Model Context Protocol (MCP) is a lightweight way to expose tools, resources, and prompts to any compliant agent or bot. Atlas Agent treats MCP as the shared control plane so every chat surface can invoke the same trusted tools with identity, consent, provenance, and tracker guard baked in.
Origins of the protocol
OpenAI introduced MCP to decouple model providers from tool providers. Instead of every assistant defining bespoke tool hooks, MCP gives us a portable schema for catalogs, auth, and telemetry.
Atlas adopted the spec early because it aligns with our mission: furnish broker-led teams with transparent tools, clear scopes, and reusable consent. We extend the base spec with policy stores, budgets, and provenance fields so regulated teams can pass audits.
Core primitives
An MCP server exposes three primitives: tools, resources, and prompts. Tools are executable actions with structured inputs/outputs. Resources are files or documents clients can fetch. Prompts are reusable instructions clients can load to keep assistants consistent.
Atlas adds metadata such as cost hints, latency hints, and consent scope names to each tool entry. Clients can decide whether to call a tool based on price, data access, or compliance requirements—before they spend a token.
- Tools: e.g., “Generate sales comps PDF” with property, radius, and time-window arguments.
- Resources: e.g., signed comps artifacts or MLS reference sheets.
- Prompts: e.g., buyer intake flows or listing prep checklists.
Why Atlas layers governance on MCP
MCP focuses on the transport. Atlas adds the controls: identity, consent, tracker guard, provenance, and the marketplace. This makes MCP safe for industries that need to prove who approved what, which vendor executed the call, and how much it cost.
Every MCP event we emit includes the fields listed in the Knowledge Base usage schema. That lets finance, compliance, and ops share the same JSON without complex joins.
- Tool library (declared inputs/outputs/scopes; cost & latency hints; rate caps)
- Tracker & guard (preflight checks; postflight signed events; budgets & alerts; cross-vendor protection)
- Identity & consent (explicit, time-bound scopes; approvals logged; brokers can prove who approved what and when)
- Provenance (chain-of-custody tags linking to exact tool calls)
- Local booster (run light checks on-device; escalate heavy work to cloud to save tokens/time)
- Marketplace (partner tools with simple spec; revenue share)
Where MCP fits in your stack
Treat the MCP server as the contract between your assistant and every downstream vendor. Whether you run a Custom GPT, VS Code agent, or Slack bot, they all talk to the same server URL and inherit the same budgets and scopes.
When you add a new tool to the server, every client can discover it without code changes. When you revoke a scope, every client respects it immediately.
- Custom GPT or team GPT
- Web chat on brokerage site
- Slack or Teams for staff
- CRM chat (e.g., Follow Up Boss)
Next steps
Use the Atlas Server Builder to spin up your first MCP instance, then follow the Python or VS Code tutorials to add more tools. The Pricing page outlines when to move from Server Builder to Bot Starter Kits or Hosted Atlas Bot seats.