A quantitative and qualitative comparison of Fulcrum, Vibe Kanban, Auto-Claude, and OpenClaw
The AI-assisted development tool space is exploding. In the span of a few months, multiple projects have emerged with overlapping but distinct visions for how developers should interact with AI coding agents. I took four of them — Fulcrum, Vibe Kanban, Auto-Claude, and OpenClaw — and put them under the microscope using scc for code metrics, git log for history, and manual code review for everything else.
What follows is a framework-driven comparison across eight dimensions, from raw LOC counts to architectural philosophy to "are they punching above their weight?"
| Fulcrum | Vibe Kanban | Auto-Claude | OpenClaw | |
|---|---|---|---|---|
| One-liner | AI agent command center for shipping software | Visual orchestrator for 10+ coding agents | Autonomous multi-agent coding pipeline | Personal AI assistant gateway across messaging channels |
| Target user | Solo dev / CTO with multiple AI agents | Dev teams using diverse coding agents | Developers wanting hands-off AI implementation | Power users wanting AI in their existing chat apps |
| First commit | Dec 19, 2025 | Jun 14, 2025 | Dec 4, 2025 | Nov 24, 2025 |
| Age | ~2 months | ~8 months | ~2 months | ~2.5 months |
| Total commits | 1,162 | 1,717 | 698 | 10,684 |
| Contributors | 4 | 51 | 20 | 20+ |
| License | PolyForm Perimeter | Apache 2.0 | (not identified) | (see repo) |
| Primary language(s) | TypeScript | TypeScript + Rust | TypeScript + Python | TypeScript + Swift + Kotlin |
| Backend runtime | Bun (Hono.js) | Rust (Axum) | Python (Claude Agent SDK) | Node.js (Express) |
| Frontend | React 19, TanStack Router | React 18, Zustand | Electron + React 19 | Native apps (SwiftUI, Kotlin) + Web + TUI |
The age and contributor counts already tell a story. Vibe Kanban is the oldest at 8 months with the largest community (51 contributors). OpenClaw compensates for its 2.5-month lifespan with staggering velocity — 10,684 commits, 73% from a single developer. Fulcrum and Auto-Claude are both ~2-month-old projects with smaller teams but ambitious scope.
All metrics from scc. "Source LOC" excludes JSON, Markdown, YAML, and other config/documentation formats. Auto-Claude's 260k LOC of bundled JavaScript output is excluded.
| Metric | Fulcrum | Vibe Kanban | Auto-Claude | OpenClaw |
|---|---|---|---|---|
| Source LOC | 111,880 | 166,219 | 255,586 | 682,009 |
| Total files | 680 | 1,615 | 1,333 | 5,317 |
| Avg LOC/file | 165 | 103 | 192 | 128 |
| Cyclomatic complexity | 12,846 | 13,526 | 25,586 | 86,312 |
| Complexity/kLOC | 114.8 | 81.4 | 100.1 | 126.6 |
| Comment density | 8.6% | 5.0% | 17.9% | 2.6% |
| Docs (Markdown LOC) | 1,113 | 8,197 | 20,805 | 105,021 |
OpenClaw is a behemoth. At 682k source LOC across TypeScript, Swift, Kotlin, and Go, it dwarfs the field by 2.7x over the next largest project. Its complexity density (126.6/kLOC) is also the highest — the inevitable cost of routing messages across 45 channels with edge-case handling for each.
Vibe Kanban has the lowest complexity density (81.4/kLOC) despite a non-trivial 166k LOC codebase. The Rust backend helps here — the type system eliminates classes of bugs that would otherwise require defensive branching in TypeScript.
Auto-Claude's comment density (17.9%) is 3.5x higher than the field average. A likely fingerprint of AI-assisted code generation: LLMs tend to over-explain. By contrast, OpenClaw's 2.6% comment density suggests a "the code is the documentation" philosophy despite its massive scale.
Fulcrum is the most compact. At 112k LOC it's the smallest codebase, yet as we'll see in the feature inventory, it punches far above this weight class.
| Fulcrum | Vibe Kanban | Auto-Claude | OpenClaw | |
|---|---|---|---|---|
| Largest file | settings.tsx (2,820 LOC) | claude.rs (2,746 LOC) | workspace.py (80.6k) | audit.test.ts (2,133 LOC) |
| Hotspot pattern | Settings UI, route handlers | Agent executors, log normalizers | Agent impls, workspace mgmt | Agent runners, channel impls |
Auto-Claude stands out with workspace.py at 80.6k LOC — a God Object that manages the entire workspace orchestration. This single file is larger than many entire projects.
Fulcrum (20 features): Task management (3 types: git worktree, scratch, manual) · Kanban board with drag-drop · Task dependencies + DAG visualization · Parallel agent orchestration (Claude Code, OpenCode) · Persistent terminal tabs (dtach) · Project/repository management · Calendar sync (CalDAV + Google OAuth) · Messaging (WhatsApp, Telegram, Slack, Discord, Email) · Proactive AI concierge (morning/evening rituals, hourly sweeps) · Two-tier memory system (master file + ephemeral FTS5) · Docker Compose deployment · Cloudflare tunnel/DNS integration · System monitoring (CPU/mem/disk) · Built-in AI assistant · MCP server (113 tools) · Desktop app (Neutralino) · Unified FTS5 search · Multi-channel notifications · PR monitoring + auto-close · Job management (systemd/launchd)
Vibe Kanban (14 features): Task management + Kanban board · Agent orchestration (10+ agents: Claude, Copilot, Cursor, Gemini, Amp, Codex, OpenCode, Droid, Qwen, ACP) · MCP integration with tool approval · Code change review (git diffs) · Git worktree isolation · Branch management (merge, rebase, conflict detection) · PR creation/tracking · Editor integration (VS Code, Cursor, Windsurf, IntelliJ, Zed) · Terminal emulation · Dev server execution · Real-time cloud sync (ElectricSQL) · Multi-user organizations · GitHub/Azure DevOps integration · Docker self-hosting
Auto-Claude (14 features): Spec creation pipeline (3 complexity tiers) · Multi-session agent pipeline (planner → coder → QA reviewer → QA fixer) · Git worktree isolation · Graph memory system (Graphiti + LadybugDB) · Kanban visualization · Roadmap planning with AI · GitHub/GitLab integration · Linear task sync · E2E testing via Electron MCP · Ideation engine (6 improvement categories) · Complexity assessment · Automatic merge conflict resolution · Auto-generated release notes · Cross-platform desktop app (Electron)
OpenClaw (15 features): Multi-channel messaging (8 built-in + 37 extension plugins) · Multi-agent orchestration with nested sub-agents · 54 built-in skills · Memory system (QMD + vector embeddings) · Sandboxed code execution (Docker) · Browser automation (Playwright) · Text-to-speech · Terminal UI · Web gateway + control panel · Auto-reply triggers + cron · Configuration management + state migrations · Native apps (macOS/iOS/Android) · Plugin SDK · Hook automation system · Security audit framework
| Features | Source kLOC | Features/kLOC | Complexity/Feature | |
|---|---|---|---|---|
| Fulcrum | 20 | 112 | 0.179 | 642 |
| Vibe Kanban | 14 | 166 | 0.084 | 966 |
| Auto-Claude | 14 | 256 | 0.055 | 1,828 |
| OpenClaw | 15 | 682 | 0.022 | 5,754 |
Fulcrum delivers 2.1x more features per kLOC than the next closest project and 8x more than OpenClaw. Its complexity-per-feature (642) is also the lowest — each feature is implemented with minimal overhead.
OpenClaw's 5,754 complexity points per feature reflects the reality of supporting 45 messaging channels with platform-specific edge cases. Each "feature" requires enormous implementation surface.
| Pattern | Frontend | Backend | Database | Real-time | |
|---|---|---|---|---|---|
| Fulcrum | Integrated monolith | React 19 + TanStack | Hono.js on Bun | SQLite (Drizzle ORM) | WebSocket |
| Vibe Kanban | Rust + React monorepo | React 18 + Zustand | Axum (Rust) | SQLite (SQLx) + PostgreSQL | ElectricSQL + WebSocket |
| Auto-Claude | Python + Electron hybrid | Electron + React 19 | Python native | LadybugDB (graph) | IPC + process streams |
| OpenClaw | Plugin-extended monolith | Native apps + Web | Express (Node.js) | SQLite + LanceDB (vector) | WebSocket + channel protocols |
Fulcrum bets on Bun's performance and Hono's minimalism. Single runtime, single database, single process. The MCP-first design means every feature is a tool that agents can discover on-demand, avoiding context bloat. Clean but coupled.
Vibe Kanban makes the boldest tech bet with its Rust backend. SQLx provides compile-time SQL validation; ts-rs auto-generates TypeScript types from Rust structs. The payoff: lowest complexity density in the field and fearless refactoring. The cost: higher barrier to contribution, dual-language maintenance.
Auto-Claude is the only project built primarily in Python, reflecting its AI-first philosophy — Python is the lingua franca of ML/AI tooling. The Electron frontend adds TypeScript, creating a dual-language codebase. The graph database (Graphiti/LadybugDB) is unique to this project and enables cross-session learning.
OpenClaw has the widest platform reach: Node.js backend + native apps in Swift (macOS/iOS), Kotlin (Android), and Go tooling. The plugin architecture (37 channel extensions + 54 skills) is the most extensible of the four, but the monolith core is massive.
| Estimated Generated/Vendored | Notes | |
|---|---|---|
| Fulcrum | ~1-2% | Drizzle migrations, minimal |
| Vibe Kanban | ~2-5% | ts-rs type generation from Rust |
| Auto-Claude | ~1-2% (excl. bundles) | 260k LOC bundled JS excluded entirely |
| OpenClaw | ~5% | Protocol schemas, i18n docs |
All four projects are predominantly hand-written. Auto-Claude's scc output is misleading without excluding the /out/ build artifacts.
This is where the projects diverge most sharply.
| Supported Agents | Primary SDK | MCP Tools | |
|---|---|---|---|
| Fulcrum | 2 (Claude Code, OpenCode) | Claude Agent SDK + custom | 113 |
| Vibe Kanban | 10+ (Claude, Copilot, Cursor, Gemini, Amp, Codex, etc.) | Custom executors per agent | Per-executor |
| Auto-Claude | 1 (Claude, configurable model) | Claude Agent SDK | ~6 MCP servers |
| OpenClaw | Multi-provider (Anthropic, OpenAI, Bedrock, Ollama, etc.) | PI (Prompt Integration) runner | Minimal (plugin/skill system instead) |
Vibe Kanban leads in agent breadth with native executors for 10+ coding agents. Each has a dedicated Rust executor module (Claude alone is 2,746 LOC) with streaming protocol handling. This is the project's core differentiator.
Fulcrum leads in MCP tool coverage. 113 tools with deferred loading via a search_tools discovery pattern. Agents find what they need without drowning in context. The MCP-first philosophy means everything in Fulcrum — tasks, projects, apps, calendar, memory — is an agent-accessible tool.
Auto-Claude takes the deepest single-agent approach. Rather than supporting many agents, it implements a sophisticated multi-session pipeline: planner → coder → QA reviewer → QA fixer. Agents coordinate across sessions, sharing graph-based memory. The "describe what you want, agents handle everything" model is the most autonomous.
OpenClaw sidesteps MCP almost entirely in favor of its own plugin/skill architecture. With 54 built-in skills and 37 channel extensions, it has the richest tool ecosystem — just not through MCP.
| Approach | Persistence | Search | |
|---|---|---|---|
| Fulcrum | Two-tier: master file (always in prompt) + ephemeral memories | SQLite + MEMORY.md | FTS5 full-text |
| Vibe Kanban | Per-workspace file context + git history | SQLite (local), PostgreSQL (cloud) | File indexing |
| Auto-Claude | Knowledge graph (Graphiti + LadybugDB) | Graph database | Semantic (embeddings, multi-provider) |
| OpenClaw | QMD (Queryable Markdown) + vector embeddings | SQLite + LanceDB | Semantic + FTS |
Auto-Claude's graph memory is the most sophisticated. Cross-session insight extraction means agents learn patterns, gotchas, and architectural decisions. The graph structure captures relationships between concepts that flat text can't.
Fulcrum's two-tier approach is the most practical. The master memory file (always injected into every system prompt) provides standing instructions, while ephemeral tagged memories with FTS5 search handle transient knowledge. Simple, effective, fast.
OpenClaw combines approaches with QMD structured memory and LanceDB vector search, plus session transcript compaction for token budget management.
Fulcrum: Parallel independence. Run multiple agents on separate tasks in isolated worktrees. Agents don't coordinate — the human is the orchestrator, aided by the proactive concierge.
Vibe Kanban: Agent-agnostic parallelism. The developer picks which agent to use per task, reviews diffs, and merges back. The tool stays out of the way.
Auto-Claude: Sequential pipeline coordination. Agents hand off to each other: spec → plan → implement → review → fix. The most automated but least flexible.
OpenClaw: Nested delegation. Agents can spawn sub-agents with configurable depth limits, model selection, and permission scoping. The most dynamic orchestration model.
| Dimension | Fulcrum | Vibe Kanban | Auto-Claude | OpenClaw |
|---|---|---|---|---|
| Auth model | None (single-user, localhost) | JWT + OAuth 2.0 (remote mode) | OAuth + keychain (desktop) | Password/token (gateway) |
| Secrets at rest | age-encrypted via fnox (ChaCha20-Poly1305) | Env vars only | OS keychain encrypted | Env vars + fnox integration |
| Input validation | Drizzle ORM, type-aware config validators, prompt injection warnings | CORS, path traversal prevention, compile-time SQL | 3-layer sandbox (OS + FS + dynamic allowlist) | Zod schemas, SSRF prevention, traversal detection |
| Execution sandbox | Process-level (worktrees) | Process-level (worktrees) | OS sandbox + FS restriction + command allowlist | Docker containers |
| Env hygiene | Clears sensitive env vars on startup | N/A | N/A | N/A |
| Audit trail | Observer invocations table + circuit breaker | N/A | N/A | N/A |
| Dep audit | None visible | None visible | VirusTotal scanning for releases | detect-secrets baseline, Oxlint |
| Security contact | N/A | N/A | N/A | Dedicated POC (Dvuln founder) |
| Known CVEs | None | None | None | CVE-2026-25253 (CVSS 8.8) + 512 vulns found in Jan 2026 audit |
| Incidents | None | None | None | 40k+ exposed instances, 12.8k exploitable via RCE |
Fulcrum takes a defense-in-depth approach anchored by encrypted secrets and minimal attack surface. All 24 sensitive settings (API keys, bot tokens, webhook URLs, OAuth secrets) are encrypted with age via fnox, using ChaCha20-Poly1305. Non-sensitive config (ports, theme, language) stays in plaintext for easy debugging. The age key file is created with 0600 permissions. On startup, clearSensitiveEnvVars() strips inherited credentials from the process environment before any config is read — preventing shell environment leakage into child processes. Inbound email is observe-only (no auto-responses, read-only task/memory creation), and an Observer audit trail tracks every message processing attempt with circuit breaker status. Critically, Fulcrum's localhost-only default binding means there's no exposed gateway to attack in the first place. Zero known CVEs or incidents.
OpenClaw has the most security code — and the most security incidents. Its rapid rise to 145k GitHub stars made it a target before its security posture was mature. CVE-2026-25253 (CVSS 8.8) allowed 1-click RCE through auth token exfiltration — the Control UI trusted a gatewayURL query parameter and sent credentials over WebSocket without origin verification. A January 2026 audit found 512 vulnerabilities, 8 critical. Censys identified 21,639 exposed instances, 63% exploitable. The extensive hardening (Docker sandboxing, Zod validation, SSRF prevention, detect-secrets, dedicated security POC from Dvuln) was largely built in response to these incidents, not in prevention of them. Credit where due: the team responded rapidly, but the track record matters.
Auto-Claude's 3-layer security model (OS sandbox, filesystem restrictions, dynamic command allowlist) is clever — the allowlist adapts to the project's tech stack rather than using a static list. No known incidents.
Vibe Kanban benefits from Rust's memory safety and compile-time SQL validation (SQLx), eliminating entire vulnerability classes at the language level. No known incidents.
| README quality | CLAUDE.md | Docs volume | Onboarding friction | |
|---|---|---|---|---|
| Fulcrum | Detailed (vision + setup) | Comprehensive (safety rules, conventions) | 1.1k LOC | Low (single mise run dev) |
| Vibe Kanban | Good (features + quickstart) | Present | 8.2k LOC + Mintlify site | Low (pnpm run dev) |
| Auto-Claude | Extensive (philosophy + setup) | Present | 20.8k LOC | Medium (Python + Node + Electron) |
| OpenClaw | Comprehensive | Detailed (agents guide) | 105k LOC (i18n) | Medium (native deps: canvas, pty) |
| Framework | Test files | Test ratio | Coverage enforcement | |
|---|---|---|---|---|
| Fulcrum | Bun built-in | 60 | 12% by file | None |
| Vibe Kanban | cargo-nextest + Vitest | ~10 | ~5-10% | CI-enforced lint |
| Auto-Claude | pytest + Vitest + Playwright | 83 | ~6% | None |
| OpenClaw | Vitest | 1,131 | 37% by file | 70% coverage threshold |
OpenClaw's test discipline is in a different league. 1,131 test files, enforced 70% coverage thresholds, separate configs for unit/e2e/live/extension/gateway tests, and process-isolated test execution. This is enterprise-grade testing infrastructure.
Fulcrum's 60 test files punch above their count — focused on business logic (channels, memory, search) rather than trivial component tests. But the absence of frontend tests is a gap.
| Task runner | Linter | Formatter | CI/CD | |
|---|---|---|---|---|
| Fulcrum | mise | ESLint | (Implicit via ESLint) | GitHub Actions |
| Vibe Kanban | pnpm scripts | ESLint + Clippy | Prettier + rustfmt | GitHub Actions (Blacksmith) |
| Auto-Claude | npm scripts | Biome | Biome | GitHub Actions (multi-platform) |
| OpenClaw | pnpm scripts | Oxlint (Rust-based) | Oxfmt (Rust-based) | GitHub Actions (9 workflows) |
OpenClaw's choice of Oxlint and Oxfmt (both Rust-based, type-aware) over ESLint/Prettier is forward-looking — dramatically faster on large codebases.
| Dimension | 1st | 2nd | 3rd | 4th |
|---|---|---|---|---|
| Feature density | Fulcrum | Vibe Kanban | Auto-Claude | OpenClaw |
| Code quality (complexity/kLOC) | Vibe Kanban | Auto-Claude | Fulcrum | OpenClaw |
| Agent breadth | Vibe Kanban | OpenClaw | Fulcrum | Auto-Claude |
| Agent depth | Auto-Claude | OpenClaw | Fulcrum | Vibe Kanban |
| MCP/Tool ecosystem | Fulcrum | OpenClaw (skills) | Vibe Kanban | Auto-Claude |
| Memory sophistication | Auto-Claude | OpenClaw | Fulcrum | Vibe Kanban |
| Security posture | Fulcrum | Auto-Claude | Vibe Kanban | OpenClaw |
| Test discipline | OpenClaw | Fulcrum | Auto-Claude | Vibe Kanban |
| Documentation | OpenClaw | Auto-Claude | Vibe Kanban | Fulcrum |
| Community size | Vibe Kanban | Auto-Claude / OpenClaw | — | Fulcrum |
Fulcrum: Punching Way Above Weight
112k LOC delivering 20 major features is remarkable efficiency. The MCP-first architecture means every feature is immediately available to AI agents — not just a UI affordance. The proactive concierge (morning/evening rituals, message triage, memory system) adds a layer that none of the others attempt. The security model is solid: age-encrypted secrets, env var isolation, observe-only email, observer audit trail, and a localhost-only default that eliminates the exposed-gateway attack vector entirely. For a 2-month-old project built primarily by one developer, the feature density is unmatched.
Standout: Features-per-kLOC (0.179) is 2x the nearest competitor. Weak spot: Small contributor base. Single-user assumption limits multi-tenant use cases.
Vibe Kanban: The Engineer's Choice
The Rust backend is a genuine differentiator — lowest complexity density (81.4/kLOC) and compile-time guarantees that TypeScript/Python can't match. Supporting 10+ coding agents with dedicated executors is the widest agent support in the field. The 51-contributor community and 8-month history make it the most mature. The tradeoff: it's focused narrowly on "orchestrate agents, review diffs, merge code" — no messaging, no calendar, no deployment.
Standout: 10+ agent support with the lowest code complexity. Rust's type system as a quality multiplier. Weak spot: Thin test coverage relative to codebase size. Narrow feature scope.
Auto-Claude: The Autonomy Maximalist
Auto-Claude is the only project that aims for true hands-off implementation. The planner → coder → QA pipeline with graph-based memory is architecturally unique. But the 80.6k LOC workspace.py God Object and 17.9% comment density suggest a codebase that grew faster than it was refactored. The Electron + Python dual-runtime adds deployment friction. For teams that want to describe a feature and walk away, nothing else comes close.
Standout: Multi-session agent pipeline with persistent graph memory. Most autonomous. Weak spot: Highest complexity per feature (1,828). Dual-runtime friction.
OpenClaw: The Platform Play
At 682k LOC across four languages with native apps, 45 messaging channels, 54 skills, and a 37% test ratio, OpenClaw is playing a different game. It's not a dev tool — it's a personal AI assistant platform that happens to have coding capabilities. The security posture, test discipline, and documentation volume (105k LOC) are enterprise-grade. The 10,684 commits in 2.5 months from a small team (73% from one developer) represent extraordinary velocity, likely AI-assisted.
Standout: Test discipline (1,131 test files, 70% coverage threshold). Platform breadth. Weak spot: Highest complexity per feature (5,754). Massive maintenance surface.
Small & Sharp ◄───────────► Large & Comprehensive
│ │
Fulcrum ─────────────┤ │
│ │
Vibe Kanban ─────────┼──────────┤ │
│ │ │
Auto-Claude ─────────┼──────────┼────────┤ │
│ │ │ │
OpenClaw ────────────┼──────────┼────────┼──────────┤
│ │ │ │
112k 166k 256k 682k LOC
Each project occupies a distinct niche:
- Need max features in min code? → Fulcrum
- Need to support many different AI coding agents? → Vibe Kanban
- Need fully autonomous feature implementation? → Auto-Claude
- Need AI in every messaging app you use? → OpenClaw
There's no single winner because they're solving different problems. But the feature-density numbers don't lie: Fulcrum at 0.179 features/kLOC is doing more with less than anyone else in this comparison. Whether that density is sustainable as the project scales is the open question.
Analysis performed February 2026. Code metrics via scc, git history via git log, qualitative assessment via manual code review. Feature counts are editorial judgments — your mileage may vary.