verify-mcp
Skill19 starsChanged 49 days agoApache-2.0View on GitHub
Run only the live MCP suite (tool list, no daily quota, MCP writes that sync to a second device, has_conflict, locked and open_elsewhere errors, two agents side by side, audit log) against the real Conduit app, its MCP server and a local Appwrite. Use after changing mcp/src/, electron/ipc-server/, the MCP gatekeeper, agent session ownership, tier or quota handling, or vault lock errors, or when asked to verify MCP live. Fully isolated; takes about 1 minute.
- Reads credentials
1 other project uses a version of this.
Adopted by
Most recently taken up by
- illinigirl/claude-code-toolkittheir own version, first seen Jun 23, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- illinigirl/claude-code-toolkit★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository5 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- verify-dataSkill
- verify-ownershipSkill
- verifySkill
- verify-syncSkill
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 2 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 36
When anything under appwrite/ changed since the last deploy (stamp file .deployed-conduit-dev), it redeploys the schema, the seed rows …
- Very specific: line 81
Read the failing scenario's steps, error, <device>.mcp.log and <device>.main.log.
- Distinctive: line 87
Never make it pass by weakening an assertion, removing a check, skipping a scenario, lowering the call count or raising a timeout without a …
- Very specific: line 125
Only the local stack on localhost.
- Very specific: line 127
Never touch the user's dev profile (~/Library/Application Support/conduit/conduit-dev, "Conduit Dev"), their running dev app or port 1420, …
- Distinctive: line 130
Never answer a macOS keychain prompt.
- Very specific: line 131
Never run docker compose down -v on the local stack or delete its env file unless the user asks.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/verify-mcp: MCP suite
Runs npm run verify:mcp (node scripts/verify/run.mjs mcp). Real Electron devices unlock vaults on
a local Appwrite, and a real MCP client (the built mcp/dist server over stdio) connects to each
device's own socket. The harness reference is scripts/verify/README.md. Every suite at once is the
verify skill.
What it proves
| Scenario | In plain words |
tools-list | A Free user with an unlocked vault: tools/list has the vault tools (entry_list, entry_info, entry_search, entry_update_notes, entry_edit_notes, document_create, document_update, credential_list) and entry_list works. |
no-daily-quota | A Free user makes 60 tool calls (the old daily cap was 50), spaced under the per-minute rate limit; none is refused and no mcp-quota.json exists in the device's HOME, app data or the cloud folder. |
writes-sync | On Pro, entry_update_notes (its !!secret!! is encrypted into a hidden secret and a ref), entry_edit_notes (a one-line edit that keeps the ref), document_create and document_update show on the device (IPC and the entry tree) and reach a second device, the encrypted secret included. |
notes-secrets | The app encrypts !!secret!! into refs on save; MCP reads show refs, never values; edits keep refs, add new secrets encrypted, and refuse bad writes; a dropped ref leaves its secret flagged unused; documents do the same. An agent types a secret into a shell by ref and reads back only the ref; page scripts refuse refs. credential_read returns refs, and reveal waits for the reveal dialog (Deny refuses, Allow once returns the value). The entry page shows chips by name. |
has-conflict | A same-field edit on two Pro devices makes entry_info return has_conflict: true; after it is resolved in the review panel it is false on both devices. |
locked-and-elsewhere | A manually locked vault returns code: VAULT_LOCKED with no reason; after a Free take-over from another device it returns reason: open_elsewhere and the device shows the locked-out dialog. |
two-agents | Two MCP clients on one device are two agents. Each runs commands in its own local shell; the other is refused with SESSION_IN_USE in it but can still read it; connection_list shows owner you, other_agent and free (a shell the user opened); once agent A's MCP process exits, B can use A's old shell. |
knowledge | entry_info suggests moving heading-structured notes; kb_import_notes makes articles and keeps the notes; folder and tag-matched vault articles reach the asset in contract order; kb_write edits flag the article for review with reasons in its history; kb_log, kb_verify and kb_search work; articles stay out of entry_list and the renderer's lists and are deleted with their asset. |
knowledge-sync | On Pro, an article and its encrypted secret reach a second device; the same article edited on both at once shows one conflict (the text) in the review panel, the metadata and history settle themselves, and both devices end with the same article and a merged history. |
knowledge-team | In a team vault, an asset with a secret chip and an article upload to the server with no value in plain text; a second team member added as editor opens the vault, finds the article under its asset, decrypts the chip and lists only the asset. |
audit | Every MCP call of the run is in <HOME>/.config/conduit/audit.log in order with the right outcome, an api_key argument is logged as [REDACTED], and failed calls are logged as errors. It also passes alone (--only audit). |
Prerequisites (the harness checks or fixes these itself)
- macOS or Linux, Node 20+, Docker running,
mcp/dependencies installed (preflight fails fast with the fix otherwise). The build phase compilesmcp/intomcp/disteach run. - Local Appwrite at localhost:8085, project
conduit-dev, made once withnpm run appwrite:setup(docs/LOCAL_APPWRITE.md). The harness readsconduit-dev.envin the stack folder, refuses any endpoint that is not localhost, and starts the stack (docker compose up -d) when it is down. - When anything under
appwrite/changed since the last deploy (stamp file.deployed-conduit-dev), it redeploys the schema, the seed rows (tiers,app_config) and theapifunction. There is no SQL, no dump and no schema check: the rules come fromappwrite/src/api/in this checkout. The seed tiers keepmcp_daily_quota=-1and Freevault_max_open_devices=1. - Isolation: its own Vite on a free port (never 1420); each device in
/tmp/cv-<id>/<name>/with its own HOME (so its own audit log), app data, Chromium profile and MCP socket,--use-mock-keychain. The MCP client never touches~/.claude.jsonor the user's MCP config. Test users areverify-<runId>-<n>@conduit.local.
Run it
npm run verify:mcp
node scripts/verify/run.mjs mcp --only locked-and-elsewhere # one scenario
node scripts/verify/run.mjs mcp --only tools-list --only audit
node scripts/verify/run.mjs mcp --only audit --keep # keep /tmp/cv-<id> to read the audit logsExit code 0 means all passed including cleanup, 1 means a failure, 2 means bad arguments. Test windows appear on screen: leave them alone.
How long
About 1 minute (setup about 8 s, scenarios about 47 s; no-daily-quota alone is about 17 s because
its calls are spaced for the rate limiter). Use a 10 minute timeout. A cold Appwrite start, or a redeploy
after appwrite/ changed, can add a few minutes.
Read the results
The output ends with a table and RESULT: PASS|FAIL in <time> (.verify/<runId>/results.json).
In .verify/<runId>/:
results.json:phases[],scenarios[](status, timestampedsteps,error,failure[]with per device screenshot and log tails),cleanup.logs/<device>.mcp.log: the MCP server's stderr for that device;<device>.main.logshows the app side (IPC server, gatekeeper, vault guard).shots/: numbered PNGs;<n>-<device>-FAIL-mcp-<scenario>.pngon failure.- A tool error body looks like
{"error":"VAULT_LOCKED: ...","code":"VAULT_LOCKED","reason":"open_elsewhere"}(built inmcp/src/tool-error.ts). /tmp/cv-<id>:<id>is the last six characters of the run id (run20260926-164136-eebd6buses/tmp/cv-eebd6b); after--keep,results.jsoncleanup.keptRootsholds the exact path.
When something fails
- Read the failing scenario's
steps,error,<device>.mcp.logand<device>.main.log. Decide: app bug, harness bug or environment (see Troubleshooting). - Re-run only it:
node scripts/verify/run.mjs mcp --only <scenario>(add--keepto read/tmp/cv-<id>/<device>/home/.config/conduit/audit.log). - Fix the root cause and add a unit test (
mcp/src/__tests__/,electron/ipc-server/__tests__/) where the code allows. - Never make it pass by weakening an assertion, removing a check, skipping a scenario, lowering the call count or raising a timeout without a measured reason.
- A scenario that passes only sometimes is a real race: report it.
- Finish with the full
npm run verify:mcp. Run every suite (npm run verify) only when the user asks for it or for a release candidate.
Cleanup guarantees
On pass, fail or Ctrl+C (once; twice aborts cleanup) the run closes every MCP client, quits every
device, stops its Vite, deletes its test users, removes /tmp/cv-<id>/ and
.verify/<runId>/dist-electron/, then kills and reports any process still mentioning the run; any
leftover makes the run FAIL. Stale verify-* users older than 6 hours are deleted at the start, and
.verify/ is pruned to the 20 newest run folders before the new one is made (21 after a run). Local
Appwrite keeps running. The build rewrites the gitignored mcp/dist/; that is expected. --keep
keeps only the device roots and dist-electron/ (test users are still deleted); afterwards delete
/tmp/cv-<id> and .verify/<runId>/dist-electron yourself.
Quick check afterwards: find /tmp/ -maxdepth 1 -name 'cv-*' (the trailing slash matters on macOS,
where /tmp is a symlink) and ps ax | grep -F /tmp/cv- | grep -v grep both print nothing, unless
another run is going.
Troubleshooting
| Symptom | Cause | Fix |
Docker is not running | Docker Desktop is off | Start it, wait for docker info, rerun. |
| appwrite phase fails or the API does not answer | Stack never set up, half-started, or short of memory | If the message names a missing conduit-dev.env, run npm run appwrite:setup. Check curl -s http://localhost:8085/v1/health/version and docker ps --filter name=appwrite; run npm run appwrite:stop and rerun (the harness starts it again). Give Docker Desktop at least 4 GB of memory. |
appwrite deploy fails in the appwrite phase | Build or schema error in appwrite/ | npm run test:server, then npm run appwrite:deploy to see the error; fix the code, not the check. Deployment and execution logs are in the console (http://localhost:8085). |
| Port 8085 busy | Another Appwrite stack | lsof -nP -iTCP:8085 -sTCP:LISTEN; ask the user before stopping it. |
EINVAL or connect ENOENT on the MCP socket, or too long for the MCP socket path | Socket path over 104 bytes, or the device's app is not up yet | Device names 1 to 12 characters, roots stay at /tmp/cv-<id>/<name>. Call ctx.connectMcp(device) only after launchDevice returns: the MCP server connects to the app once at startup and keeps saying "Not connected" if the app was not there. |
Rate limit exceeded in a scenario that did not expect it | Calls faster than mcp/src/rate-limiter.ts allows | Space the calls; do not raise the limiter to pass. |
| Audit lines missing | The audit log is per HOME, written by the MCP process | Read the device's home/.config/conduit/audit.log with --keep; check mcp/src/audit.ts. |
| Page call timeout or a macOS keychain prompt | A code path reached the real keychain | Never answer it. Let cleanup kill the device or SIGKILL only that device's pid; fix the isolation in scripts/verify/lib/launcher.mjs. |
| Dependencies missing | Install step skipped | npm install, (cd mcp && npm install). |
[cleanup] FAIL or leftover process killed | A device or MCP client did not shut down | Read logs/cleanup.log; fix the quit path. |
Safety rules
- Only the local stack on localhost; never production Appwrite (
appwrite.cloudwise.ca) or any cloud project; neverCONDUIT_ENV=production. - Never touch the user's dev profile (
~/Library/Application Support/conduit/conduit-dev, "Conduit Dev"), their running dev app or port 1420,~/.claude.json, or the Conduit MCP server registered in the user's own Claude Code setup. Only the MCP clients the harness starts are used. - Never answer a macOS keychain prompt; only kill processes that belong to the run.
- Never run
docker compose down -von the local stack or delete its env file unless the user asks; never print the API key, the service secret or tokens.
Report back
One line with the result and time, then passed / total, then each failure with its root cause and
the fix or open question, and the run folder. Quote the RESULT: line as printed; never claim a pass
you did not run.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

