MCP 服务器

Veris

Behavioral verification intelligence for AI coding agents. Reads a TypeScript or JavaScript repo, clusters functions into 25 semantic workflows (Authentication, Payments, Webhooks, Caching, Queue, and more), and emits concrete adversarial probes per workflow. 17 MCP tools, local SQLite state, zero cloud.

README

<h1 align="center">Veris</h1>

<p align="center"><strong>Behavioral Verification Infrastructure for autonomous coding agents.</strong></p>

Veris is the verification intelligence layer that sits between AI coding agents and production reliability. It does not run your tests. It tells any MCP-compatible coding agent or CI pipeline what behaviors are at risk, what to verify, and how confident the result actually is — backed by a behavioral graph, semantic workflow grouping, persistent run history, drift detection, and explainable confidence math.

Today: TypeScript + JavaScript repos. Python and Go adapters on the roadmap.

Works with any MCP client. CLI works standalone. Fully open source. Local-first. No cloud. No telemetry. No paid tier.

Plug-and-play install

Option A — As an MCP server (one config line)

Veris speaks the Model Context Protocol. Drop this into any MCP-compatible client config:

{
  "mcpServers": {
    "veris": {
      "command": "npx",
      "args": ["-y", "veris-core", "mcp"]
    }
  }
}

Restart the client. 17 tools light up: analyze_pr_behavior, list_workflows, detect_drift, generate_adversarial_probes, allocate_budget, what_if_revert, report_execution, and more.

Option B — As a CLI

npx veris-core .                                 # analyze current repo
npx veris-core . --base-ref=origin/main          # explicit git base ref
npx veris-core . --budget=10 --onboarding        # 10-min verification plan + onboarding map
npx veris-core init                              # scaffold .veris/ with plugin slot
npx veris-core doctor                            # health check

Reports land in veris-reports/:

veris-dashboard.html — interactive single-file dashboard (graph, heatmap, drift, probes, budget, history)
veris-report.md — markdown executive summary
onboarding/ — workflow-first markdown package for new engineers (with --onboarding)

Option C — From source

git clone https://github.com/vighriday/Veris
cd Veris
npm install && npm run build
node dist/cli.js .

What it gives you

Surface	What lands
Behavioral graph	Classes, methods, functions linked by `DependsOn` and real `Invokes` edges (call-expression resolution)
Semantic workflows	Auto-clustered into 25 domains (Authentication, Billing, Checkout, Caching, Queue, Webhooks, AI, ...)
Real git diff	Worktree-based diff vs any base ref. Not a placeholder
Risk scoring	Blast radius, fragility, runtime criticality + plain-English explanations
Confidence math	Half-life decay over real execution history. Failed runs reduce confidence; flaky = half credit
Drift detection	SHA-256 workflow fingerprints. Silent rewrites caught (same members, different topology)
Counterfactual mode	`what_if_revert(nodeIds)` simulates rollback impact
Adversarial probes	Concrete Tier 3 hypotheses per workflow kind (idempotency, replay, retry storms, cache stampede)
Budget allocator	Knapsack on `(tier × criticality × risk) / cost`. Highest-leverage subset within N minutes
Knowledge transfer	Workflow-first onboarding markdown package
Cross-repo view	Register multiple services; one MCP call for fleet-wide confidence
Interactive dashboard	Single-file HTML. Vis-network graph. Click workflow → filter everything. ESC to clear. Click-to-copy directives

Example agent prompts

Any MCP-compatible agent can drive Veris with prompts like these:

veris: analyze_pr_behavior with baseRef=origin/main
veris: list_workflows then detect_drift
veris: generate_adversarial_probes for the highest-risk workflow, then allocate_budget minutes=15
veris: what_if_revert nodeIds=[...]

After your agent runs the verifications it executed externally, close the loop:

veris: report_execution executions=[{nodeId:..., tier:'Tier 3', result:'pass'}, ...]

Confidence math now reflects what actually ran.

Privacy

Local-first. Everything runs on your machine.
No telemetry. Veris does not phone home.
Zero-retention mode. VERIS_STATE_DISABLED=1 skips all .veris/state.db writes.
No network calls. The MCP server speaks only over stdio.

Plugins

Drop a .js file into .veris/plugins/:

module.exports.register = function (api) {
    api.addWorkflowRule({
        kind: 'Payments',
        importTokens: ['stripe', '@yourorg/billing-sdk'],
        weight: 3
    });
    api.addRuntimeRisks('Payments', [
        '3DS challenge response lost on tab close'
    ]);
};

Full plugin API: docs/PLUGINS.md. Example: examples/plugin-fintech.js.

MCP tool reference

17 tools across categories: ingest, diff, plan, semantic, drift, counterfactual, verification, feedback, history, fleet.

See docs/MCP_TOOLS.md for the full reference with recommended flows.

Architecture

Source -> AST (ts-morph)
       -> Behavioral Graph (DependsOn + Invokes)
       -> Real git-worktree diff vs base ref
       -> Risk model (blast / fragility / criticality + explanations)
       -> Workflow classifier (25 semantic kinds, plugin-extensible)
       -> Fingerprints -> Drift detector (vs SQLite history)
       -> Adversarial probe generator
       -> Verification plan (Tier 1/2/3)
       -> Budget allocator (leverage / cost)
       -> Confidence engine (half-life decay over execution history)
       -> Reports + interactive dashboard
       -> MCP (17 tools) -> autonomous agents close the loop via report_execution

See ARCHITECTURE.md for the deep dive.