Aura – a self-hosted AI agent with temporal graph memory, written in Go

Aura, a self-hosted, multi-user AI agent written in Go, ships as a single binary hosting the runtime, tools, CLI, Telegram gateway and embedded web cockpit, with each user getting their own memory database, workspace and sandbox. The default stack measured 7 GB of RAM with speech-to-text and text-to-speech running on a 16 GB mini PC as of 2026-09-02, and the project's latest tagged prerelease is v1.0.2-rc1, checked on 2026-10-03, while the edge branch tracks master. Aura defaults to DeepSeek-V4 Flash via OpenRouter and also supports a ChatGPT plan, a bundled llama.cpp server running Gemma 4 12B QAT, or Ollama, with the active profile hot-reloaded from cockpit settings without a restart.

A local-first, provider-neutral AI agent platform — in Go. An agent for ongoing work: tools, document retrieval, temporal memory, scheduled jobs, and a web cockpit on infrastructure you control. What is Aura? what-is-aura · Features key-features · Studio studio · Compare how-aura-compares · Architecture architecture-one-screen · Quick Start quick-start · Docs documentation · Development development Aura is a self-hosted, multi-user AI agent. Its Go binary hosts the runtime, tools, CLI, Telegram gateway, and embedded web cockpit; each person signs in to their own identity, with their own memory database, workspace and sandbox. Docker Compose runs Postgres, ArcadeDB, Garage, embedding, ingestion, web search, voice and the bundled integrations alongside it. The model is chosen in the cockpit settings: OpenRouter the default route , a ChatGPT plan, the bundled local llama.cpp server, or Ollama. Local storage does not make cloud inference offline: a cloud provider receives the context sent to its model; with a local server nothing leaves the host for inference. Hardware: a mini PC with 16 GB of RAM is enough. The default stack measured 7 GB with speech-to-text and text-to-speech running on a 16 GB mini PC 2026-09-02 . No local LLM runs by default: inference goes to the provider you choose. <sub A real run on a local stack: Aura stores the fact in its memory graph, loads the deferred task tool and schedules the reminder; a new chat then answers from memory, with provenance. Model replies were written by Claude through an OpenAI-compatible endpoint; waiting time is trimmed.</sub | Language | Go 1.27 | | Tests | Unit, property, race, leak, mutation, live integration, and browser tests | | Test coverage | Owned-surface aggregate ≥85% , with package policies and separate live memory/sandbox coverage authorities | | CI | build/vet/lint · CodeQL · -race + goleak · db/ArcadeDB/embed integration · MUSR two-identity E2E · web lint/test/mutation/Playwright · critical mutation ≥70% killed | | Persistence | Postgres sqlc, pgx + ArcadeDB graph memory, full-text + LSM vector index + Garage S3 object store | | Models | Default DeepSeek-V4 Flash via OpenRouter; also a ChatGPT plan, the bundled llama.cpp server Gemma 4 12B QAT, localllm profile or Ollama. The active profile provider, model, budgets is hot-reloaded from the cockpit settings, no restart | | Distribution | edge tracks master; v1.0.2-rc1 is the latest tagged prerelease checked on 2026-10-03. See Releases for current availability | - Streaming agent loop with shared step/time budgets and repeated-call controls to bound work. - Deferred tools and tool search — discover tools and load their schemas when needed, including tools from mounted MCP servers. - Adaptive reasoning router — selects reasoning effort using the configured classifier and the active model's supported capabilities. - Full host terminal + filesystem tools — real operating power, with destructive-command approval gates and secret redaction. - Graph-native memory — facts, sources and validity windows in ArcadeDB; temporal paths return supporting evidence. Postgres-authoritative conversations have a derived recall projection and managed context compaction. - Document retrieval — indexed passages with source hashes and citations, plus access to the original file for calculations and whole-file tasks. - Self-extension — author and run skills, use bundled memory/PIM/WhatsApp integrations, and connect additional MCP servers. - Scheduler and self wake-ups — one task tool at | every | cron for reminders and agent job runs, with job policy, operator controls and outcomes delivered to the owning conversation. - Per-identity sandbox — a full-capability box per operator opt-in sandbox profile; gVisor runsc on native Linux , with deliverables handed back over the channel send file , never as a path. - Multi-user — Authula sign-in password, plus a TOTP step for accounts enrolled in it , one isolated ArcadeDB database per identity enforced by the server, capability grants, and an admin audit view. - Multi-channel — CLI REPL, Telegram voice/photo/docs/HITL , and a web cockpit over AG-UI/SSE with mid-turn steering, approvals, voice input/output, and live settings. - Studio — image and video generation, photo and video editing, and a multi-track video editor, all in the cockpit details studio . - Bundled integrations — calendar/e-mail PIM MCP, OAuth providers , WhatsApp unofficial client , web search through a bundled SearXNG, snapshot share links to a conversation, and Cloudflare remote access. The cockpit's creative workspace, per identity. - Generate images and video from a prompt over OpenRouter's media models. The model picker shows each model's price per image, per second or per million output tokens and the estimated cost before you press Generate. Options cover resolution, aspect ratio, seed, and duration and sound for video; advanced inputs take a start frame, an end frame and reference images from your library. Every generation lands in a searchable history you can reuse or download. Generation needs the OpenRouter route. - Edit photos and clips in the browser: a photo editor Filerobot and a quick video editor trim, crop, rotate, audio for any image or clip in a chat or in the Garage library. - Multi-track video editor : - a video lane plus overlay lanes for titles and images, with transitions between clips; - per-clip transform fill, fit, crop, flip, rotate , adjustments opacity, brightness, contrast, saturation, hue, blur , animations and speed; - audio lanes for an uploaded sound, a recorded voice, a text read aloud, or the sound extracted from a clip, with noise reduction, fades and automatic ducking under speech; - undo and redo, saved projects, a mobile layout, and an export rendered in the browser video, or the audio alone as WAV . A generated video opens in the editor with one click. Checked on 2026-10-03 against each project's own documentation. Open WebUI and LibreChat are mature, much larger projects; this table shows where Aura differs, not that it is ahead. | | Aura | Open WebUI | LibreChat | |---|---|---|---| | Backend | Go, one binary + Compose appliance | Python | Node.js | | License | MIT | Open WebUI License BSD-3 up to v0.6.5; branding must stay above 50 users | MIT | | Long-term memory | Temporal knowledge graph: facts with sources and validity windows, one ArcadeDB database per identity | Facts and notes the model can search and update | Memory with per-agent partitions | | Scheduled work | task tool at , every , cron running full agent jobs | Scheduled prompts | Scheduled Chats beta | | Tool approval HITL | Yes | Not documented | Yes v0.8.8 | | Messaging channels | Telegram, WhatsApp, e-mail/calendar | Not documented | Not documented | | Video | Generation plus a multi-track editor | Voice and video calls | Not documented | | Single sign-on | No: email/password TOTP for enrolled accounts | SSO/OIDC, LDAP, SCIM | OAuth2, SAML, LDAP | | Community | Small, one maintainer | Very large | Large | Choose Open WebUI or LibreChat for a polished multi-model chat front end with SSO and a large ecosystem. Choose Aura for a long-running personal agent that remembers over time, works on a schedule and reaches you on Telegram or WhatsApp. Transport & UX cmd/aura CLI · channels +telegram · agui SSE · webui embedded SPA · webauth Authula · setup · askuser Agent runtime agent LlmAgent, Budget, Events, hooks, workflow Seq/Par/Loop · runner · swarm · steer Tools & MCP agent/tools registry, deferred, tool search, fs/shell/web/skill · agent/mcptools · mcp +manager · mcpoauth · sandbox Intelligence llm +openai compat · chatgptplan · semindex embed-index core · reasoningtrace · scoring · multimodal · mediagen Capabilities web · skills · cron · onboarding · documents · share · retention Persistence db Postgres+sqlc · arcadedb memory + retrieval · conversations · identity · objectstore · secret · settings Observability obs · agent/panicobs · reasoningtrace · toolinvocations · cachemetrics | Doc | For | |---|---| | docs/ARCHITECTURE.md https://github.com/chetto1983/Aura/blob/master/docs/ARCHITECTURE.md | How the system is built — layers, turn lifecycle, invariants | | docs/TECHNICAL OVERVIEW.md https://github.com/chetto1983/Aura/blob/master/docs/TECHNICAL OVERVIEW.md | CTO / due-diligence overview — problem, differentiators, maturity | | docs/CAPABILITIES.md https://github.com/chetto1983/Aura/blob/master/docs/CAPABILITIES.md | Capability matrix — shipped / in-progress / roadmap | | docs/release-readiness.md https://github.com/chetto1983/Aura/blob/master/docs/release-readiness.md | How a release is cut — the twelve-report exact-SHA gate, rollback rule, operational checks | | docs/BACKUP-RESTORE.md https://github.com/chetto1983/Aura/blob/master/docs/BACKUP-RESTORE.md | Backup schedules, recovery procedures, live validation and scope | | CLAUDE.md https://github.com/chetto1983/Aura/blob/master/CLAUDE.md · prd.md https://github.com/chetto1983/Aura/blob/master/prd.md | Engineering guidance · product requirements source of truth | Aura is a self-hosted agent runtime packaged as a Docker Compose appliance. The default stack brings up Aura with its migration one-shot , Postgres, ArcadeDB and its MCP, Garage, the local embedding sidecar, document ingestion, SearXNG, speech-to-text and text-to-speech, the PIM and WhatsApp MCP sidecars, the Cloudflare tunnel supervisor idle until enabled , and Caddy in front of the Authula sign-in. Compose profiles add the rest: localllm a llama.cpp server with Gemma 4 12B QAT , ocr , observability Prometheus, Tempo, Grafana and sandbox the Docker socket proxy for per-identity boxes . Releases. ghcr.io/chetto1983/aura:<tag and the binary archives are published by the Release workflow on a v tag, and only after the exact-SHA Production Readiness check passed for that commit docs/release-readiness.md https://github.com/chetto1983/Aura/blob/master/docs/release-readiness.md . Check the Releases page https://github.com/chetto1983/Aura/releases for the current tag v1.0.2-rc1 is the latest and use it as vX.Y.Z below. Independently of releases, every master push publishes the moving ghcr.io/chetto1983/aura:edge image plus an immutable master-<sha tag — the continuous-delivery channel a default install tracks. The interactive installer supports local installation or a Linux target over SSH: npx create-aura-appliance npx create-aura-appliance --mode remote It requires Node.js 22.13 or newer on the workstation. The target needs at least 4 CPU cores, 14 GiB usable RAM, and 20 GiB free disk ; documents, models and backup retention need additional capacity. The installer detects the embedding backend on the target: CUDA for an NVIDIA GPU Docker can drive, otherwise Vulkan for an Intel or AMD GPU exposing /dev/dri , otherwise CPU. See the installer guide https://github.com/chetto1983/Aura/blob/master/packages/create-aura/README.md for supported targets and prerequisites. The npm installer carries its own payload. The source-hosted installer remains available: Install Docker, then run the installer. One command on a machine with Node 18+ npx fetches the repo and runs scripts/install.sh : sudo npx github:chetto1983/Aura -- --appliance or the curl equivalent of the same script — use master to track the edge channel, or a release tag vX.Y.Z to pin: curl -fsSL https://raw.githubusercontent.com/chetto1983/Aura/master/scripts/install.sh | sudo bash -s -- --appliance The installer checks hardware, creates .env with generated POSTGRES PASSWORD , the three ARCADEDB secrets, and AURA ACCESS TOKEN , downloads the Compose/Caddy assets, and starts the stack. Re-running it keeps an existing .env intact. A master/edge install points .env at the :edge moving tags, and --appliance also enables the aura-image-update systemd timer: from then on the machine re-pulls aura and its MCP sidecars from GHCR on its own, migrations and Compose payload included, with no operator involved. Without --appliance no systemd units, no timer , the stack still starts; updates stay manual. The default stack also starts the Cloudflare supervisor healthy-idle. Enable it after setup in Settings Remote access ; the installer requires no Cloudflare credential. Existing edge appliances receive the sidecar through the payload updater without replacing database volumes. See Cloudflare Remote Access https://github.com/chetto1983/Aura/blob/master/docs/cloudflare-remote-access.md for registered-domain prerequisites, Access OTP, organization WARP, token refresh and safe disable/delete. Add --gvisor on native Linux Docker hosts that should run Aura under runsc . Docker Desktop is intentionally not supported for that isolation tier. Use Docker Desktop and the shipped Compose files. From PowerShell in the Aura checkout or release directory: function New-Hex { -join 1..32 | ForEach-Object { '{0:x2}' -f Get-Random -Maximum 256 } } @" POSTGRES PASSWORD=$ New-Hex POSTGRES USER=aura POSTGRES DB=aura ARCADEDB PASSWORD=$ New-Hex ARCADEDB APP PASSWORD=$ New-Hex AURA ARCADEDB TENANT SECRET=$ New-Hex AURA IMAGE=ghcr.io/chetto1983/aura:vX.Y.Z AURA ACCESS TOKEN=$ New-Hex AURA AUTHULA SECRET=$ New-Hex SEARXNG SECRET=$ New-Hex AURA OBJECTSTORE ACCESS KEY=GK$ New-Hex .Substring 0,24 AURA OBJECTSTORE SECRET KEY=$ New-Hex GARAGE RPC SECRET=$ New-Hex AURA GARAGE ADMIN TOKEN=$ New-Hex AURA BACKUP DIR=./backups AURA EMBED REVISION=0f741b5a6585bd53aeb15cd1372c56f2a0f65e12 AURA EMBED FINGERPRINT=b5ce9d77a3fc4b3b39ccb5643c36777911cc4eb46a66962eadfa3f5f60490d63 AURA EMBED NGL=99 "@ | Set-Content -Path .env -Encoding ascii docker run --rm --gpus all nvidia/cuda:12.8.0-base-ubuntu24.04 nvidia-smi docker compose up -d This .env targets an NVIDIA GPU: the embedding sidecar reserves one, so fix Docker/NVIDIA before starting Aura if the nvidia-smi container check fails. Without an NVIDIA GPU, run embeddings on the CPU instead: set AURA EMBED NGL=0 and COMPOSE FILE=compose.yaml;compose.cpu.yaml ; is Compose's path separator on Windows , which drops the GPU reservation and selects the CPU build of the pinned llama.cpp server. The model route, the OpenRouter key and the Telegram bot token are not .env settings: choose them in the first-run web setup, and Aura keeps them in aura.settings . For local development images, replace AURA IMAGE with aura:local after building the image. Postgres 18 is the default Compose image for new installs. When upgrading an existing Aura deployment from Postgres 17, migrate the data with pg dump / pg restore or pg upgrade ; a Postgres 18 container cannot reuse a Postgres 17 data volume directly. Aura listens on loopback; Caddy serves the cockpit on HTTPS at https://<host , behind the Authula sign-in. On a fresh install the sign-in page offers Create first user : that account is the operator, and after signing in the cockpit's first-run setup finishes the configuration. The Telegram bot is connected through the setup wizard, gated by the access token the installer prints: https://<host /setup/?token=<AURA ACCESS TOKEN Caddy uses tls internal . Browsers on other LAN machines will warn until they trust the local CA root from the caddy-data volume: docker compose exec caddy cat /data/caddy/pki/authorities/local/root.crt aura-caddy-root.crt Trust on the Ubuntu server does not make remote browsers trust that CA. Cloudflare named-tunnel hostnames use the public edge certificate; direct port 443 bypasses Cloudflare Access and retains Authula. Quick Tunnels are temporary testing only and cannot validate Aura chat because they do not support SSE. An edge appliance installed with --appliance updates itself: the aura-image-update.timer 5-minute cadence, flock-guarded pulls the moving tags and recreates only what changed, running migrations first. The aura image also carries the installation payload — the Compose files, the updater and its units, the sidecar configuration — and each tick installs whatever differs from /opt/aura backing up what it replaces under backups/payload- and brings the whole stack up on it. A version pin changed in compose.yaml therefore reaches every appliance on its own. sha256sum -c payload manifest.txt inside /opt/aura shows whether a host matches its payload. Watch it with journalctl -u aura-image-update.service -f . Manual update pinned installs, or no systemd . Volumes persist, and the aura-migrate one-shot runs the Postgres migrations before the Aura service starts ArcadeDB needs none — the MCP creates each identity's database on first use : docker compose pull docker compose up -d Scheduled backups run inside the socketless Aura box. Postgres is dumped over the Compose network with pg dump into AURA BACKUP DIR : ./backups/postgres-YYYYMMDDTHHMMSSZ.dump Memory is backed up automatically. ArcadeDB loads docker/arcadedb/backup.json https://github.com/chetto1983/Aura/blob/master/docker/arcadedb/backup.json and backs up every database every 60 minutes, including newly-created identity databases. Archives live in the separate aura-arcadedb-backups volume. The configuration sets maxFiles=60 and tiered hourly/daily/weekly/monthly retention of 24/7/4/6. The database and backup volumes are separate, but both are on the same host by default. Preserve off-host copies and the deployment configuration separately. Garage objects, workspaces, and other runtime files need their own backup policy. Run the restore drill against the current Compose stack: set -a . ./.env set +a scripts/restore drill.sh The drill tests four planes: Postgres, conversation sidecars, Garage and an ArcadeDB database shaped like a tenant. It verifies restored checksums and cleans up its disposable resources. All four passed on 2026-09-07. A separate restore of an existing scheduled operator-memory archive recovered 93 entities, 75 facts and 40 mentions, including a historical fact. This is a dated recovery check, not a complete host-loss rehearsal or an RPO/RTO guarantee. See Backup and restore https://github.com/chetto1983/Aura/blob/master/docs/BACKUP-RESTORE.md for scope and evidence. Manual restore commands: docker compose exec -T -e PGPASSWORD="$POSTGRES PASSWORD" postgres \ pg restore -U "${POSTGRES USER:-aura}" -d "${POSTGRES DB:-aura}" \ --clean --if-exists --no-owner --no-acl /backups/postgres-YYYYMMDDTHHMMSSZ.dump Take a fresh backup before restoring over a live database. The whatsapp service is part of the default stack, mounted through Aura's MCP catalog. It uses an unofficial whatsmeow-based client, so it carries WhatsApp Terms of Service and account-ban risk. First pairing is headless: docker compose logs -f whatsapp Scan the QR code shown in the logs. Aura boot never depends on this service. The host needs no Python MCP runtime at all: memory is served by Aura's own ArcadeDB MCP, a Go binary in the image. Old host-level Python installs and the earlier WSL WhatsApp MCP install can be removed after migrating to the Compose appliance. aura serve run the long-lived agent runtime channels, cockpit, scheduler aura shell | chat <sub interactive REPL / chat conversations against the agent loop aura doctor | config <sub environment diagnostics / effective configuration aura agent dry-run drive a mock LoopAgent through the Budget tree aura tools print the tool manifest aura task <sub operator parity with the model-facing task tool: schedule | list | cancel | run now | approve | runs | doctor aura mcp <sub managed MCP servers: install | add | list | doctor | tools | enable | disable | remove aura memory <sub ArcadeDB memory administration aura identity <sub identities, capability grants, operator break-glass recovery aura gateway grants <sub AG-UI gateway approval grants aura paused-states <sub HITL pauses aura skills <sub | pack <sub skill lifecycle · packs: list | show | install | trust aura retention <plan|apply retention sweep aura db <sub Postgres lifecycle: migrate | ping | status | reset aura objectstore <sub Garage object-store administration aura web <doctor|tool ... web tools search/fetch from the CLI aura docs <sub document ingestion aura version build metadata For source builds, install Go from go.mod https://github.com/chetto1983/Aura/blob/master/go.mod , Docker, and a POSIX shell. Linux is the supported source-build and quality-gate runtime. On Windows, use WSL; the native Windows Go binary is not a release target because Windows ACLs are not represented by POSIX FileMode bits. Docker Desktop remains supported for running the shipped Linux Compose appliance. git clone https://github.com/chetto1983/Aura.git cd Aura cp .env.example .env make tools lefthook install make db-migrate memory-up go run ./cmd/aura version go run ./cmd/aura agent dry-run --request-id auto To run a local source build instead of a published image, build the image and point .env at it the image builds web/ in its own stage; the committed internal/webui/dist only feeds a host go build and is refreshed from that stage, never from a host vite build : docker build -f docker/aura/Dockerfile -t aura:local . then in .env: AURA IMAGE=aura:local Quality gates: make quality make db-migrate memory-up make quality-full | Target | Does | |---|---| | make tools | install the quality toolchain | | make lint / make vet | lint and go vet | | make vuln | govulncheck supply-chain scan | | make test-race | go test -race ./... | | make coverage | owned-surface coverage floor | | make restore-drill | four-plane restore drill Postgres, sidecars, Garage, ArcadeDB | cmd/aura/ CLI entry and subcommands cmd/arcadedb-mcp/ Aura's own ArcadeDB memory MCP server cmd/aura- / sidecar binaries Cloudflare and ingest supervisors, media index, file cards internal/ the runtime, one package per concern see Architecture web/ React cockpit, embedded into the binary from internal/webui/dist services/ingest/ document ingestion sidecar packages/create-aura/ the npx installer create-aura-appliance docker/ deploy/ caddy/ image builds, systemd units and the updater, Caddy front door scripts/ install, smoke, restore drill, coverage, file-size cap docs/ architecture, capabilities, release and backup guides .planning/ GSD planning artifacts Aura is PRD-first. Persistence is Postgres plus an ArcadeDB graph, with graph access through MCP for model-facing tools. The default packaged deployment keeps Aura socketless: no Docker socket is mounted into the Aura container. The opt-in sandbox profile reaches Docker only through a socket proxy that allows the box lifecycle verbs. See CONTRIBUTING.md https://github.com/chetto1983/Aura/blob/master/CONTRIBUTING.md . Report vulnerabilities privately per SECURITY.md https://github.com/chetto1983/Aura/blob/master/SECURITY.md , never in a public issue. MIT https://github.com/chetto1983/Aura/blob/master/LICENSE Copyright 2026 Davide Marchetto.