Introducing ADAM — Agentic Data Access Model

Your AI can search the internet.
It can't query your own database.

Security blocks it. Auth doesn't work. Adam is different. It writes real queries through Eve's execution layer, uses your existing auth, and every action is traced.

Natural Langauge Data Access

AI That Lives Next to Your Data

Adam sits alongside your database endpoints as siblings in the same system, querying your data natively, governed by the same permissions as the user.

Your App
↓
ADAM
Same Auth · Same RBAC · One API
↓
Databases
↓
LLMs
↓
APIs
↓
Functions

ADAM queries all your data natively via Tools. No glue code. No separate systems. One unified access layer.

Provider-Agnostic AI

OpenAI, Anthropic, Ollama, OpenRouter — one API, swap providers without changing code. Self-host with Ollama for air-gapped deployments.

RBAC-Scoped Querying

AI queries your databases via tool calling, constrained by user permissions. No privilege escalation — the AI has exactly the same access as the user.

Custom Skills & Templates

Give your AI domain expertise with runtime-loadable skills. Version prompts, swap templates mid-conversation, no redeployment needed.

Unified Gateway

Separate API keys, separate auth, separate everything. It's exhausting.

Adam puts LLMs and databases behind one system. Same auth, same API, same place.

Multi-Provider LLM Access

Four LLM providers normalized through a single abstraction layer. Swap providers without changing a line of code.

  • OpenAI, Anthropic, Ollama, OpenRouter
  • Streaming & tool calling on all providers
  • Custom base URL support for self-hosted models

Database Proxying & Pooling

10+ endpoint types register in the same system. Connection pools for async operations, configurable per endpoint.

  • Native wire protocols , RESP, PostgreSQL, MySQL, MongoDB, ClickHouse, Oracle
  • Connection pooling with per-endpoint managers
  • PSYNC and AOF replication support for Redis

Shared Auth & RBAC

One identity system for AI and database access. JWT-based authentication with configurable expiry, organization-based multi-tenancy, and role-based access control for every endpoint.

  • No privilege escalation , AI has same permissions as user
  • Organization-scoped credential management with secure storage
  • Permission-verified on every request

Custom Templates & Skills

Define prompt templates stored in memory and toggle domain expertise on or off without redeployment.

  • Versioned prompts
  • Runtime-loadable skills
  • Swap templates mid-conversation

Credential Management

Separate secure storage for database credentials. TLS termination, automatic credential rotation, and more as standard.

  • Credentials isolated from application code
  • TLS everywhere by default
  • Automatic rotation support

Policy Enforcement

Command-level policy enforcement with three modes: observe, warn, or block. Restrict dangerous commands on managed services like ElastiCache.

  • Block ACL, CLUSTER, CLIENT, CONFIG, DEBUG on ElastiCache
  • Audit logging with service/command filtering
  • Per-connection context tracking for compliance

Provider-Agnostic AI Access

One API, every model. Zero vendor lock-in.

ProviderDefault ModelStreamingTool Calling
OpenAIgpt-5.2
Anthropicclaude-sonnet-4-5
Ollamallama3.2
OpenRoutermoonshotai/kimi-k2
AI Database Querying

Need data? File a ticket. Bug an engineer. Wait.

Adam lets anyone ask in plain English and get real answers from real databases.

How ATI Tool Calling Works

1

User sends chat request with database endpoint references

2

Adam connects to each endpoint's ATI server and discovers tools

3

Tool definitions sent to LLM alongside the conversation

4

LLM decides whether to call tools based on context

5

Adam verifies RBAC: does this user have access?

6

Tool result fed back into conversation (max 12 iterations)

Same permissions as the user. Always.

RBAC-Scoped AI Access

The AI operates under the exact same permissions as the requesting user. No privilege escalation , because the AI is authenticated through the same identity system.

  • Read-only if user is read-only
  • Invisible endpoints stay invisible
  • Same identity system for AI and data

Natural Language Queries

Ask questions about your data in plain English. AI queries your databases directly through ATI, constrained by your RBAC permissions.

  • No SQL required
  • Context-aware responses
  • Multi-database support

ATI Tool Orchestration

Up to 12 tool passes per turn. AI discovers available tools via cursor-based pagination, calls databases mid-conversation, and feeds results back into the conversation.

  • Namespace-qualified tools , {client}__{tool} for multi-tenant isolation
  • Cursor-based tool discovery with cached snapshots
  • Two tool sources: managed Adam endpoints + user-registered ATI servers

Tool Execution Metrics

Every tool call is instrumented with OpenTelemetry. Track attempts, failures, and latency per tool and per server.

  • Attempt & failure counters
  • Latency histograms
  • Per-tool observability

Conversation Persistence

Conversations stored in Postgres with two modes: persistent for long-running analysis, ephemeral for one-off queries. Full sequence tracking across sessions.

  • Persistent and ephemeral conversation modes
  • SSE streaming with heartbeat keep-alive
  • Resume conversations across sessions

Token Accounting

Prompt and completion tokens tracked per conversation, per user, per organization. Know exactly what your AI usage costs across every provider.

  • Per-conversation token tracking
  • Prompt and completion breakdowns
  • Organization-level usage aggregation
Intelligent Analytics

Slow queries and hot keys. You find out after the incident.

Adam watches your traffic and flags problems before your customers notice.

Automatic Anti-Pattern Detection

Adam watches your traffic and identifies performance anti-patterns in real time at two levels: per-connection burst analysis and per-endpoint aggregate analysis.

  • N+1 query detection
  • Hot key concentration via Space-Saving sketch (64-entry, O(1) lookup)
  • Unbounded range warnings

PII Exposure Alerts

Real-time detection of personally identifiable information flowing through your data layer. Regex-based scanning for 5+ patterns with automatic redaction before storage.

  • SSN & credit card detection
  • API key & password scanning
  • Email, phone, IP detection

30+ Rule Recommendation Engine

Not just metrics, but actionable recommendations with confidence scoring, severity classification, and estimated monthly cost savings. Covers write-heavy patterns, missing TTLs, oversized values, hot keys above 15% traffic, large hash operations, and read replica candidates.

  • Pattern & context rules
  • Confidence-scored output
  • Cost savings estimates

Real-Time Streaming Analytics

SSE-powered real-time monitoring with per-endpoint health snapshots every 60 seconds including ops/sec, error rate, and p50/p99/p999 latency percentiles. Anomaly transitions emit immediately.

  • Zero-overhead when unused
  • Anomaly-only filtering
  • 64-event broadcast buffer

Anomaly Detection

Three-stage confirmation using Jensen-Shannon divergence. Compares command distribution to baseline , watch at 2% divergence, critical at 5%. Includes latency spike and error rate detectors.

  • 3-window confirmation before alerting
  • 5-window cooldown before reset
  • Confirmation memory with 5-minute bridging windows

Adaptive Metadata Polling

Three-frequency metadata collection , High (60s), Medium (30min), Low (24h) , with four escalation levels that accelerate up to 8x during anomalies.

  • Normal → Watch → Elevated → Critical escalation
  • Hysteresis prevents flapping between states
  • ClickHouse storage with SummingMergeTree rollups

Detected Anti-Patterns

Automatically flagged with confidence scores and actionable recommendations.

N+1 Queries

5 repeated commands

Hot Keys

> 15% traffic

KEYS Command

Any occurrence

Unbounded Range

Any occurrence

Large Responses

> 64 KB

High Fanout

100+ keys in MGET/MSET

High Error Rate

> 10% rate

Missing TTLs

> 80% no TTL

Slow Queries

> 10ms

Dangerous Commands

FLUSHDB, SHUTDOWN, CONFIG

Large Hash Ops

> 5 KB per operation

Write-Heavy Never-Read

0 reads on pattern

Explore Request Analysis
Eden Chat

Leave terminal. Open browser. Copy-paste. Lose your thought.

Adam lives in your terminal. Query databases, load skills — without switching.

Features

  • Multi-provider support: Anthropic, OpenAI, Ollama, OpenRouter
  • Auto-endpoint discovery from your Adam deployment
  • Database querying via ATI with the same RBAC as your user account
  • Real-time SSE streaming with heartbeat keep-alive and token-by-token rendering
  • Conversation persistence: persistent mode for long-running analysis, ephemeral for one-off queries
  • Skill system: toggle domain expertise on/off at runtime
  • Tool approval gates for safe AI execution with server-side polling
  • Model override: switch providers and models mid-conversation

Built-in Tools

eden_curl

Make authenticated API calls to Eden service from within AI conversation

load_skill

Dynamically load and enable a skill at runtime

show_form

Display interactive forms dynamically within the terminal

Sign in to try Eden Chat directly in your browser, connected to your real data with your real permissions.

Under the Hood

Framework

Built on ratatui, a real terminal application, not a web wrapper

Architecture

Elm Architecture (TEA) for predictable state management

Screens

Login → LLM Provider Select → Chat, full UX flow

Tool Execution

Server-side execution with polling, up to 12 tool passes per turn

See what your team has been missing.

Everyone who sees Adam wants it for a different reason. The common thread: they can't reach their own data.