01
MCP Inspector
Discover tools, inspect schemas, and confirm that a server behaves on the happy path.
Local-first MCP reliability testing · v0.3.1
Make recovery
reproducible.
MCP Inspector shows what a server does. ResiliReplay proves what happens when it fails, whether it recovers safely, and whether that recovery remains fixed.
- No API key
- No telemetry
- No hosted backend
Start in one command
Install the public CLI.
Node.js 22 or 24. The verified local workflow needs no account or provider.
npx --yes resilireplay@0.3.1 --versionVerified 60-second workflow
From reviewed target to executable regression.
This fixture-backed Studio capture is real. The measured workflow completed in 2.4 seconds and the GIF is assembled from six browser-driven frames.
Complementary tools
Inspect first. Break safely second.
02
ResiliReplay
Inject bounded faults, observe recovery, approve evidence, and export a regression.
ResiliReplay is not endorsed by MCP Inspector or the tested projects.
Campaign workflow
A reviewable path from config to CI.
-
01
Review
Import an Inspector-shaped config and dry-run a value-free execution plan.
-
02
Declare
Pin the target, tool allowlist, seed, budgets, faults, and expected outcomes.
-
03
Run
Execute a clean control, bounded fault, and recovery path through the real server.
-
04
Compare
Approve complete evidence and fail closed on later baseline differences.
-
05
Keep fixed
Generate and execute the minimized causal regression in Node or CI.
Independent field validation
Three public servers. Three reproducible cases.
Each run used a pinned public package, one read-only tool, and public ResiliReplay v0.3.0.
LF Projects · stdio
MCP Everything
Echo control, one recovered tool error, and one declared canary failure.
- 2.7 s
- 3 / 3 expected
- 0 baseline differences
Microsoft · stdio
Playwright MCP
Isolated blank-page snapshot with no navigation, profile, or remote target.
- 4.8 s
- 3 / 3 expected
- 0 baseline differences
UI5 · stdio
UI5 MCP Server
Bundled guidelines only—no project inspection, linting, generation, or writes.
- 4.6 s
- 3 / 3 expected
- 0 baseline differences
Authorization before automation
Fault injection needs a hard boundary.
Discovery calls no tool. Tool campaigns require an explicit allowlist and the exact hash of the reviewed campaign. Remote HTTP requires separate ownership confirmation.
Read the security model-
Local by default
Studio binds only to 127.0.0.1 and has no cloud account.
-
Contained evidence
Reports are local and credential-shaped values are redacted before persistence.
-
Honest claims
A matched campaign is bounded reliability evidence, never universal certification.
-
Your authority
Commands are not OS-sandboxed. Test only targets you own or are authorized to audit.
Keep recovery fixed
Put the baseline in GitHub Actions.
Discovery-only campaigns need no tool confirmation. Review tool-calling hashes in code.
- uses: aliengineering-byte/resilireplay@v0.3.1
with:
campaign: reliability.campaign.yml
- run: npx --yes resilireplay@0.3.1 campaign compare runs/latest \
--baseline baselines/reliability.jsonDocumentation
Read the contract before the campaign.
01
02
03
04
05
06
Field-test guide
Your server, about five minutes.
Campaign schema
Seeds, budgets, faults, assertions.
Inspector import
Supported config and transport subset.
Studio security
Loopback, session, CSRF, containment.
Known limitations
What v0.3.0 does not prove.
Contribute
Bring a sanitized reproduction.
Honest limits
Evidence has edges.
- Commands and servers are not OS-sandboxed.
- Pattern redaction cannot prove every application secret is gone.
- Hashes prove integrity, not signer identity.
- v0.3.0 does not model incremental streaming or interactive OAuth.
- The three case studies cover bounded tools, not whole projects.
Bring a real server
Can another developer reproduce your recovery?
Run the five-minute guide, sanitize the evidence, and share what did—or did not—recover.