Evidence · bounded receipt proof

How to Test MCP Server Behavior | Reelier

Evidence state: explained against released reelier@0.32.0 on 2026-08-12.

▸ Direct answer

Record real MCP calls, compile the observed workflow, and replay declared checks against observed outcomes. This tests selected behavior beyond a static schema, but only for recorded calls and declared checks. Schema compatibility, unrecorded branches, and complete behavioral coverage remain separate questions.

Evidence note

The linked guide and recording reference document the record, compile, and replay workflow. This cluster retains no claim of total behavioral coverage.

Record representative calls

Start with tool calls that exercise the behavior you care about. The resulting workflow covers those observed calls, not unrecorded server branches.

Declare checks

Compilation turns observed calls into a fixed workflow. Assertions name the response conditions that replay should enforce; an undeclared property is not silently checked.

Replay against the server

Replay re-executes the fixed calls and evaluates the declared checks against observed outcomes. This complements schema validation but does not replace it.

Non-claims

This does not prove that an action was safe or semantically correct.

This does not prove that every write was observed or receipted.

Falsifier

A materially changed recorded tool outcome that is covered by a declared check but cannot be observed by replay would falsify this claim for that case.

Citation-friendly definition

A recorded MCP workflow tests selected calls by replaying them and evaluating declared checks; unrecorded branches and undeclared behavior remain outside its coverage.

Sources

  1. Released compiler implementation
  2. Released runner contract
  3. Test MCP servers guide
  4. MCP recording reference

Reading as an agent? This page has a markdown twin: /evidence/test-mcp-server-behavior/md