Tech specs

• Linux, on a physical server or virtual machine
• x86_64 and ARM64 architectures

Windows is not supported. macOS for development only.

• Docker Engine with Compose v2 (guided installation)
• Podman, partially supported, without guided installation

Ten containerized services: proxy, web, API, worker, PostgreSQL, Redis, S3-compatible object storage, Milvus, etcd and an isolated Docker guard.

PostgreSQL, Redis, S3-compatible storage and Milvus or Zilliz can be bundled with the installation or be services your organization already runs.

One node per installation. Several isolated installations can share a server, each with its own containers and network.

• HTTP proxy on one port, 8080 by default
• TLS terminated at your organization's load balancer or reverse proxy
• Works behind corporate proxies and TLS inspection once your CA is in the server's trust store

• Online
• Fully offline, on an isolated network, with a license file

• 9 languages, with the most complete translations in Portuguese, English and Spanish
• Light and dark themes
• Your organization's logo, name and color on the sign-in screen and menu

• Minimum: 4 GiB
• Recommended: 8 GiB

• Minimum: 10 GiB
• Recommended: 50 GiB

Runs the platform and local models on CPU alone, on x86_64 or ARM64. The machine check reports the largest catalog model that fits.

• NVIDIA, through CUDA, with the driver and the NVIDIA Container Toolkit
• Intel integrated or discrete, through OpenVINO

Without the NVIDIA Container Toolkit, the platform starts on the CPU instead of failing.

Inference on NPU.

The system panel shows CPU, memory, disks and GPUs (NVIDIA, AMD, Intel and Apple), with video memory.

Before deploying a model, the platform checks the runtime, the hardware and the available memory and refuses incompatible combinations.

OpenAI, Azure OpenAI, Anthropic and any endpoint compatible with the OpenAI API.

Encrypted keys and a connection test.

vLLM, llama.cpp, OpenVINO Model Server, TensorRT-LLM, SGLang and IPEX-LLM, in isolated containers, installed, started, stopped, restarted and removed from the panel.

IPEX-LLM is still being validated.

Running Ollama, vLLM and LM Studio servers are detected and connected as compatible endpoints.

More than 1,000 open models, in a curated catalog filterable by modality, family, quantization and runtime, with a fit estimate for the machine.

• GGUF
• OpenVINO IR
• safetensors (Hugging Face)

Each format is offered only to the runtimes that can serve it.

Q4_K_M, Q5_K_M, Q8_0, FP16, INT8 and INT4

Global models, enabled per workspace, with default models for embeddings, reranking, vision and transcription.

Local or cloud base model, system instructions, temperature, max tokens, repeat penalty, top-p and top-k. Draft, published or archived, with access per workspace and per person.

• 41 file extensions
• Text: txt, md, csv, tsv, json, xml, html and log
• PDF and Office: pdf, docx, xlsx and pptx
• Images, in 10 extensions, described by a vision model
• Audio, in 9 extensions, and video, in 8, transcribed

Hybrid: vector (Milvus) and keyword (PostgreSQL), combined with Reciprocal Rank Fusion and reordered by a reranking model.

Chunks of 900 characters with a 150-character overlap.

Served locally by models compatible with the OpenAI API. Each context pins its own model and can be reindexed.

• Types: text, integer, decimal, boolean, date, timestamp and JSON
• Primary key, searchable columns and sensitivity labels
• CSV, TSV or XLSX loads up to 100 MB, appended or replaced, with rejected rows and the reason

Read-only, on a separate database role: one SELECT statement at a time, with a timeout and up to 1,000 rows. Can be turned off per context.

Documents, images, audio and video, up to 100 MB per file, with types and limits set by the administrator.

.xlsx, .pptx, .pdf, .csv, .md and .txt

Optional, per message: DuckDuckGo (default), Tavily or Brave.

The browser's speech recognition or a transcription model.

Visual, with 13 block types: Input, Decision, Agent, AI model, Integration, Storage search, Data query, Data write, AI over rows, Load files, Transform, Sub-flow and Output.

JSON Schema with closed enums and a configurable reasoning effort in AI blocks.

Up to 50 rows per call and up to 8 concurrent calls.

Up to 8 levels of nesting.

Draft, published and archived, with an input and output contract in JSON Schema.

• From chat
• Through an HTTP endpoint, with each flow's own token
• On a recurring interval of 60 seconds or more, with a fixed input, a start time and a maximum number of runs

Every run records the input, the sources consulted with their relevance, the decisions and the output of each step.

30, across 6 areas: credit and risk (8), documents and back office (7), customer service (5), compliance and legal (4), HR (3) and sales (3).

Flow export and import as JSON.

• More than 99. Examples by category:
• E-commerce: Amazon Selling Partner, Mercado Livre, Shopify, Meta Product Catalog and TikTok Shop
• Messaging: WhatsApp Cloud API, Telegram, Twilio and Slack
• Productivity: Microsoft 365 and Google Workspace
• ERP: TOTVS Protheus
• CRM: Salesforce
• Database: SAP HANA
• DevOps: Jira and Confluence
• ITOM: HPE OpsRamp
• Backup: Veeam
• Security: Netskope
• Market data: AwesomeAPI

• REST and GraphQL, executed in flows
• Other types, such as database, file transfer and gRPC, are registered with a connectivity test

API key, bearer, basic, OAuth 2.0 client credentials and OAuth 2.0 refresh token with automatic renewal, plus a no-authentication option.

Credentials are encrypted and decrypted only at run time.

HTTP endpoint (POST) per published flow, with its own bearer token, shown once, stored as a hash and rotatable, and a list of allowed origins (CORS).

Outbound calls validated against SSRF, without following redirects.

OIDC (Authorization Code with PKCE) with Microsoft Entra ID, Google Workspace or any OIDC provider, such as Keycloak, Okta and Auth0.

MFA through the identity provider. Optional SSO-only sign-in policy.

Directory groups grant administration, modules and workspaces.

Owner (root), administrator and member.

Granted per person: Chat, Agents, Storage, Data, Flows, Schedules and Integrations. New users start with none.

A personal workspace for each user and shared workspaces, with members, models and agents enabled per workspace.

• Passwords of at least 12 characters, hashed with Argon2
• HttpOnly, SameSite=Strict session cookies

After 7 failures from one source or 30 on one account, for 15 minutes.

A password-only root account, never federated, recoverable on the server with a single-use key verified offline.

Sign-ins and failures, users and workspaces, SSO and access policy, providers and secrets, license, root account, data queries and loads, settings with before and after values, and every deletion.

Excluded from automatic purges and cannot be edited or deleted through the API.

Every chat, agent and flow interaction, with content, model, workspace, user, tokens, latency and rating, with filters.

The model and sources of every answer: documents with their relevance, or the SQL query with its row count. Indicators for answers with sources and for indexed documents.

Irregularity register with action plans, a semiannual report and a machine-readable data dictionary, aligned with CMN/BCB Joint Resolution No. 18/2025.

Optional: purge of flow runs, messages and data loads after the period you set. Without a policy, data is kept. Every deletion is recorded.

Conversations, messages, input and output tokens and ratings, in total and per person.

Structured JSON logs, /health and /ready checks and export of traces and logs over OpenTelemetry (OTLP) to your organization's collector.

XChaCha20-Poly1305 for keys and credentials. Each installation generates its own secrets and service passwords.

The API has no direct access to the server's Docker. An isolated, network-less guard allows only model runtime operations, with no privileged containers, host mounts or host network.

Static binary for Linux x86_64 and ARM64, including RHEL 8, Ubuntu 20.04 and Debian 11, in Portuguese and English.

aptabit check: system, memory, disk, Docker and Compose, GPUs and NPUs, network, clock and proxy, with JSON output and an offline mode.

aptabit install: checks, configures and starts the platform. Running it again updates the existing installation and never deletes data.

aptabit update: a backup first, a staged update behind health checks and an automatic rollback if the new version fails. Release channels and version pinning.

Local snapshots of the configuration and the installation volumes, with listing, rollback and cleanup.

The platform pauses briefly during the snapshot. External services are not included.

JSON status, diagnostics and a support bundle without passwords, keys or tokens. Every error has a stable code.

Applied automatically when the platform starts.

Private registry with short-lived, license-bound credentials and a signed provenance statement in every product image.

• Only for flows published to production
• No per-user, per-token, storage or usage fees

Per customer, digitally signed (Ed25519), with a key pair generated on each installation.

A check at least every 6 hours. Without contact, the platform keeps running for up to 7 days, a default period set in each license.

License file valid for up to 120 days, renewed the same way. The installation never contacts Aptabit.

License and installation metadata. Depending on the license's telemetry level, also aggregate usage counts, CLI failure reports without passwords, keys or tokens and, at the full level, a hardware and environment inventory. No documents, conversations, prompts or files.

A 7-day grace period by default. An expired or suspended license blocks access but never deletes data or containers.