BEFORE
The happy path works.
Clean tool call ✓
Recovery behavior ?
Duplicate effects ?
Regression missing
v0.7.0 · MCP first · local by default
Break the MCP call.
Prove the recovery.
ResiliReplay injects deterministic MCP failures, verifies bounded recovery, and turns failures into executable regression tests.
- No account or API key
- No telemetry
- No remote target in the demo
Ten-second local workflow
One command. One result.
Node.js 22 or 24. No checkout, pnpm, config, account, credential, or paid model.
npx --yes resilireplay@latest mcp demoExact packed-package output
Clean call → fault → one retry → regression pass.
The bundled fixture contacts no network target. Temporary state is removed after the generated regression executes; only the final SHA-256 evidence identity is printed.
Before and after
Replace recovery guesses with executable evidence.
AFTER RESILIREPLAY
The failure path is reviewable.
Deterministic failure ✓
Recovery bounded ✓
Duplicate effects 0
Regression executed ✓
Three jobs
Focused MCP reliability testing.
01
Inject
Reproduce a declared MCP fault deterministically at a reviewable boundary.
02
Verify
Bound recovery and record duplicate-effect behavior under the approved policy.
03
Regress
Compile causal failure evidence into an executable test and run it in isolation.
Your reviewed MCP server
Plan first. Execute the exact digest.
Dry-run starts nothing and writes nothing. It prints the selected transport, tool, fault, retry/time bounds, and SHA-256 approval identity.
npx --yes resilireplay@latest mcp test --config ./mcp.json --server my-server --tool echo --safety inert --dry-runReal public MCP subject
Official Everything server. Packed ResiliReplay client.
Cross-platform CI installs `@modelcontextprotocol/server-everything@2026.8.18` and the packed CLI in a clean npm project. A real SDK 1.30.0 client discovers and invokes inert `echo` over local stdio, injects `mcp-tool-error`, retries once, generates the regression, verifies cleanup, and scans evidence for credentials and private paths.
Pinned
Exact version, npm integrity, SDK runtime, protocol revision, and operation.
Bounded
One inert operation, one deterministic fault, one approved retry, finite timeouts.
Honest
Product-owned validation, not an independent adopter or vendor endorsement.
MCP-only support
Evidence labels describe the boundary.
| Surface | Evidence | Boundary |
|---|---|---|
| Bundled fixture | Fixture verified | Local, deterministic, zero network |
| Inspector stdio | Live verified | Real SDK process, reviewed tool, bounded recovery |
| Streamable HTTP | Live verified | Authenticated loopback fixture; explicit remote ownership |
| SSE | Protocol verified | Supported SDK transport and config import |
| Everything 2026.8.18 | Install + live verified | Pinned local stdio, inert echo, packed CLI |
Supporting credibility · MCP-RES v0.2
Portable reliability evidence, independent of one tool.
MCP-RES is the project-defined open evidence standard behind the result vocabulary. ResiliReplay is a reference implementation, not a required dependency. MCP-RES is independent of the official MCP specification and is not an official MCP standard, security certification, or endorsement.
Safety and scope
Keep the target, effects, and evidence reviewable.
Explicit authority
One selected config entry, one allowlisted tool, and an exact plan digest.
Metadata-only persistence
Tool bodies and arguments become bounded hashes before artifacts are written.
Fail closed
Path escape, mismatched overwrite, missing approval, secrets, and exhausted bounds fail.
ResiliReplay is not a sandbox or authorization layer. Commands run with your OS authority. Security policy · limitations
Secondary workflow
Agent-runtime capture remains available.
Genuine local/no-key evidence exists for LangGraph and OpenAI Agents SDK. Claude Code and Codex are installation- and fixture-verified; Hermes is installation-verified. Authenticated hosted models, billed calls, production APIs, and vendor endorsement are not claimed.