Local-First SQLite WAL Zero Backend Runtime Dependencies Meta Muse & Llama Reasoning Ready

Git for AI Context.
End Coding Agent Amnesia.

Git tracks your source code. VectorForce tracks your AI coding agent’s architectural decisions, non-negotiable constraints, and verifiable workspace evidence—so you can switch seamlessly between Google Antigravity, Meta Muse, Claude Code, Cursor, Windsurf, GitHub Copilot, and OpenAI Codex without losing a single thought.

Free OSS Installation ☁️ Live Cloud Demo ↗ Watch Demo Video

$ npx vectorforce init && npx vectorforce studio

Works natively via MCP with:

⚡ Google Antigravity 🦙 Meta Muse ⌘ Claude / Command Code ◈ OpenAI Codex 🖱️ Cursor IDE 🌊 Windsurf Cascade 🐙 GitHub Copilot ⚡ Kiro IDE

Understand It In 10 Seconds

Why AI Coding Agents Break Your Architecture

Every time you close a chat window or switch tools, your agent forgets why your code was designed that way. VectorForce replaces ephemeral chat history with cryptographic engineering memory.

✗ Without VectorForce

Ephemeral Chat = Agent Amnesia

Each coding agent session starts blind, guessing your rules and silently breaking prior invariants.

Vanishing Decisions: Architectural trade-offs and security boundaries disappear when the context window resets.

Silent Workspace Drift: Teammates or other agents edit files, and your current agent continues with stale assumptions.

Locked Into One Tool: Starting a feature in Antigravity and finishing it in Command Code or Codex requires re-explaining everything.

Hallucinated Ground Truth: Agents confuse their own unverified claims with actual Git and file state.

✓ With VectorForce

Cryptographic Context Continuity

Immutable checkpoints bind model intent to verifiable Git commits and SHA-256 file hashes.

✓

Immutable Checkpoints: Goals, non-negotiable constraints, and ADRs are committed to a local SQLite WAL store.

✓

Automatic Drift Detection: Before resuming work, VectorForce compares current Git HEAD and file SHA-256 hashes against the last checkpoint.

✓

Instant Cross-Harness Resume: Start in Antigravity, run vf resume in Command Code, Muse, or Codex, and pick up immediately.

✓

Claims vs. Evidence Fencing: Model assertions (claims) are strictly separated from cryptographic observations (evidence).

Product Walkthrough

See VectorForce in Action

Watch how VectorForce captures checkpoints, detects silent workspace drift, reasons over engineering memory with Meta Llama/Muse, and hands off context across harnesses.

1. The Problem: Coding Agent Amnesia

Click Play or select any chapter below to preview the workflow (or drop in demo.mp4 / YouTube URL in index.html).

# Without VectorForce: Switching from Antigravity to Command Code $ claude / command-code > "Continue working on the OAuth PKCE task" ⚠️ Agent has zero memory of prior session: ✗ Forgot invariant constraint: "Zero external runtime dependencies" ✗ Unaware that src/policy/authorizer.js was modified 10 mins ago ✗ Re-introduces npm package violating ADR-002 security boundary

💡 Video Placeholder Ready: Replace the placeholder in index.html (line 195) with your ./demo.mp4 file or YouTube/Loom embed URL anytime.

Core Capabilities

Engineered for Privacy, Speed & Verifiability

Built from the ground up using pure Node.js built-ins (node:sqlite, node:crypto, node:http) with zero third-party runtime dependencies.

🔒

Claims vs. Evidence Fencing

Model assertions (goals, constraints, ADRs) are recorded as claims. Git commit SHAs and SHA-256 document hashes are recorded as evidence—preventing AI hallucinations from masquerading as ground truth.

🧭

Workspace & Document Drift

Automatically detects if a commit changed or an allowlisted specification drifted since the last checkpoint, alerting your agent before it writes conflicting code.

🧠

Meta Muse & Llama Reasoning

Run vf reason to synthesize where you left off, what changed in the codebase, architectural risks, and ranked next steps—with graceful offline fallback.

🔌

Universal MCP Protocol

Exposes a compliant JSON-RPC stdio Model Context Protocol server with 9 specialized tools for Google Antigravity, Command Code, Meta Muse, and OpenAI Codex.

📦

Tamper-Evident Handoffs

Export selective context packages with SHA-256 manifest verification. Share task continuity with teammates or remote workspaces without leaking secrets.

🛡️

Zero-Leak Allowlist Policy

Strict workspace containment blocks path traversal (..), symlinks, binary files, and sensitive credentials (.env, *.key). Receipts never store raw file bodies.

VectorForce Studio (Web UI)

Interactive Architecture Control Center

Launch npx vectorforce studio on port 3210 to inspect module dependencies, blast radius, governing ADRs, and live AI memory. Click any node below to inspect it:

VectorForce Studio • http://127.0.0.1:3210 Health: 98%

Architecture Topology Graph Claims vs Evidence Memory Timeline Safe Handoff

👆 Click any architecture node above to inspect its blast radius & ADR contract 75/75 Tests Passing

VERIFIED • ZERO DRIFT

src/mcp/mcp-server.js

Harness Bridge Layer (JSON-RPC stdio)

Blast Radius

High (4 AI Harnesses: Antigravity, Command Code, Muse, Codex)

Governing Architecture Decision (ADR)

ADR-004 & ADR-012: Multi-Harness MCP Protocol

Contract & Security Role

Exposes 9 deterministic tools over JSON-RPC stdio. Enforces destination harness authorization before serving briefings.

Step-by-Step Setup

Installation & Quickstart Guide (Free Open-Source Edition)

Get the local-first Open-Source Edition running in under 60 seconds (requires Node.js 22.5.0+). Need Multi-Project Cloud Sync, Developer Task Assignment, or the PM & Customer Web Portal? Explore VectorForce Cloud (Launching Post-Diwali • After Nov 15) →

STEP 1

Install VectorForce (npm or GitHub)

Install globally via npm to use the vf and vectorforce CLI commands anywhere, or clone directly from GitHub:

Terminal — Global npm Install or Source Clone

# Option A: Install globally from npm (Recommended)
npm install -g vectorforce

# Option B: Run directly via npx without installing
npx vectorforce --help

# Option C: Clone from GitHub (Zero dependencies to install!)
git clone https://github.com/getvectorforce/VectorForce.git
cd VectorForce && npm test

STEP 2

Initialize Your Project & Create an Active Task

Navigate to your repository and initialize VectorForce. This creates a local, git-ignored .vectorforce/vectorforce.db SQLite WAL store inside your workspace.

Terminal — Project & Task Initialization

# 1. Initialize VectorForce in your repository root
vf init

# 2. Create and activate an engineering task
vf task auth-module "Implement PKCE OAuth Authentication"

# 3. Allowlist architecture documents for cryptographic hash tracking
vf policy allow AGENTS.md
vf policy allow README.md

STEP 3

Connect Your AI Coding Agent (Zero Manual JSON Editing)

How MCP actually works: You never have to run vf mcp manually in a terminal! Instead, each AI coding tool reads its own configuration file and automatically spawns VectorForce in the background over stdio.

⚡

Google Antigravity: Reads .agents/mcp_config.json (workspace) and ~/.gemini/antigravity/mcp_config.json (global IDE).

🦙

Meta Muse: Reads ~/.config/muse/settings.json and workspace .muse/mcp_config.json.

⌘

Claude Code & Command Code: Auto-detects .mcp.json and commandcode.json in project root.

◈

OpenAI Codex CLI: Reads .codex/config.toml and ~/.codex/config.toml.

🖱️

Cursor IDE: Reads .cursor/mcp.json and steers via .cursorrules.

🌊

Windsurf Cascade: Reads .windsurf/mcp_config.json and .windsurfrules.

🐙

GitHub Copilot: Reads .vscode/mcp.json (official servers key) and .github/copilot-instructions.md.

⚡

Kiro IDE: Reads .kiro/settings/mcp.json and kiro.json.

Auto-Configure All Harnesses + Verify with vf status

# 1. Auto-configure Antigravity, Claude Code, Command Code, Meta Muse, Codex, Cursor, Windsurf, Copilot & Kiro:
npx vectorforce config install all

# 2. Check which MCP config files were created and verify readiness:
npx vectorforce status

# Note: You do NOT need to run "vf mcp" manually!
# Each AI tool reads its config file and spawns the MCP server automatically.

✅ How do I know it's working inside my AI agent?
Open your AI coding agent (Antigravity, Claude Code, Command Code, Muse, or Codex) and type in the chat:
"List active VectorForce tasks and check workspace drift"
You will immediately see the agent call the list_tasks and check_workspace_drift MCP tools!

STEP 4

Capture Checkpoints & Detect Workspace Drift

At the end of a work session—or before switching from one AI agent to another—record an immutable checkpoint. When resuming, check for drift and retrieve the authorized briefing:

Terminal — Checkpoint, Drift & Cross-Harness Resume

# Commit an immutable checkpoint with Claims + Verifiable Evidence
vf checkpoint --task auth-module \
  --summary "Completed PKCE OAuth handshake and token validation" \
  --goal "Ship secure OAuth flow" \
  --constraint "Zero external auth libraries" \
  --decision "DEC-AUTH-01:PKCE:Use SHA-256 code challenge with local storage" \
  --next "Implement refresh token rotation" \
  --doc "AGENTS.md"

# Check if Git commits or allowlisted files drifted since the checkpoint
vf drift --task auth-module

# Generate an authorized briefing for your destination harness
vf resume --task auth-module --dest antigravity

STEP 5

Configure AI Memory Models (Meta Muse & Muse Spark Open 1st Priority + Fireworks, Ollama, OpenRouter, OpenAI, Anthropic, BYOK)

Built for the Meta Muse Hackathon: VectorForce makes Meta Muse (muse-spark-1.3) and the open-weights Meta Muse Spark Open (muse-spark-open • 1M token context window) its #1 priority reasoning models—accessible directly via Meta API, Fireworks AI, OpenRouter, or Local Ollama:

Meta Muse Direct API + Open-Source Muse Spark Open (1st Priority Default)

# Option A: Configure Meta Muse Direct API (api.meta.ai/v1 • muse-spark-1.3)
vf config muse <YOUR_API_KEY>
vf reason --task auth-module

# Option B: Use Open-Weights "Muse Spark Open" (1M Context) via Fireworks AI:
vf config fireworks <YOUR_FIREWORKS_API_KEY> accounts/fireworks/models/muse-spark-open
vf reason --task auth-module

# Option C: Use "Muse Spark Open" via OpenRouter:
vf config openrouter <YOUR_OPENROUTER_API_KEY> meta/muse-spark-open

# Launch VectorForce Studio Web Control Center on http://127.0.0.1:3210
vf studio

🖥️ Configure Visually in VectorForce Studio: Run vf studio (opens http://127.0.0.1:3210) and click the Settings tab in the sidebar to switch between Meta Muse, Fireworks AI (muse-spark-open), OpenRouter, Ollama, OpenAI, or Anthropic. Need deterministic output with zero LLM calls? Run vf reason --task auth-module --raw.

VectorForce Cloud & Remote Context Hub

The Open-Source CLI is built for a single developer on one machine. VectorForce Cloud connects multiple devices (Lenovo, MacBook, Linux) and teams with end-to-end zero-knowledge encryption, live drift detection, and automated cross-agent briefings.

🌐 Open Live Cloud Dashboard (Meta AI Showcase) ↗ 📁 Open Restaurant Billing Workspace ↗

Commercial & Cloud Platform Live on Modal.com • Zero-Knowledge E2EE

Everything Beyond a Single Machine — Built for Teams, PMs & Customers

No manual database setup or file copying. Subscribe, grab your vf_live_... key, and unlock seamless multi-developer collaboration:

🗂️ Available Projects Hub

Switch across all company & client projects from a single web portal—each with isolated policies, health scores, and decision ledgers.

👥 Developer Task Differentiation

Assign tasks per developer (Assignee, Role: Frontend/Backend/QA, Priority) with scope locks so agents never collide.

🔄 Two-Machine Live Context Sync

Move between your office workstation and laptop—or sync with teammates via vf push / vf pull without manual bundle files.

📊 PM & Customer Interpreter Portal

PMs & clients log in on the web (zero CLI needed) to view plain-English progress, decision rationale, and upload to Jira, Linear, Notion, or Slack.

🔑 Subscription Cloud MCP

Connect Antigravity, Claude Code, Command Code, Meta Muse, and Codex via remote serverUrl using your subscription API key.

🚀 One-Command Service Deploy

Prefer self-hosting in your own cloud VPC? Run vf deploy to spin up the full multi-project hub on Docker, Railway, or Render in 1 command.

Zero-Install Cloud Subscription MCP (.mcp.json / Antigravity / Muse)

{
  "mcpServers": {
    "vectorforce-cloud": {
      "serverUrl": "https://mcp.vectorforce.ai/v1/sse",
      "headers": {
        "Authorization": "Bearer vf_live_sub_98a7bc65...",
        "X-VF-Project-Id": "proj_bistroforce_prod",
        "X-VF-Developer": "rahul@company.com"
      }
    }
  }
}

CUSTOMER VALIDATION & EARLY ACCESS 🔥 48 Teams Interested

Are You Interested in VectorForce Cloud?

We are launching the VectorForce Cloud & Team Subscription 10 days after Diwali (right after 15 November 2026). Tell us if you or your team want early access so we can prioritize your onboarding:

1. Are you interested in the Cloud Subscription launching after Nov 15?

2. Which Cloud features are most critical for you? (Click to select)

3. Your Role

4. Team / Machines

Your Name / Company

Work Email (For Nov 15 Invite)

Anything specific you want in the Nov 15 Cloud release? (Optional)

Model Context Protocol

9 Native MCP Tools for Autonomous Agents

Once connected via MCP, your AI coding agent automatically calls these tools to stay aligned with your project's architectural invariants.

save_checkpoint WRITE

Commits an immutable checkpoint combining model claims with verifiable Git commit SHAs and allowlisted document hashes.

get_resume_briefing READ

Retrieves a harness-authorized markdown briefing separating invariant constraints, active goals, and verified evidence.

check_workspace_drift VERIFY

Compares the live working tree and allowlisted files against the latest checkpoint to flag silent changes.

reason_about_task AI MEMORY

Generates a Llama / Meta Muse narrative summarizing where work left off, architectural risks, and ranked next actions.

list_tasks QUERY

Discovers and filters active, paused, or completed tasks tracked inside the workspace SQLite WAL database.

set_document_policy POLICY

Manages the strict project allowlist governing which markdown or architecture files can be collected as evidence.

export_handoff BUNDLE

Packages selected checkpoints and verified evidence into a tamper-evident SHA-256 bundle with preview mode.

import_handoff RECONCILE

Cryptographically verifies a handoff manifest and safely imports external task context into a destination workspace.

get_timeline AUDIT

Synthesizes a chronological continuity recap of checkpoints, decisions, and handoffs across sessions.

Git for AI Context. End Coding Agent Amnesia.