MAEQ vs Haystack / deepset

OpenAI, Claude, Gemini, Mistral (EU), Azure OpenAI and Azure AI Foundry, side by side. Your own keys or usage-billed access.

EU company (Amsterdam). EU storage and vector cloud by default; US or Asia, or your own S3 or Azure Blob, on request.

Hosted SaaS today.

Agents, portals and automations run on MAEQ; there is nothing to deploy or scale.

Agents are generated from a plain-language description and automations are compiled from plain language, reviewed before saving.

The portal orchestrator routes each question to specialist agents, runs them in parallel and merges their answers.

Not yet. Agents carry an A2A-style description card, but there is no A2A endpoint.

Shared and per-person memory; portal users can see and delete what is remembered about them.

E2B, Daytona or Blaxel workspaces, with git for GitHub, GitLab, Bitbucket and Azure DevOps.

1,600+ toolkits through Composio, first-party OAuth for GitHub, Slack, Jira, Teams, Azure DevOps and Gmail, and your own MCP servers.

Claude, Cursor and other MCP clients can search the knowledge and ask the portals a scoped key allows.

Scheduled, run-now, chained and platform-event triggers; inbound webhook triggers are not available yet.

The reviewer sees the exact pending action, such as the email about to be sent, and can edit, swap or regenerate it.

SharePoint, Google Drive and Dropbox sync incrementally, following changes and deletions; S3 and Azure Blob are read in place.

Dropped folders and archives are de-duplicated, grouped by topic and proposed as knowledge bases; duplicates merge into restorable golden records.

Snowflake Cortex Analyst and Databricks Genie: the warehouse writes and runs the SQL, and MAEQ shows it.

Branded portals for staff or customers, an embeddable website assistant, and SSO with Entra ID, Okta or any OIDC provider.

25 policy types (PII, prompt injection, toxicity, quotas, IP and time limits, tool-chain control), plus Lakera Guard, Bedrock Guardrails, NeMo and Guardrails AI.

A versioned registry of your MAEQ agents with auto-written description cards; other vendors' agents are not covered.

Every call with its steps and cost; export traces to your own Langfuse. LangSmith, Arize and Helicone export are coming.

Automatically built test sets, recall@k and MRR, regression checks, AI-judge groundedness, citation checks, and Comparison Lab to test models and search settings side by side.

Users rate answers with a reason; a triage console turns each complaint into a recommended fix and where to make it.

Not yet: no simulated users or task-completion graders.

The cost of every call is recorded in a credit ledger, with token, cost and invocation quotas.

Daily discovery of new and retiring models across four providers, and a watch on deprecation pages. Guided migration with an impact preview and rollback, started by a person.

Each client gets an isolated tenant with its own branding, data and plan; a partner console across clients is not available yet.

Intake already prepares the hand-off. Next is a model of your entities, terms and relationships, built automatically from your documents and data.

Agents already combine warehouse figures and documents in one answer. Next, MAEQ decides per question which source should answer what.

Detect conflicting figures and statements across datasets, data types and local, regional and global sources; score confidence and send low-confidence cases to a reviewer.