Agentic AI Comparison:
Dust AI vs Natoma MCP Platform

Dust AI: 8/10

Evidence confidence: high

Dust has a substantial documented end-to-end agent workflow: users can select models, create custom agents, provide instructions, connect company knowledge, invoke tools, and use agents for questions, drafting, analysis, and actions in connected systems. The documentation also identifies managed connections, MCP tools, public-web ingestion, uploaded folders, custom API connections, schedules, triggers, and multi-agent workflows. Dust therefore meets the rubric's substantial-workflow standard. The score is not 9 or 10 because the available evidence does not establish unusually complete coverage of every governance, deployment, testing, or operational-control requirement, and documentation of a feature does not independently establish performance or reliability.

Natoma MCP Platform: 8/10

Evidence confidence: high

Natoma documents a substantial workflow for its narrower purpose: registering or deploying MCP servers, connecting MCP applications to workspaces, routing communication through a gateway, and applying governance and access controls. The documentation identifies MCP servers as the principal application type and describes Natoma Apps as providing external data, APIs, tools, and systems to agents. The pricing materials add documented controls including agent IAM, OAuth 2.1 authentication, tool-call authorization, access policies, activity monitoring, reporting, SSO/SAML/SCIM, an MCP builder and inspector, self-hosted MCPs, and a runtime/proxy. The score is below 9 because the evidence establishes a strong MCP-management scope but does not prove exceptionally complete coverage across all possible MCP deployment, policy, observability, and lifecycle scenarios.

Dust AI: 8/10

Evidence confidence: high

Dust presents an active product with a documented getting-started path for creating and using agents, and its pricing page lists a free tier rather than only a sales-led deployment. The product is therefore accessible to intended team users, although adoption still involves workspace configuration, model and credit choices, connection authorization, and potentially administrator controls. The score is 8 rather than 9 or 10 because the evidence does not establish unusually frictionless setup for every integration or enterprise environment, and some workflows may require administrative configuration or custom API work.

Natoma MCP Platform: 7/10

Evidence confidence: medium

Natoma advertises a hosted platform with one-click deployment and intuitive setup, and its pricing page states that users can start on a Free plan without a credit card. Those facts support a current, accessible adoption path. However, the intended workflow is infrastructure- and governance-oriented: users may need to configure MCP servers, authentication, permissions, tool-call policies, clients, and possibly self-hosted components. The score is therefore below Dust's despite Natoma's stated one-click deployment, because a usable governed MCP environment generally involves more technical and administrative setup than a general agent workspace; this is a rubric judgement, not a measured setup-time claim.

Dust AI: 8/10

Evidence confidence: high

Dust publishes understandable entry pricing and allowances: Free includes 500 lifetime credits; Pro is listed at 30€/month monthly or 24€/month annually with 8,000 credits per seat per month; Max is listed at 150€/month monthly or 120€/month annually with 40,000 credits per seat per month. Dust also explains that credit consumption depends on model choice, task complexity, and tool use, and that credits measure AI usage rather than representing unlimited execution. This gives prospective users a comparatively clear purchasing model, though total cost can vary with seats, usage, connected services, and any external model or system costs; the score is not 9 or 10 because the practical cost of complex workflows is usage-dependent.

Natoma MCP Platform: 7/10

Evidence confidence: high

Natoma publishes a Free plan with concrete limits—5 MCP servers, 5 users, and 5,000 tool calls per month—and explains that tool calls are metered and that pricing is based on servers, users, and calls rather than agent count. This is useful cost information and supports a meaningful free evaluation path. However, Pro and Enterprise pricing is not fully published and is described as tailored or contact-based, so budgeting for larger deployments remains incomplete. The score reflects clear entry allowances but conservatively discounts the unresolved commercial terms for scale; it does not treat the free plan as proof that model, hosting, MCP-server, or downstream service costs are zero.

Dust AI: 8/10

Evidence confidence: high

Dust documents multiple relevant connection mechanisms: managed integrations including Google Drive, Notion, Slack, and Confluence; MCP-based tools including Slack, Salesforce, Gmail, and HubSpot; public websites; manually uploaded folders; and custom connections built through its API or scripts. Its product materials additionally state that users can connect Slack, Notion, GitHub, Drive, more than 20 additional services, or any tool through MCP, while the changelog documents Dust as a remote MCP server for MCP-capable clients. This supports a broad and extensible integration ecosystem. The score is below 9 because the evidence confirms documented options rather than verified successful operation for every named connector, and the rubric does not count internal dependencies or roadmap items as integrations.

Natoma MCP Platform: 8/10

Evidence confidence: high

Natoma's integration model is intrinsically extensible through MCP: its catalog describes Apps as MCP servers that expose external data, APIs, tools, and systems, and its platform materials describe access from ChatGPT, Claude, Cursor, and internal agents. The documented platform also includes an MCP runtime and proxy, personal and shared MCP connections, an enterprise MCP registry, desktop MCPs, BYO MCP through Docker Hub, self-hosted MCPs, and an MCP builder and inspector. These are strong, task-relevant connection mechanisms. The score is not 9 or 10 because the available evidence does not establish the breadth, maturity, or successful operation of every catalog application and client connection, and platform capabilities should not be confused with a claim that every downstream MCP server is natively supplied or supported.