vnem

quality contract

Triple-check before code.

VNEM prevents the lazy trade-off where an agent makes one metric better by quietly making the product worse somewhere else.

01

Analyze

Identify the stated goal, hidden requirements, visible surfaces, interaction needs, and risk domains.

02

Architect

Plan performance and visuals/playability together using profiles, toggles, adaptive effects, and fallback paths.

03

Review

Verify no important domain was sacrificed before final output, and report any remaining trade-off plainly.

orchestration

Route the right workflow before context collapses.

VNEM chooses the smallest deterministic pattern: single agent for simple work, orchestrator-worker for app and game builds, split-and-merge for deep research.

01

Route

Classify the prompt and return the orchestration pattern, reasons, worker count, and strict JSON contracts.

02

Coordinate

Use shared-state task claims, worker reports, artifacts, facts, decisions, blockers, and ordered events.

03

Reflect

Run Generator and Evaluator roles for up to three schema-valid iterations, then pass, revise, or block honestly.

precision

Exact patches. Current docs. Real feedback.

The default MCP stays read-only. When a project explicitly enables the precision server, VNEM gives agents controlled mutation, local context, and proof loops instead of a full-file rewrite button.

01

Patch

Apply exact SEARCH/REPLACE blocks or unified diffs only after dry-run context verification.

02

Ingest

Fetch current HTTPS framework docs and inject compact context before workers write code.

03

Verify

Run allowlisted build, test, and check commands with stdout, stderr, cwd state, and timeouts.

omniscient context

Find the code. Prove the logic. Clean up the helper.

Stage 4 adds a local semantic code index and test-driven healing loop to the opt-in precision server, so agents stop guessing file paths and stop claiming logic works without evidence.

01

Search

Use local private semantic search to return paths, line ranges, snippets, and scores before manual traversal.

02

Heal

Run red/green verification loops with hard attempt limits before reporting feature or logic success.

03

Script

Run short temporary Node or Python helpers for unique local roadblocks, then delete the sandbox.

intelligence pipeline

Research. Protect. Stage. Review.

The local app server can turn a dashboard target into an auditable dispatch flow: Research AI gathers a source-backed signal, Protection AI reviews provenance and risk, and Giving AI stages markdown for explicit owner approval.

01

Smart queue

OpenRouter inference is serialized. If a 429 happens, VNEM waits for the exact retry window, keeps the active payload queued, and shows a resume countdown.

02

Review modal

Staged findings open from the matrix into a markdown review view with Research, Protection, and Giving notes plus the raw dispatch body.

03

Safe approval

Approve only moves markdown from .vnem/staging/ to .vnem/approved/. Reject deletes the staged file. Neither action commits code or executes scripts.

install

Install vnem: one curl.

The public installer clones the full vnem repo into ./vnem, so you get the CLI, MCP server, registry, generated pack, and docs.

01

Open the folder where you want the vnem checkout to live.

02

curl -fsSL https://vnem.pages.dev/i | sh

03

Run cd vnem, then use the CLI installer for any project that should receive the read-only pack.

Existing repo: managed install.

If the repo already has an AGENTS.md, use the local installer so vnem adds a managed block instead of replacing the file.

git clone https://github.com/Ovvuhy/vnem.git
cd vnem
npm install
npm run install:project -- /path/to/project
npm run doctor -- /path/to/project

Project pack only.

For a clean project that only needs the read-only agent guidance files, extract the pack directly in that project folder.

curl -fsSL https://vnem.pages.dev/install.tgz | tar -xz

Need dashboard review?

Run the ARD operator console with the primary dashboard command. Use it to inspect repo/local Research AI v2 lanes, Protection AI v2 branch eligibility, Giving AI work packages, and Changes by ARD protected branch previews. npm run dev:all remains a secondary broad launcher.

npm.cmd run dashboard
npm.cmd run ard:dogfood
npm.cmd run test:current

Need MCP too?

From the same checkout, generate an absolute-path config instead of hand-writing paths.

node scripts/vnem-cli.mjs mcp-config
node scripts/vnem-cli.mjs mcp-config --server-json

Need precision execution?

Enable this only for the project workspace where exact patching, current-doc ingestion, safe terminal checks, semantic search, proof loops, and temporary helpers are allowed.

node scripts/vnem-cli.mjs mcp-config --precision --workspace /path/to/project
npm run precision:mcp

what the agent gets

Current agent tooling

Searchable guidance for MCP servers, skills, agent frameworks, evals, memory layers, search tools, identity, payments, and workflow patterns.

Source radar

Official docs, registries, MCP sources, eval frameworks, and verification sources are mapped so agents research from high-signal upstreams first.

Repo-aware review protocol

The agent inspects manifests and configs first, then separates stale or risky choices from better current options.

Holistic quality gate

Performance, visuals, playability, accessibility, maintainability, and safety are checked together instead of traded away silently.

Real MCP server

Use vnem as a local stdio MCP for search, recommendations, orchestration plans, entry lookup, comparison, and best-practice routing.

Precision MCP server

Opt in to exact patching, current documentation ingestion, semantic code search, bounded verification, and temporary helper scripts when a project needs controlled mutation.

mcp

Use vnem as a read-only tool server.

For MCP clients, point the default server at a local vnem checkout. It reads generated artifacts and never installs packages, edits code, calls upstream services, or collects secrets. Precision execution is a separate opt-in server with an explicit workspace root.

{
  "mcpServers": {
    "vnem": {
      "command": "node",
      "args": ["/path/to/vnem/scripts/vnem-mcp-server.mjs"],
      "env": {
        "VNEM_ROOT": "/path/to/vnem"
      }
    }
  }
}

inside the pack

Local guidance pack. Better agent judgment.

AGENTS.mdroot pointer that tells agents to read vnem automatically
.vnem/AGENTS.mdhow the agent should read and report
.vnem/install-guide.mddownload, managed install, MCP config, and verification guide
.vnem/operating-protocol.mduniversal loop and task contract
.vnem/quality-contract.mdholistic excellence, Triple-Check Workflow, and intelligent trade-offs
.vnem/orchestration-protocol.mdrouting, reflection, Magentic coding, split-and-merge research, shared state
.vnem/precision-execution-protocol.mdexact patching, current docs, safe terminal feedback, and mutation boundaries
.vnem/omniscient-self-healing-protocol.mdsemantic code search, verification loops, and temporary helper scripts
.vnem/coding-protocol.mdrepo sensing, implementation rules, and verification ladder
.vnem/coding-playbooks.jsonmode-specific execution loops for coding tasks
.vnem/design-architecture.mdvisual, motion, sound, dashboard, and game-feel guidance
.vnem/visual-qa-protocol.mdrendered screenshot, mobile, and interaction quality checks
.vnem/task-rubrics.jsonbroad quality bars, gates, and verification checks
.vnem/search-index.jsontools, MCPs, scores, freshness, risk flags
.vnem/source-radar.jsonofficial docs, registries, MCP sources, evals, verification inputs
.vnem/best-practices.mdcurrent stack guidance by project area
.vnem/agent-workspace.mdMCP gateways, memory files, and agent modes
.vnem/prompt-engineering.mdprompt enhancement protocol
.vnem/prompt-patterns.jsonmachine-readable task patterns

blog

New: MCP, Skill, and one-command repo install.

vnem now ships a real read-only MCP server, a Codex skill, a clean-folder curl install, and a broader registry for agentic coding, memory, search, and browser-game stacks.

Read the update

AI-improvement forge for agents.