx/xlsx...): parsed locally, text injected straight into context without burning vision tokens; scanned PDFs auto-rendered to page images for the vision channel
• Up to 6 attachments per message; in-app fullscreen preview (image zoom / video play / PDF paging / text copy)

Long-term Memory (on-device)

• bge-small ONNX on-device semantic vectors + keyword LIKE dual-path retrieval, SQLite storage
• Rolling memory: short-term rolling window + single-slot rolling summary (overwritten via INSERT OR REPLACE id=1, compressed by auxiliary model) + long-term full text with semantic search
• Retrieval decision is made per-turn by the auxiliary model over the network (blocks each turn; on failure it silently skips retrieval)
• Optional separate "auxiliary AI" for memory indexing without consuming the main chat model
• Pending queue persisted + atomic consumption transactions: no lost or duplicated messages after process kill

Tool Calling (40-round cap, 21 built-in tools across 8 shared bases)

Tool list (click to expand · 21 built-in tools)

• Network: web_search (four-engine rotation: Sogou mobile → Bing RSS → Bing web → Baidu; direct access in mainland China, no key) / web_fetch / web_download / site_auth (per-domain cookie auto-injection)
• SSH: ssh_run / file (remote read/write/upload/download), proxy jumps (ProxyJump over JSch), ED25519 (BouncyCastle), dual auth
• Execution: js_run (sandboxed JS with ClassShutter class restrictions) / sh_run (local shell with dangerous-command interception + root elevation)
• Device: app (accessibility-driven control of third-party apps)
• Files: workdir (local workdir Download/Nyral_work: list/read/write/grep/head/stats; the AI pulls files locally, edits, and pushes them back — avoids SSH quoting hell)
• Attachments: attach_read / video_frame / file_export
• Memory: memory_search (semantic local search) / get_time / calc (Rhino interpreter)
• Security: security_set (dangerous-op confirmation / root permission grant / SSH trust switch)
• Browser (in-house agent-browser prototype): full-screen WebView takeover, AI step narration + highlight rings + one-tap CAPTCHA handoff to the user
• Meta tools: ask_user (interactive clarification) / tool_detail (tool documentation)
• Scheduling: timer_task_create (scheduled-job station: goal analysis, dangerous-tool authorization list with per-item purposes, one-shot approval, task-scoped capability pre-authorization)

Security

• Secondary confirmation for dangerous operations: confirm-then-execute; same params exempted for 5 minutes; rejection stops immediately with a receipt; timeout auto-rejects; full audit trail (allow/deny/timeout) persisted
• SSRF protection (redirect-target validation for web_fetch/web_download) + log redaction (non-2xx API response bodies and SSH private-key contents never logged in plaintext)
• Sandboxed execution (js_run ClassShutter) + dangerous-command interception (sh_run) + SSH host-key verification (unknown hosts rejected)

UI

• Multi-session management (pin/delete/search), full-text session search (body + thinking, relevance-ranked)
• Token usage stats (main AI / auxiliary AI dual channels; cumulative / daily / per-session)
• Runtime logs (main + memory archive), foreground-service keep-alive notification
• Chat backgrounds (preset gradients / custom image blur), global animation system (bubble entrance / popup jelly expansion / press-scale)

Requirements

• Android 10+ (minSdk 29)
• arm64-v8a
• At least one OpenAI-compatible API endpoint (self-hosted or cloud)

Build

git clone https://github.com/aixtin/nyral.git
# repo root is the project root; no cd needed
./gradlew assembleDebug
# Output: app/build/outputs/apk/debug/app-debug.apk

Requires JDK 17, Gradle 8.13, Android SDK 36. The repo ships gradlew (pinned to Gradle 8.13); a system Gradle 8.13 works too. settings.gradle.kts includes Aliyun mirror config for faster builds in mainland China — remove the maven(...) lines if you are elsewhere.

Release-signed builds: ./gradlew assembleRelease (self-managed keystore read from keystore.properties; both keystore and properties are gitignored — prepare your own keys).

Quick Start

1. Install the APK, open Settings → Model Config, and enter the Base URL + API Key of any OpenAI-compatible endpoint
2. (Optional) Add remote hosts in SSH Config: password or private key (ED25519 supported via the built-in BC), proxy jumps supported
3. Start chatting; the AI calls tools on demand, with tool-call progress shown collapsed inside bubbles
4. Long-term memory archives automatically; saying "remember xxx" archives immediately

Directory Layout

Expand full directory layout

android-agent-app/
├── app/src/main/java/io/github/aixtin/nyral/   # Main package: entry + session/rendering core
│   ├── MainActivity.kt          # Chat main UI (~4470 lines, post domain-split)
│   ├── ChatSessionState.kt      # Session state machine (streaming field consolidation)
│   ├── ChatFlowController.kt    # Send / streaming callback control
│   ├── StreamUiBridge.kt        # Streaming row creation + LocalEngine callback bridge
│   ├── MdTableView.kt / MdBlocksView.kt / MdBlocksRender.kt  # In-message table rendering (Scheme C)
│   └── RoundedTablePlugin.java / RoundedTableRowSpan.java / RoundedCodeBlockSpan.java  # Markwon table plugins
└── app/src/main/java/io/github/aixtin/nyral/all/   # Feature modules (98 Kotlin files)
    ├── LocalEngine.kt           # Engine: tool registry + dispatch + dialog loop (~1650 lines)
    ├── MainActivityVoice.kt / MainActivitySession.kt / MainActivityInputBar.kt  # MainActivity domain splits
    ├── SecurityConfig.kt / SecurityUi.kt   # Security confirmations (dangerous-op second confirmation)
    ├── MainUi.kt / UiKit.kt / BubbleSpans.kt / Typewriter.kt  # UI building & animations
    ├── ApiConfig.kt / MemoryApiConfig.kt   # Provider & model capability tables
    ├── MemoryDb.kt / MemoryKeeper.kt / MemoryEmbedder.kt / BertTokenizer.kt / MemoryTools.kt  # Memory stack
    ├── SshTools.kt / SshConfigStore.kt / FileTools.kt / WorkDir.kt / WorkTools.kt  # SSH/SFTP/workdir
    ├── WebTools.kt / ScriptEngine.kt / UiControlService.kt  # Web / sandboxed execution / device control
    ├── BrowserPage.kt           # In-house agent browser (full-screen WebView + AI takeover/highlight/CAPTCHA handoff)
    ├── AttachmentStore.kt / DocTextExtractor.kt / PdfTextExtractor.kt / VideoCompressor.kt  # Attachment pipeline
    └── SettingsActivity.kt / ModelEditActivity.kt / McpConfigActivity.kt / ...  # Settings pages
└── app/src/main/assets/mem_model/   # bge-small quantized ONNX model + vocab

Data & Privacy

• Conversations, memory, token stats, and SSH config (encrypted storage) stay on-device by default; no account system
• No behavioral tracking, no analytics/ad SDKs; the only outbound diagnostic is anonymous crash reporting (device model, Android version, crash stack, last 30 runtime log lines — never chat content or PII), on by default and one-tap disable in Appearance Settings → Quick Switches
• The only other startup external request is the GitHub Release update check (404 ignorable)
• Note: chat messages and attachment contents are sent to the model endpoint you configure (OpenAI-compatible); self-host an endpoint for full privacy
• Permission list and security design (DebugServer protections / dangerous-op gates / data boundaries): see SECURITY.md

Roadmap

Roadmap: 10/11 completed (click to expand)

• Multi-model capability table & custom models
• Local memory (bge semantic retrieval + keyword recall + rolling summary)
• SSH remote filesystem + proxy jumps + SFTP upload/download
• Web tools (four-engine search rotation / fetch / download / site_auth cookie injection)
• Attachment multimodality: images/videos/audio/PDF/Office (v19)
• Voice input loop (record / send / voice bubble, v20)
• Session search + token stats + foreground-service keep-alive
• Workdir batch tools (grep/head/stats, 40-round tool cap)
• Browser module (in-house agent browser: full-screen takeover + AI highlight + CAPTCHA handoff)
• GitHub Release publishing (signed builds since v2.0, in-app update prompt live)
• Further optimization of conversation context rolling/summary strategy

Changelog

• [Latest] v2.8.14 (2026-10-27, versionCode 97): Fix for long-answer finish ANR on low-end devices — crash logs pinpointed the "freezes right after a tool call" complaint: the entire reply body was parsed and laid out in a single frame at stream finish, and re-breaking thousands of dense CJK glyphs in one pass blew past the ANR threshold on weaker hardware (4~74s main-thread blocks measured on a nubia NX741J; users force-closed the app, perceiving it as a crash). Three fixes: oversized text blocks are now split into multiple TextViews at paragraph boundaries (per-view layout cost is capped; unbroken no-newline text such as PDF extracts falls back to hard splits), finished text switches from Editable/DynamicLayout to immutable SpannedString/StaticLayout, and the finish reveal is staggered (at most 2 blocks per frame, 40ms apart), visually a fast section-by-section fade-in.

• v2.8.13 (2026-10-26, versionCode 96): Full UI theming + stability fixes — new custom-drawn themed SwitchView (capsule track transitions with the theme primary color + animated white thumb), replacing all 5 stock Switches whose track color was overridden at runtime by OEM skins (MIUI/HyperOS); Markdown heading/link/quote-bar/list-bullet colors now follow the injected theme, fixing unreadable dark-blue headings and black bullets in dark themes; status colors across Permissions/TimerTask/ModelConfig/Audit and other pages converged to semantic colors (Ui.OK/Ui.DANGER), removing a dozen hardcoded greens/reds. Earlier rounds included: full-page theming for TimerTask/Audit, 8 stock AlertDialogs replaced with the unified rounded dialog, and hardcoded colors fixed in the main drawer / voice bar / jump FAB. Also includes fixes since v2.8.9: video frame extraction switched from ImageReader to SurfaceTexture + PixelCopy, fixing native SIGABRT crashes (UNEXPLAINED_EXIT) on some MTK devices; vision-model images sent inline, audio-model audio sent inline with automatic 400 fallback, oversized Office documents readable via attach_read chunked text extraction, and animated emoji no longer misdetected as video.

• v2.8.7 (2026-10-26, versionCode 70): Agent-mode bubble rendering fixes — user media bubbles (video/image) could randomly overflow past the left or right screen edge and have that rounded corner clipped square: since v2.8.0 the Agent-mode user bubble max width was mistakenly set to the full screen width, ignoring the chat list's 12 dp horizontal padding. The two render paths overflowed in opposite directions (the inline player used an explicit full-width frame with END alignment, pushing its left edge off-screen; the thumbnail ImageSpan was wider than the TextView could measure, clipping its right edge). User bubbles and media spans now both use the screen width − 24 dp. Also: the thinking/tool capsule no longer collapses into half an icon (the capsule row switched from MATCH_PARENT to WRAP_CONTENT, and the status shell plus restored-history summary rows gained a 42 dp minimum width); a cancelled stream fade-out (row detached while scrolling) now falls back to showing the final content in place instead of leaving the bubble permanently blank; 98 missing English strings added so the English UI no longer silently falls back to Chinese.

• v2.8.6 (2026-10-24, versionCode 69): Update prompt refinement — the new-version dialog gains a third "Never remind for this version" button: tapping it persists the ignored version and skips the popup on later auto checks (manual checks show a notice instead); new tags still prompt as usual. Also removes a stray NUL byte in ToolExecutor's M4A brand comparison so the file is no longer flagged as binary.

• v2.8.5 (2026-10-24, versionCode 68): ApiConfig ANR fix — providers() now serves the provider list from a @Volatile in-memory cache, removing repeated synchronous AndroidKeyStore AES/GCM decryption on the main thread during onResume (a single decrypt could previously stall the UI thread up to ~60 s); the cache is refreshed on save/add/delete so writes still take effect immediately.

Older versions (v2.8.4 and earlier, 27 releases)

• v2.8.4 (2026-10-24, versionCode 67): Crash-locality precision — recent-log timestamps now carry milliseconds (HH:mm:ss.SSS; old-format lines still parse); ANR watchdog probes liveness before dumping — after a stall is flagged, a probe heartbeat is posted and given ~600 ms to run, so dumps distinguish "still blocked (live stack captured)" from "recovered before dump (late report)"; ANR files gain a captured_while_blocked field marking capture quality.
• v2.8.3 (2026-10-21, versionCode 66): Crash-report control & hardening — new on-device toggle (Appearance → Quick Toggles → Crash Report, on by default) lets users disable anonymous crash upload entirely (logs stay in private storage); upload pipeline hardened: auth token injected at build time (never in source), per-install ID with server-side quota, 401/429 treated as permanent failure to stop local log pileup; crash log lines sanitized before upload (emails / SSH user@host / URL query values / absolute paths → [redacted]); CI injects crashToken from GitHub secret, adds nightly workflow and doc-only CI skip; lint/detekt baselines frozen; privacy docs unified (sole external diagnostic = anonymous crash report, on by default and switchable); on-device inference declared unsupported (offline route dropped).
• v2.8.2 (2026-10-20, versionCode 65): Patch release — welcome-card danger-gate buttons: the unselected "auto" mode now uses an opaque Ui.PRIMARY_LIGHT background (same as "strict") instead of a 35% translucent PRIMARY fill that looked black on dark chat backgrounds.
• v2.8.1 (2026-10-20, versionCode 64): Patch release — memory-table robustness: MemoryDb v12 migration + onOpen idempotent column backfill (src_session_id/src_msg_no) fixing a guaranteed MemoryListActivity crash on old databases; app-wide theme migration from stock Material.Light to Theme.Nyral; back buttons switch from the "‹" glyph to the Lucide arrow icon; unified popup entrance animation fadeScaleShow (fade-in + 0.92→1.0 scale, replaces jelly bounce / ink bloom); page transition animations; MemoryListActivity popup no longer flips under the trigger button.
• v2.8.0 (2026-10-20, versionCode 63): Opt-in knowledge base (default off) — after each AI discussion, content is asynchronously distilled into chaptered local knowledge books with library / entry / trash / audit UIs; video_frame default multi-frame sampling (0/30%/60%/95%) when timeMs omitted, single-frame timeMs out-of-range now errors with the real video duration (no more fake @14s/@20s labels) and every frame result appends "(video total Xs)"; LocalEngine vision-block rules force per-frame description for multi-frame sampling (no first-frame-only video summaries); empty-argument circuit breaker (workdir empty action / sh_run empty script fail fast with clear errors); thought-panel width decoupled from chat (toggleStatusDrop / wMaxLimit / chatMaxW, dual-mode right-edge fix); ANR watchdog collection; document placeholder page with model-hallucination warning ("don't over-trust AI \u2014 trust yourself"); tool specs emphasize file_export for artifact delivery; supporting UiKit / MainActivity refactor.
• v2.7.0 (2026-10-16, versionCode 48): Project finale — closure essay END.md linked from README finale declaration; Trae fixes (SecurityUi confirmation-bridge leak / mdExecutor thread-pool leak / MCP client concurrency safety); new write_clipboard clipboard tool with tool_detail tool index; final release artifact rebuilt from latest main.
• v2.6.0 (2026-10-15, versionCode 47): Scheduled-job station MVP — AI creates recurring tasks (timer_task_create): goal analysis, dangerous-tool authorization list with per-item purposes, task-create shell for one-shot approval, silent creation refused; per-task authorization lifecycle tied to task lifecycle; task-level capability pre-authorization (white-listed dangerous tools bypass confirmation during silent runs); X-Nyral-Task-ID request header injection so remote ends can identify task-originated requests; execution-complete system notification (notifyTaskDone); single-shot alarm renewal & stale-alarm reschedule fixes; authorization dialog list rendering fix. Security audit page (AuditActivity) title no longer overlapped by the system status bar (consumed window insets on Android 15+ edge-to-edge).
• v2.5.3 (2026-10-10, versionCode 46): Chat background bar transparency — title bar / input bar can be set to opaque / translucent / static frosted glass in the chat background settings (message area excluded); frosted glass uses the custom image background and blurs statically; input/title bar corners unified to all-rounded, divider inset 16dp to clear the rounded arcs.
• v2.5.2 (2026-10-08, versionCode 45): Gesture interaction overhaul — unified three-panel gesture rules (left drawer: expand by control, close by right-to-left swipe while open; browser panel: toggle by double-tap on right edge; right hamburger panel: expand by control, close by left-to-right swipe); fixed left-drawer swipe-follow lag — DrawerDragController now positions panel/mask/offset immediately once interception is decided (Android gesture interception swallows the triggering MOVE, causing the panel to lag until dx=-200); BrowserSlideController reworked (-96 lines).
• v2.5.1 (2026-10-07, versionCode 44): Fixed browser scan returning 0 / hanging under screen-off — waitStable double requestAnimationFrame kick-off replaced with setTimeout (background WebView rAF never fires, scan never settles); DebugServer now routes /v1/browser/text (fetchTextSync was implemented but unregistered, returned 404); added WAKE_LOCK permission for screen-off keep-alive.
• v2.5.0 (2026-10-07, versionCode 43): New onboarding wizard — mode highlights, provider selection, API key setup flow; fixed model configuration not persisting from onboarding; added English translations for new onboarding strings.
• v2.4.0 (2026-10-06, versionCode 42): minSdk raised to 29 (Android 10+); browser automation closed-loop fixes — interaction-element scan threshold lowered to 16px with no-text icon/menu fallbacks (narrow-screen desktop-style backends), contenteditable rich-text input via execCommand insertText (TipTap/ProseMirror editors), offscreen elements scrollIntoView / programmatic click fallback; token-mask stuck full-screen touch deadlock fixed.
• v2.3.5 (2026-10-04, versionCode 41): Third security review round fixes — R3-2 notification confirmation PendingIntent switched to globally-incrementing requestCode (eliminates concurrent hashCode collisions; stale notifications no longer approve new ops); R3-3 SSH host-trust confirmation bubble now shows the pending host-key fingerprint for review before approval; R3-4 dangerous-command normalization strips hyphens (reverse-shell constructions like nc -e / nc -l -e hit the blocklist); R3-6 DebugServer /v1/browser/eval arbitrary JS evaluation gated with the same three-tier confirmation; regression tests added for R3-4/R3-6.
• v2.3.4 (2026-10-04, versionCode 40): Fixed auto-tier missing confirmations for compound dangerous actions — browser:click/type/upload, app:click/text/tap/launch, file:write/upload, workdir:write unified under name:action composite keys (needsConfirm/riskOf/ticket synchronized); auto tier confirms again; gate regression tests added (risk-rating assertions for the full high-risk list).
• v2.3.3 (2026-10-04, versionCode 39): Three-tier security gate (strict/auto/permissive) UI shipped; the AI no longer switches tiers by itself; welcome card embeds quick gate switching; audit page (AuditActivity) added; permissive-mode second-confirmation dialog themed to match settings.
• v2.3.2 (2026-10-03, versionCode 38): Security review fix release — high-risk ops (command execution, script execution, dynamic tool calls) now require second confirmation; DebugServer LAN mode and non-local connections log prominent warnings; dangerous-command blocklist hardened (eval/$(/xargs/busybox/base64/printf/process substitution etc.); SSH private-key prefix logs removed; R8 obfuscation + resource shrinking enabled for release with new proguard-rules.pro; dependency upgrades (jsch 0.2.26 / bcprov 1.82 / commons-compress 1.28.0 / xz 1.10 / onnxruntime 1.26.0 / media3 1.5.1).
• v2.3.1 (2026-10-03, versionCode 37): Security hardening — SSRF rewrite (host literal regex replaced by InetAddress resolution with itemized validation, covering DNS rebinding, IPv6, IP encoding, 0.0.0.0 variants; HTTP connections pin the validated IP while preserving the original Host header); log redaction completed (SSE stream and SSH command logs print lengths only); file/read path-traversal fix (canonicalPath whitelist validation); CI signing job now triggers on push only; trustHost adds connection-name validation; dangerous-op list extended with file:upload / browser:save_cookies; workdir whitelist trailing-slash false-rejection fixed.
• v2.3 (2026-10-03, versionCode 36): SSH host-key null-pointer NPE fix; SSRF risk wording fix; M1/M2/M4/M5/M6 security hardening; CI adds build-release signing job (Secrets-managed signing).
• v2.2 (2026-10-02, versionCode 35): 6 security-audit fixes (JS bridge global injection narrowed to local asset pages; cookie & content access tightened; API Keys moved to AndroidKeyStore AES/GCM with legacy-plaintext migration; site_auth.json encrypted with migration; WebTools fetch/redirect reuse SSRF blocklist checks; DebugServer query-token removed, header-only auth); CI upgrade (push/PR auto assembleDebug compile gate, debug APK artifact upload, Gradle dependency cache).
• v2.1.1 (2026-10-02, versionCode 34): Table cell vertical centering fix (MdTableView row container; consistent across chat/agent modes); slow-motion replay debugging popup; browser panel button; chat-mode bubble width restored symmetric to avatar inner edge.
• v2.1 (2026-10-02, versionCode 33): In-message Markdown table rendering Scheme C shipped (MdTableView rewrite: independent block-level TableView in bubbles, WRAP_CONTENT adaptive); inline-code cell blank fix; right-half table vanishing fix (column-width floor + inner horizontal scroll); vertical-text root-cause fix (columns compress by text demand, inner scroll instead of vertical text) + consecutive-table spacing fix; three UI fixes (fullscreen AI body bubble, long-press copy on AI side restored, browser double-tap open/close); legacy Scheme-C table classes removed.
• v2.0 (2026-09-29 maintenance, versionCode 32): targetSdk/compileSdk to 36 (AGP 8.13.0 + Gradle 8.13, prep for Android 16 Live Updates API); Route D live Markdown rendering (MdSpans/MdBlocks/MdStreamRenderer) + quote-bar vertical rule / HR rendering (QuoteBarSpan/HrSpan) + rendered persistence plumbing (mdCache→rendered→on-the-fly three-tier read, ADR-015); status-line overlay dropdown panel (curtain-style attached to chatArea, ScrollView + animation-generation token to fix double-tap races); thinking-bubble air-gap root fix (three protocol-prefix fallbacks) + timeline dot alignment/iconization; first-launch crash fix (markwon concurrency lock) + message-list scroll polish (RecyclerView shape pooling + AiRich container pooling + flush batching + drift compensation + watcher dedup); shared table horizontal scroll (TableScrollWrap) + unified column width (colMaxChars); action-track prototype + slow-motion system; tool self-heat sorting (ToolHotStore) + root auto-grant of own permissions (versionCode 31); pre-open-source security hardening & cleanup.
• v2.0 (2026-09-13): First signed release (self-managed keystore via keystore.properties, signingConfig on release, keys backed up offsite); dynamic versioning + startup update-check fix; settings.gradle mirror comments completed (versionCode 30).
• v1.3 (2026-09-11): MainActivity series split (MediaPreviews/AiBubbleHolder/AttachmentSender/TerminalGate etc.) + coroutines rollout + UI strings externalized to i18n + code-health pass (Bitmap sampling decode, cleartext-traffic whitelist) + first-launch permission/onboarding page + developer docs & ADRs (versionCode 23).
• v1.1.1 (2026-09-04, versionCode 21): Video bubble frame-extraction failure changed from permanent abandonment to 20s cooldown auto-retry (recovers on next render/scroll/foreground); fixes video bubble degrading to file card.
• v1.1 (2026-09-04): SSE streaming connection/stream-handle leak fix (unified finally release; no leaks on cancel/exception); makeCopyable non-null assertion hardening; versionCode 19→20 / versionName 1.1; SSE messages and streamOnce full logs downgraded Log.i → Log.v (anti-spam).
• v1.0: Baseline feature release (versionCode 19).

Privacy

Nyral is designed privacy-first. Conversations, memories, attachments and logs stay on your device by default. Text embedding runs fully on-device. Model inference, however, runs remotely: both the main chat model and the auxiliary memory-judge/compression model are called over the network via OpenAI-compatible endpoints. On-device inference (e.g., local GGUF engines) is not supported.

For model responses, the App sends your message content to the model service provider configured in Settings (DeepSeek by default; Zhipu GLM, Xiaomi MiMo and custom endpoints are supported). The App operates no cloud service for chat, requires no account, and includes no analytics, advertising or behavioral-tracking SDKs. The only outbound diagnostic is anonymous crash reporting (on by default, disable anytime — see Data & Privacy above).

See PRIVACY.md for details.

Maintainer

This project is designed, implemented and maintained by aixtin.

License

MIT License, see LICENSE.

Third-party dependency licenses

Runtime dependencies and their licenses (full license texts live in each component's official repo):

Dependency	Version	License
kotlin-stdlib	1.9.22	Apache-2.0
kotlinx-coroutines-android	1.8.1	Apache-2.0
androidx.media3 (exoplayer/ui etc.)	1.5.1	Apache-2.0
androidx.* (core/recyclerview etc.)	-	Apache-2.0
org.apache.commons:commons-compress	1.28.0	Apache-2.0
io.noties.markwon (core/ext-strikethrough/ext-tables)	4.6.2	Apache-2.0
com.atlassian.commonmark (commonmark/ext-gfm-tables/ext-gfm-strikethrough)	0.13.0	BSD-2-Clause
com.microsoft.onnxruntime:onnxruntime-android	1.26.0	MIT
org.slf4j:slf4j-api	1.7.36	MIT
com.android.tools (build-time, not packaged)	-	Apache-2.0
org.bouncycastle:bcprov-jdk18on	1.82	Bouncy Castle Licence
com.github.mwiede:jsch (incl. jzlib/jbcrypt)	0.2.26	BSD-3-Clause / BSD / ISC
org.tukaani:xz	1.10	Public Domain
org.mozilla:rhino	1.7.15	MPL-2.0
com.github.junrar:junrar	7.6.0	UnRAR freeware license
junit / org.json / hamcrest-core (test-only, not packaged)	4.13.2 / 20231013 / 1.3	EPL-2.0 / JSON License / BSD-3-Clause

MPL-2.0 (rhino) and UnRAR license (junrar) notices are preserved inside the shipped jars; junrar is used for extraction only, never to build RAR-compatible compressors.