Connectors
Connect your agent the way it already works: MCP, REST/OpenAPI, or document and email intake.
Integrations
Connect your agent over MCP. No rewrite.
Agents connect to AutoStack environments through the Model Context Protocol (MCP): the environment serves its tools from an MCP server backed by a real database. In an engagement we also mirror your agent's existing tools over REST. An agent that can point at a different endpoint runs unchanged. Agents that act only through a browser UI aren't supported yet.
Connection
From your agent to a seeded database.
tools from a real episode · 8 of 15
readclaims_getauthority_getpolicy_getmanual_search
writecoverage_recordreserves_setpayments_issueclaims_escalate
MCP · live
Tools over MCP
- serverAn MCP server per environmentThe environment serves its tools from an MCP server backed by a real database, not canned responses.
- tools15 tools, typed6 read tools and 9 write tools, each with a schema and description your agent can discover.
- stateWrites change stateEvery write is recorded in the audit log and checked by the grader, including calls the system rejects.
- resetReset between episodesThe database is rebuilt from a seed before every episode, in about 6 ms.
REST · in an engagement
What if our agent calls REST APIs, not MCP?
In an engagement we mirror the operations your agent already calls, up to 3 in a feasibility week, behind the same paths and payloads, backed by the environment's database. Your agent keeps its client code; only the base URL changes.
base_url = mirror\n→ same paths\n→ same payloads\n→ runs unchanged
tools/list\n→ 6 read\n→ 9 write\n→ seeded database
Frameworks
Agent frameworks that can connect
Any framework with an MCP client can call the environment's tools, and any agent that calls HTTP APIs can use a REST mirror. We haven't tested every framework, so tell us yours and we'll confirm before an engagement starts.
What we need
What we need from you to connect
- 01Your agent's tool interfaceTool schemas and descriptions, as your agent sees them today.
- 02Your workflow rulesThe operating rules the work must follow. We write tasks and graders against them.
- 03An endpoint switchA way to point your agent at a different endpoint. If it has one, it runs unchanged.
- --Not needed: production dataEverything in the environment is synthetic.
Limits
Not supported yet
Agents that act only through a browser or desktop UI. We test agents that act through tools or APIs.
In development
Coming next for integrations
Run in your own tools
Export tasks to Harbor, Inspect and OpenEnv, including for RL training.
Runs on Microsoft's open-source ThinkingBox harness
Not affiliated with or endorsed by Microsoft.
Full roadmap → Roadmap items are plans, not commitments.
FAQ
Questions about connecting.
Which agent frameworks work?
Any framework with an MCP client can call the environment's tools. We haven't tested every framework, so tell us yours and we'll confirm before an engagement starts.
Do you support agents that use a browser?
Not yet. We test agents that act through tools or APIs; UI-only agents aren't supported.
Do we have to change our agent's code?
Usually no. If your agent can point at a different endpoint, it runs unchanged against the mirrored tools.