verifyax-mcp 0.3.5

verifyax-mcp 0.3.5

Security

Catch up to @verifyax/mcp-server 0.3.4, which this plugin had never picked up. The plugin stayed pinned to server 0.3.3 while 0.3.4 shipped a security hotfix: keyed HMAC-SHA-256 API-key fingerprints replacing synchronous PBKDF2 in the HTTP transport, a direct-peer session-initialization limiter, and patched transitive dependencies. Anyone on plugin 0.3.3 was running below that fix — update.

Added

• Tracked the server 0.3.5 release: MCP Tasks for the long-running generate_scenario and evaluate_agent tools. Task-capable clients get a pollable handle immediately instead of holding tools/call open for the whole run (evaluations can take ~30 minutes); other clients keep the existing blocking behaviour. tasks/cancel maps to the VerifyAX job and simulation cancel APIs. Plugin 0.3.5 ↔ server 0.3.5.

Changed

• scripts/check-manifests.mjs now fails CI when a plugin's pinned MCP server version doesn't match the plugin's own version, or when a server is launched unpinned — the drift that let 0.3.3 sit below the 0.3.4 fix is now caught before merge.

Note on where this runs

This plugin launches a local npx MCP server, so it works in Claude Code but not in Claude.ai chat or Cowork, where plugins with local MCP servers don't load. On Enterprise plans, admin policy may also restrict marketplaces or disable local MCP servers.

Install/update: /plugin marketplace update verifyax-plugins then /plugin install verifyax-mcp@verifyax-plugins

Full changelog: CHANGELOG.md

verifyax-claude-agent 0.1.2

verifyax-claude-agent 0.1.2

Security hardening and test coverage for the adapter that exposes your own Claude Code agent over A2A.

Security

• Replace the static tools-off deny-list with Claude Code's fail-closed --tools "" capability, and test it against the exact pinned CLI.
• Pin cloudflared 2026.8.3 with source-controlled SHA-256 values for every supported platform; custom versions now require an explicit checksum.
• Pin the sandbox base image by digest, Claude Code by exact version, and the complete Python dependency graph with hashes.
• Make the tools-on launcher deny networking by default, require an explicitly marked restricted network, mount the project read-only, and provide a dedicated ephemeral scratch volume.
• Pin the sandbox agent user to UID/GID 1000 (reclaiming the base image node account) so the launcher tmpfs mounts are writable.

Testing

• Add executable-plugin CI covering app imports, bearer auth, forwarded-host validation, subprocess flags and stdin, timeout cleanup, context eviction, tools-on gating, tunnel integrity, and the sandbox definition.
• Pin CI actions by commit SHA and enable dependency update automation.

Note on where this runs

This plugin needs local Python, the claude CLI, and a tunnel, so it works in Claude Code but not in Claude.ai chat or Cowork.

Install/update: /plugin marketplace update verifyax-plugins then /plugin install verifyax-claude-agent@verifyax-plugins (auto-installs verifyax-api)

Full changelog: CHANGELOG.md

verifyax-api 0.3.0

verifyax-api 0.3.0

Rework the skill around the canonical OpenAPI contract as the single source of truth. The API surface is no longer transcribed into the skill (where it drifted); it is fetched from console.verifyax.com/openapi.yaml on demand.

Changed

• SKILL.md shrinks ~476 → ~130 lines. It keeps the workflow and behavioural rules a spec can't express — async 201-then-FAILED, async tag compatibility, open enums, error-status-is-truth, rate limits, base path — and instructs the agent to download and grep openapi.yaml for exact endpoint shapes. The canonical human-readable companion is served at console.verifyax.com/SKILL.md.
• Add explicit secret-handling guidance inside the skill: read the key from VERIFYAX_API_KEY, never inline, log, or commit it.
• Add a robust polling example with a deadline and backoff.
• Tighten the trigger description; note that the one-time-login token is a live credential.

Install/update: /plugin marketplace update verifyax-plugins then /plugin install verifyax-api@verifyax-plugins

Claude.ai: download verifyax-api.skill below and upload it via Settings → Capabilities → Skills.

Full changelog: CHANGELOG.md

verifyax-claude-agent 0.1.1

Security + robustness fixes from code review (Cursor Bugbot + multi-agent review).

Fixed

• tools-off now truly disables tools — was --allowedTools "" (auto-approval only, agent could still run shell); now removes tool availability via --disallowedTools + --strict-mcp-config (verified: shell exec blocked). A genuine no-host-access mode.
• Prompt is fed on stdin, not argv — turn text starting with - was parsed as CLI options (broke the turn) and was a flag-injection surface into the same argv as the security flags; now over stdin so untrusted text can never be interpreted as flags.
• Timeouts kill the whole claude process tree — new process group + killpg / taskkill /T, done asynchronously (Windows taskkill no longer blocks the event loop); no orphaned/zombie processes under --pids-limit.
• Bearer check no longer 500s on a non-ASCII token (bytes compare + guard → clean 401).
• Tunnel checksum/version pin (CLOUDFLARED_SHA256 / CLOUDFLARED_VERSION) enforced on PATH and cached binaries too, not only fresh downloads; verifies the installed binary consistently; archive members validated as regular files before extraction.
• CLAUDE_TURN_TIMEOUT env now wires through the create_app factory (was fixed at 240s); parsed defensively (bad/non-positive → 240s).
• Sandbox uses an ephemeral tmpfs home — writable so claude --resume works under --read-only, but discarded with the container so no state (or adversarial write) persists across runs; auth is per-run.
• Session id stored before the error check (resume survives an error turn); per-context caches bounded (_MAX_CONTEXTS=512) with no lock/--resume eviction race; derive_base re-validates the host on the fallback path.

Install: /plugin marketplace add verifyax/verifyax-plugins-claude then /plugin install verifyax-claude-agent@verifyax-plugins.

verifyax-api 0.2.0

verifyax-api 0.2.0

Expanded the API reference to match the current VerifyAX gateway surface. Additive and
backward-compatible — no breaking changes to existing workflows.

Added

• Direct Line (Copilot Studio) agents — agent_type: DIRECTLINE, with
  agent_parameters.directline { secret, region }, region→URL mapping, and the
  api-agent-test-directline probe.
• MCP agents — agent_type: MCP, with agent_parameters.mcp { url, auth_method, token, transport, enabled_tools } and the mcp-connection discovery/probe endpoint.
• Extra connectivity probes: a2a-connection and a2a-message.
• POST /v1/scenarios/generate-from-qna for inline Q&A interview scenarios.
• Simulation runs: batch scenario_uuids run groups and per-run timeout_minutes.
• Evaluation shortcuts (…/evaluation, …/evaluation/scores, batch …/scores) and
  structured JSON run output (…/output).
• New sections: GET /v1/billing/balance, audit logs (GET /v1/logs), and
  POST /v1/client-tags/register-qna.

Changed

• Skill-tag discovery moved from the browser-session /web/api/v1/tags route to the
  public GET /api/v1/tags (Bearer key). Response is now a bare JSON array, and the per-tag
  flag client_specific is renamed custom.
• generate / generate-from-qna forward only documented public fields — internal
  engine/model/DAG knobs are stripped at the gateway.
• Run timeout (timeout_minutes) is now set on POST /v1/engine/simulate/scenario.
• agent_type documented as the open set A2A | API | DIRECTLINE | EXTENSION | MCP.

Install/update: /plugin marketplace update verifyax-plugins then
/plugin install verifyax-api@verifyax-plugins

Full changelog: CHANGELOG.md

v0.1.0 — Initial release

First public release of the VerifyAX Claude plugin and skill.

What's included

• verifyax-api — drives the VerifyAX REST API end-to-end: register agents (A2A or REST), discover skill tags, generate scenarios, trigger simulations, poll async jobs, and fetch evaluation results.

Install (Claude Code)

/plugin marketplace add verifyax/verifyax-plugins-claude
/plugin install verifyax-api@verifyax-plugins

Install (Claude.ai)

Download verifyax-api.skill from the assets below and upload it via Settings → Capabilities → Skills.

What's covered

• Full agent registration flow with connectivity pre-checks
• Skill-tag discovery via /web/api/v1/tags with scenario-type compatibility rules
• Async job polling with the standard PENDING → PROCESSING → COMPLETED|FAILED|CANCELLED lifecycle
• Credit estimation before triggering runs
• Simulation runs (single and batch) with optional auto-evaluation
• Evaluation result and usage-event retrieval

Requirements

• A VerifyAX API key (VERIFYAX_API_KEY env var) — get one from Settings → API Keys in the VerifyAX console.