An Adversarial Quality Gate for AI-assisted IDEs: the IDE agent and a Critic LLM debate code; the human decides only on deadlock.
Quick start (npm / npx)
1. Choose a Critic provider
Use a direct API provider with its key, or use a local CLI that is already installed and signed in. Local CLI providers do not need a separate provider API key. If is omitted, Vibe-Gate selects the first installed local CLI in this order: Codex, Claude Code, Cursor Agent, then OpenCode CLI. OpenCode CLI moves to the front when is in form. The first review response includes a notice naming the selected CLI. If none is found, OpenAI remains the default and requires . Set to choose explicitly.
Adversarial Quality Gate MCP for vibe-coding: IDE AI vs Critic AI, human decides on deadlock
# Direct API example
CRITIC_PROVIDER=openai
OPENAI_API_KEY=YOUR_OPENAI_API_KEY
# Or use the signed-in Codex CLI account (no API key)
# CRITIC_PROVIDER=codex-cli
# Or another signed-in local CLI (no separate API key)
# CRITIC_PROVIDER=claude-code
# CRITIC_PROVIDER=cursor-agent
# CRITIC_PROVIDER=opencode-cli
# CRITIC_MODEL=provider/model # required for opencode-cli; see `opencode models`
# Or OpenCode (https://opencode.ai/auth)
# CRITIC_PROVIDER=opencode
# OPENCODE_API_KEY=...
# OPENCODE_PLAN=go
# CRITIC_MODEL=minimax-m3
Provider
CRITIC_PROVIDER
Authentication
OpenAI
openai
OPENAI_API_KEY
Anthropic
anthropic
ANTHROPIC_API_KEY
Google Gemini
google
GOOGLE_GENERATIVE_AI_API_KEY
MiniMax
minimax
MINIMAX_API_KEY
OpenCode
opencode
OPENCODE_API_KEY (+ optional OPENCODE_PLAN)
Codex CLI
codex-cli
Existing codex login session
Claude Code
claude-code
Existing Claude Code account session
Cursor Agent
cursor-agent
Existing agent login session
OpenCode CLI
opencode-cli
Saved opencode auth login credentials + model
For Cursor Agent, Vibe-Gate runs agent first and falls back to the legacy cursor-agent executable if the primary command is unavailable.
opencode is still the separate Zen/Go HTTP provider and needs OPENCODE_API_KEY. opencode-cli runs the local CLI and requires a provider/model value in CRITIC_MODEL; see the CLI guide for details.
CRITIC_MODEL is passed to the selected provider or CLI without a Vibe-Gate model allowlist; the provider must support that model ID. The OpenAI API provider uses the Responses API. See provider configuration.
For direct providers, prefer keys in a local .env next to the package or in the consumer project under VIBE_WORKSPACE_ROOT (never commit secrets). MCP env overrides .env. For a CLI provider, install and sign in to that CLI as the same OS user running the MCP server.
Pass changed source paths in files[]. MCP reads the complete files from disk and serializes FILE/CONTENT internally for the Critic. The agent supplies paths and a completion report; it does not need to paste file bodies or generate a corpus. The example uses readOnly: true to keep review state untouched. See docs/SEMANTIC_DIFF_PAYLOAD.md.
Local development (this repo)
corepack yarn install
npm run build
npm test
cp .env.example .env # fill Critic key
npm start # stdio MCP
After npm run build, restart the vibe-gate MCP server in the IDE.
Publish
Releases are automatic. Every push to main runs .github/workflows/publish.yml: it runs yarn quality, lets
semantic-release choose the version from the conventional commits (fix: is a patch, feat: a minor), publishes
vibe-gate-mcp to npm with provenance, and lists that version in the
MCP Registry as io.github.mustafacagri/vibe-gate-mcp. No token is
stored: npm and the registry both trust the workflow through GitHub OIDC. server.json is the registry entry and
is kept at the release version by scripts/sync-server-json.mjs.
yarn quality # what the workflow checks before it tags anything
npm pack --dry-run # inspect the exact tarball contents
Consumers then use npx -y vibe-gate-mcp as above.
Payload sources — prefer files[]
Priority
Field
Use
1
files[]
Normal batches
2
semanticDiffPath
Existing compatibility carrier
3
semanticDiff
Existing compatibility carrier
Use readOnly: true for probes. updateStatus: false and mcp-smoke- / vibe-gate-probe- phase prefixes only skip phase status writes.
Set readOnly: true on submit_phase_review for a check that must leave the workspace untouched. It overrides updateStatus: true and prevents session clearing/saving, status updates, debt appends, conflict counter updates and deadlock case writes. Verdicts and concern verification still follow the ordinary review rules; deadlock case data is returned without saving it. Existing matching sessions can be read on later rounds, but read-only calls do not save a new round. logToDebt still expresses acceptance of debt when required, without writing the log. The default is false. updateStatus: false alone only disables phase status updates.
Every carrier (files[], inline semanticDiff, raw/JSON semanticDiffPath) is limited to ten actual FILE/CONTENT source blocks, including additional REQUEST: context on later rounds. Duplicate blocks and both rename endpoints count separately, even if their paths or bytes match. An oversized corpus fails before the enlarged Critic request; it is never reduced to the first ten blocks. Requested line ranges are read as complete files, subject to the existing file and aggregate size limits. Markdown, JSON and other document/data endpoints are rejected as source context. Callers positively classify their selected source endpoints; files[] lets MCP supply the full source bodies. For deleted files and old rename endpoints, an orchestrator such as Vibe-Pilot materializes pinned Git base blobs at real snapshot paths before passing those paths to files[]. MCP reads those files from disk; it does not read Git history itself.