Sentra for agents: the MCP server and API
Claude, ChatGPT, Cursor and internal agents reach the same company memory as people do, through the Sentra MCP server and REST API, with the same permissions applied to every call.
In short
- MCP server at https://api.sentra.app/mcp/ over Streamable HTTP, with API keys or OAuth.
- Agents get extracted facts about people, accounts and projects, not raw text, plus a research agent that returns a written answer with citations.
- Every public REST endpoint has a matching MCP tool, and tools a caller is not entitled to are hidden from that caller.
- On Terminal-Bench 2.1, a coding agent with Sentra scored 88.31% against an 83.37% baseline, at 72.6% lower model cost with 41.2% fewer tokens.
Why not pipe company data into Claude yourself?
Engineering teams often ask how this differs from connecting their databases and tools to a model through MCP. The difference is what the model receives. A pipeline hands the model raw documents and messages, and the model has to work out on every call who is who, what is current and what conflicts. That costs tokens and gives different answers each time.
Sentra does that work once, when the data arrives. It resolves names to entities, keeps the source and time on every fact, and applies each source's permissions. Agents then ask for the facts about an account or a person and get a short, cited, current answer.
What the MCP server exposes
Tools are grouped by what they read. A caller only sees the tools it is entitled to.
- Search and fetch across meetings and entities
- Entities, entity types, the facts about an entity and its relationships
- Meetings, with summary and transcript
- A research agent: start a run with a natural-language question and get back a written answer with citations
- Playbooks, connected sources, stalled threads and follow-up load
- Project views, with tools to create, update and resolve items
- Email threads and raw email bodies, only with a separate email access permission
Authentication and limits
Agents authenticate with a Sentra API key as a bearer token or through OAuth. Keys issued through OAuth integrations cannot write. Each key is limited to 120 requests a minute. Search returns at most 20 results per call and a meeting fetch is capped, so one call cannot flood an agent's context.
Better and cheaper agents
Because agents receive distilled facts rather than raw text, they spend fewer tokens finding context and more on the task. On Terminal-Bench 2.1, across 445 trials, a coding agent with Sentra reached 88.31% mean reward against an 83.37% published baseline, at 72.6% lower model cost and with 41.2% fewer tokens.
Frequently asked questions
Does Sentra have an MCP server?
Yes, at https://api.sentra.app/mcp/ (the trailing slash is required). It works with Claude, Claude Code, ChatGPT, Codex, Cursor, VS Code, Windsurf and Zed.
Do agents see everything in Sentra?
No. An agent acting for a person sees only what that person may see, under the same source permissions, and tools it is not entitled to are hidden.
How does Sentra reduce LLM token costs?
Agents receive extracted, sourced facts instead of raw documents and messages, and results are capped per call. On Terminal-Bench 2.1 that meant 41.2% fewer tokens and 72.6% lower model cost with a higher score.
Is there a REST API as well?
Yes. The REST API under /external/v1 covers meetings, email threads, sources, entities, facts, relationships and research runs, with an OpenAPI spec at https://api.sentra.app/external/v1/openapi.json.