Not a framework. Not an SDK wrapper. An OS for AI agents.
The only agent runtime that rewrites itself.
Website • Docs • Twitter / X • Community • Discussions
┌─────────────────────────────────────────────────────────────────────────────┐
│ c l a w R E F O R M · v 0 . 2 . 2 │
│ by ae.ltd │
├──────────────┬──────────────┬──────────────┬──────────────┬─────────────────┤
│ 146 K │ 141 │ 27 │ 60 │ 0 │
│ LINES OF │ MODELS │ PROVIDERS │ BUNDLED │ RUNTIME │
│ RUST │ SUPPORTED │ NATIVELY │ SKILLS │ DEPS │
├──────────────┴──────────────┴──────────────┴──────────────┴─────────────────┤
│ 40 CHANNELS · 100+ API ENDPOINTS · 1744+ TESTS · 13 RUST CRATES │
│ A2A PROTOCOL · MCP CLIENT + SERVER · OPENAI-COMPATIBLE · WASM │
└─────────────────────────────────────────────────────────────────────────────┘
What makes clawREFORM different? It's the only open-source agent runtime that ships four architectural primitives — AgentDNA, Memory Ladder, Collective Conscience, and a Self-Rewrite Engine — in a single, zero-dependency Rust binary. No Python. No Node. No Docker required.
clawREFORM now ships with a dark-first tactile interface derived from the metallic brand mark.
- Dark mode is the default: forged charcoal substrate, brushed metal controls, amber emissive feedback.
- Light mode is fully designed, not inverted: stone/aluminum surfaces, restrained gold accent, preserved depth cues.
- Shared rules across both modes: geometric typography, debossed/stamped headings for key UI chrome, and machined pill controls for primary actions, toggles, and progress rails.
These are not features bolted onto an SDK. They are architectural primitives baked into the runtime. Each one solves a category of problem that no other agent OS has addressed.
The world's first file-native agent identity system.
Every other agent OS stores identity as a prompt string — untestable, ungittrackable, and opaque. clawREFORM stores agent identity as discrete, human-readable Markdown files called the AgentDNA stack. Each file has a single responsibility. The runtime injects them in deterministic order. Change a file → the agent changes. Instantly. No recompile. No redeploy. git blame your agent's personality.
A G E N T D N A S T A C K
┌────────────────────────────────────────────────────────────────┐
│ │
│ IDENTITY.md ← Who the agent is (name, role, purpose) │
│ SOUL.md ← How it thinks (values, style, constraints) │
│ HEARTBEAT.md ← What it does autonomously on a schedule │
│ HANDS.md ← Which autonomous action packages it uses │
│ MEMORY.md ← How it accesses the memory ladder │
│ COLLECTIVE.md ← Its shared evidence pool with other agents │
│ │
│ TOOLS.md ← Permitted tool surface │
│ SKILLS.md ← Loaded skill packs │
│ AGENTS.md ← Known peer agents │
│ USER.md ← User context & preferences │
│ BOOTSTRAP.md ← First-run initialization logic │
│ │
│ CORE.md ← Durable long-term memory (high friction) │
│ OVERVIEW.md ← Strategic snapshot of current priorities │
│ PROJECT.md ← Per-initiative auto-curated ledger │
└────────────────────────────────────────────────────────────────┘
│
▼
┌────────────────┐ Git tracks every change.
│ Kernel │ Any text editor works.
│ (deterministic │ CI can lint your agent.
│ file injection)│ Diff an agent update like a PR.
└────────────────┘
Why this matters: Prompt engineering is folk art. AgentDNA turns agent design into software engineering — testable, reviewable, version-controlled.
Four-tier hierarchical memory with biological confidence decay.
Flat context windows degrade. Long-term dumps bloat. clawREFORM implements a four-tier memory promotion system inspired by how working memory actually functions in cognitive science.
HIGH ╔══════════════════════════════════════════════╗
CONFIDENCE ║ TIER 1 → CORE.md ║
║ Permanent truths. Survives project resets. ║
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ ║ Requires multi-agent consensus to edit. ║
╠══════════════════════════════════════════════╣
║ TIER 2 → OVERVIEW.md ║
▓▓▓▓▓▓▓▓▓▓▓▓ ║ Strategic priorities. Updated on goal shift.║
╠══════════════════════════════════════════════╣
║ TIER 3 → PROJECT.md ║
▓▓▓▓▓▓▓▓ ║ Per-initiative auto-curated notes ledger. ║
╠══════════════════════════════════════════════╣
║ TIER 4 → COLLECTIVE.md + ledger.json ║
▓▓▓▓ ║ Cross-agent evidence pool. Decays naturally. ║
LOW ╚══════════════════════════════════════════════╝
CONFIDENCE
Claims promoted upward by:
evidence_score × recurrence × cross_source_correlation
Why this matters: Every other agent OS gives you one flat context. clawREFORM gives your agents the same hierarchical memory structure used by expert systems since 1985 — because flat context dumps don't scale.
Cross-agent knowledge that earns its way into long-term memory.
When multiple agents share a COLLECTIVE.md AgentDNA file, every observation they make is scored against a confidence model. High-confidence claims get automatically promoted up the Memory Ladder — with a full audit trail every human can review.
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
│ Agent A │ │ Agent B │ │ Agent C │ │ Agent D │
└────┬────┘ └────┬────┘ └────┬────┘ └────┬────┘
│ │ │ │
└─────────────┴──────┬──────┴─────────────┘
│ Claims + observations
┌────────▼──────────┐
│ COLLECTIVE.md │
│ Scoring Engine │
│ │
│ evidence × 0.4 │
│ recurrence × 0.4 │
│ cross-src × 0.2 │
└────────┬──────────┘
│
┌──────────────┼──────────────┐
▼ ▼ ▼
PROJECT.md OVERVIEW.md CORE.md
(threshold (threshold (threshold
≥ 0.40) ≥ 0.65) ≥ 0.85)
Why this matters: No other agent framework has a consensus mechanism for collective knowledge. This is the difference between a hive of agents and a swarm with shared intelligence.
The first open-source agent OS that modifies its own Rust source code.
clawREFORM ships a full code-modification pipeline. Tell it to add a feature in plain English. It writes the Rust, validates it, and commits — or rolls back automatically. Zero human in the loop required (though you can gate every step).
┌─────────────────────────────────────────────────────────────────────┐
│ S E L F - R E W R I T E P I P E L I N E │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ INPUT → Natural language request (any LLM) │
│ │
│ PLAN → AST-aware change plan with affected crates │
│ │
│ BACKUP → Atomic snapshot of every file to be modified │
│ │
│ RISK → Complexity score (blast radius · cyclomatic · deps) │
│ │
│ WRITE → LLM generates idiomatic Rust, respecting all lints │
│ │
│ VALIDATE → cargo build → cargo test → cargo clippy │
│ │ │ │
│ ▼ all pass ▼ any fail │
│ COMMIT → patch applied live ROLLBACK → snapshot │
│ │
└─────────────────────────────────────────────────────────────────────┘
$ clawreform chat "Add Prometheus histogram for LLM latency per provider"
→ Writes crates/clawreform-api/src/metrics.rs
→ Adds histogram registration to server.rs
→ All 1744 tests pass. Committed. Done.
Why this matters: Self-improvement is the final boss of agent autonomy. Every other framework offloads this to the user. clawREFORM eats its own tail.
clawREFORM sits at the top of a 4-tier hierarchy. Understanding where it sits explains why it exists.
┌────────────────────────────────────────────────────────────────────────┐
│ T H E A E G N T I C . A I H I E R A R C H Y │
│ │
│ TIER 0 · clawREFORM (CEO) │
│ Strategic brain. Learns your business. Directs divisions. │
│ Company Goals · Org Chart · Budget Limits · RAG │
│ │ │
│ TIER 1 · Agent-Zero (Senior Manager) │
│ Python-based hierarchical orchestrator in Docker. │
│ Owns a Division (Ops, Growth, R&D). Commands OpenClaw PMs │
│ │ │ │ │
│ TIER 2 · OpenClaw (Project Manager) │
│ Isolated Rust daemon per project. Own memory, skills, │
│ channels, gateway. Delegates to ZeroClaw staff. │
│ │────│────│ │
│ TIER 3 · ZeroClaw (Staff) ← 3.4 MB · <5 MB RAM · <10ms startup │
│ Pure execution muscle. Tool calls, file ops, browser, │
│ shell. Deny-by-default. Reports upstream. │
│ │
│ ───────────────────────────────────────────────────────────────── │
│ │
│ HOW CLAWREFORM FITS: │
│ clawREFORM IS the Tier 0 (CEO) orchestrator that governs the │
│ hierarchy. It provides the Company Dashboard (Goal Tracking, │
│ Org Charts, Budgets) to monitor all 4 tiers. │
└────────────────────────────────────────────────────────────────────────┘
| clawREFORM | OpenClaw | ZeroClaw | Agent-Zero | |
|---|---|---|---|---|
| Role in hierarchy | CEO + Dashboard | Project Manager | Staff Worker | Senior Manager |
| Language | Rust 🦀 | Rust 🦀 | Rust 🦀 | Python 🐍 |
| Binary size | ~8 MB | ~4 MB | 3.4 MB | Docker image |
| Cold start | ~200ms | ~80ms | <10ms | ~3–5s |
| RAM footprint | ~30 MB | ~15 MB | <5 MB | ~250 MB |
| Self-modification | ✅ Full pipeline | ❌ | ❌ | ❌ |
| AgentDNA system | ✅ 15 files | Partial (6) | ❌ | ❌ |
| Memory Ladder | ✅ 4-tier | ✅ 4-tier | ❌ | Flat |
| Collective Conscience | ✅ | ✅ | ❌ | ❌ |
| Web Dashboard | ✅ Alpine.js SPA | ❌ | ❌ | ✅ |
| Visual Workflow Builder | ✅ | ❌ | ❌ | ❌ |
| OpenAI-compatible API | ✅ | ❌ | ❌ | ❌ |
| MCP client + server | ✅ | ✅ client | ❌ | ✅ |
| A2A protocol | ✅ both | ❌ | ❌ | ❌ |
| Desktop app (Tauri) | ✅ | ❌ | ❌ | ❌ |
| Channel adapters | 40 | 8 | 0 | ~12 |
| LLM providers | 27 | 8 | 4 | ~10 |
| Bundled skills | 60 | 40 | 0 | ~25 |
| Test coverage | 1744+ | — | — | — |
CAPABILITY clawREFORM OpenDevin CrewAI LangGraph n8n Dify Cursor
─────────────────────────────────────────────────────────────────────────────────────
Agent OS (not a framework) ✅ ❌ ❌ ❌ ❌ ❌ ❌
File-native identity (AgentDNA)✅ ❌ ❌ ❌ ❌ ❌ ❌
4-tier memory + decay ✅ ❌ ❌ ❌ ❌ ❌ ❌
Self-rewrite of own source ✅ ❌ ❌ ❌ ❌ ❌ ❌
Collective cross-agent memory ✅ ❌ ❌ ❌ ❌ ❌ ❌
Written in Rust (single binary)✅ ❌ ❌ ❌ ❌ ❌ ✅
Zero runtime dependencies ✅ ❌ ❌ ❌ ❌ ❌ ❌
WASM sandboxed tool execution ✅ ❌ ❌ ❌ ❌ ❌ ❌
27+ LLM providers ✅ ~5 ~5 ~5 ~10 ~15 ~8
40+ communication channels ✅ ❌ ❌ ❌ ✅ ❌ ❌
A2A protocol (Google standard) ✅ ❌ ❌ ❌ ❌ ❌ ❌
MCP client + server ✅ client ❌ client ❌ ❌ client
OpenAI-compatible API ✅ ❌ ❌ ❌ ❌ ✅ ❌
Visual workflow builder ✅ ❌ ❌ ✅ ✅ ✅ ❌
Native desktop app ✅ ❌ ❌ ❌ ❌ ❌ ✅
Ed25519 signed manifests ✅ ❌ ❌ ❌ ❌ ❌ ❌
Merkle audit trail ✅ ❌ ❌ ❌ ❌ ❌ ❌
RBAC (4 roles) ✅ ❌ ❌ ❌ ✅ ✅ ❌
Prometheus /metrics endpoint ✅ ❌ ❌ ❌ ✅ ❌ ❌
Config hot-reload (no restart) ✅ ❌ ❌ ❌ ❌ ❌ ❌
─────────────────────────────────────────────────────────────────────────────────────
Novel primitives 4 0 0 0 0 0 0
PROVIDER clawREFORM CrewAI LangChain Dify n8n
─────────────────────────────────────────────────────────────────────────────────
OpenRouter (141 models) ✅ ❌ ✅ ✅ ❌
Anthropic ✅ ✅ ✅ ✅ ✅
Google Gemini ✅ ✅ ✅ ✅ ✅
OpenAI ✅ ✅ ✅ ✅ ✅
Groq (fastest inference) ✅ ❌ ✅ ✅ ❌
Cerebras (fastest chips) ✅ ❌ ❌ ❌ ❌
SambaNova ✅ ❌ ❌ ❌ ❌
DeepSeek ✅ ❌ ✅ ✅ ❌
Together AI ✅ ❌ ✅ ❌ ❌
Fireworks ✅ ❌ ✅ ❌ ❌
Mistral ✅ ✅ ✅ ✅ ❌
Perplexity ✅ ❌ ✅ ❌ ❌
xAI (Grok) ✅ ❌ ✅ ❌ ❌
Cohere ✅ ❌ ✅ ✅ ❌
AI21 ✅ ❌ ✅ ❌ ❌
Ollama (self-hosted) ✅ ✅ ✅ ✅ ❌
vLLM ✅ ❌ ✅ ❌ ❌
LM Studio ✅ ❌ ✅ ❌ ❌
HuggingFace ✅ ❌ ✅ ✅ ❌
Replicate ✅ ❌ ✅ ❌ ❌
─────────────────────────────────────────────────────────────────────────────────
TOTAL PROVIDERS 27 ~5 ~20 ~15 ~5
TOTAL MODELS (catalogued) 141+ — — — —
flowchart LR
U["👤 You"] --> CLI["CLI / TUI\n14+ subcommands"]
U --> WEB["Web Dashboard\n(:4332 · Alpine.js SPA)"]
U --> DESK["Tauri 2.0\nDesktop App"]
WEB --> API["REST / WS / SSE\naxum 0.8 · 100+ endpoints"]
DESK --> API
CLI --> K["🧠 Tier 0 (CEO)\nOrchestrator"]
API --> K
K --> DNA["AgentDNA Injector\n15-file deterministic stack"]
DNA --> LLM["LLM Router\n27 providers · 141 models\nfallback · cost metering"]
K --> MCP["MCP Client + Server\n23+ servers · JSON-RPC 2.0"]
K --> CH["40 Channel Adapters\nTelegram · Discord · Slack…"]
K --> WF["Workflow Engine\nfan-out · loops · conditionals"]
K --> SM["Self-Rewrite Engine\nbuild → test → clippy → commit"]
K --> MEM["Memory Ladder"]
K --> A2A["A2A Protocol\n(Google standard)"]
K --> WASM["WASM Sandbox\nfuel + epoch metering"]
MEM --> C1["CORE.md"]
MEM --> C2["OVERVIEW.md"]
MEM --> C3["PROJECT.md"]
MEM --> C4["COLLECTIVE.md\n+ ledger.json"]
flowchart TD
A["clawreform-cli\nclawreform-desktop"] --> B["clawreform-api"]
B --> C["clawreform-kernel"]
C --> D["clawreform-runtime"]
C --> E["clawreform-channels"]
C --> F["clawreform-wire"]
C --> G["clawreform-skills"]
C --> H["clawreform-migrate"]
C --> I["clawreform-extensions"]
C --> J["clawreform-hands"]
D --> K["clawreform-memory"]
E --> K
K --> L["clawreform-types"]
flowchart TD
A["Agent observations\n+ session summaries"] --> B["Collective Claim\nExtraction Engine"]
B --> C["Score:\nevidence × recurrence × cross-source"]
C --> D["COLLECTIVE.md\n+ ledger.json"]
C --> E{"Promotion threshold?"}
E -- "≥ 0.40" --> F["PROJECT.md"]
E -- "≥ 0.65" --> G["OVERVIEW.md"]
E -- "≥ 0.85" --> H["CORE.md"]
F --> G
G --> H
LAYER MECHANISM PURPOSE
────────────────────────────────────────────────────────────────────────────
Signing Ed25519 signed agent manifests Tamper-evident provenance
Audit Merkle hash chain Every change verifiably chained
Taint Information flow tracking Secrets don't bleed into logs
Network HMAC-SHA256 mutual auth (P2P wire) Peer identity verification
Sandbox WASM fuel + epoch + watchdog thread Tool execution with hard limits
Auth GCRA rate limiter, cost-aware Prevents token farming attacks
SSRF Private IP + metadata endpoint block Prevents cloud metadata exfil
Headers CSP · X-Frame-Options · HSTS Browser-layer hardening
Secrets Argon2 · AES-256-GCM · zeroize drop Keys purged on scope exit
Paths Path traversal guards on all 41 tools No directory escape
Roles RBAC: Owner / Admin / User / Viewer Principle of least privilege
Sessions SHA256 loop guard + circuit breaker Infinite loop prevention
────────────────────────────────────────────────────────────────────────────
CVEs: 0 known · Tests: 1744+ green · Clippy: 0 warnings
# macOS / Linux
curl -fsSL https://clawreform.sh/install | sh
# Windows PowerShell
irm https://clawreform.sh/install.ps1 | iex
# npm global (Linux/macOS/Windows)
npm install -g clawreform
# Start the daemon + open dashboard
clawreform start
# → http://127.0.0.1:4332# Run the 6-step setup wizard (picks up env vars automatically)
clawreform init
# Send your first message
clawreform chat "Summarise everything in this repo"
# Ask it to improve itself
clawreform chat "Add a /api/health endpoint that returns uptime and version" WHAT YOU GET OUT OF THE BOX
────────────────────────────────────────────────────────────────────────────
AgentDNA Templates 30 pre-built agent archetypes (analyst · coder · devops…)
LLM Drivers (3) Anthropic · Gemini · OpenAI-compatible (covers all 27)
Model Catalog 141+ models, tier-classified, cost-per-token indexed
Skills (60) 14 categories: research · code · browser · data · media
Hands (7) Browser · Clip · Lead · Collector · Predictor · Researcher · X
Channels (40) Telegram · Discord · Slack · WhatsApp · Email · Matrix…
Tools (41) Filesystem · Web · Shell · Browser · TTS · Vision · Inter-agent
MCP Servers (23+) GitHub · Linear · Supabase · Slack · aegntic/cldcde
API Endpoints (100+) REST / WS / SSE · OpenAI /v1 · Google A2A · Prometheus
SDKs (3) Python client · Python agent SDK · TypeScript/JavaScript SDK
CLI Commands (14+) init · start · agent · workflow · skill · channel · mcp · doctor
Desktop App Tauri 2.0 · system tray · single-instance · hide-to-tray
Visual Builder Drag-and-drop workflow canvas · 7 node types · TOML export
────────────────────────────────────────────────────────────────────────────
MESSAGING PROFESSIONAL SOCIAL SELF-HOSTED
─────────────────────────────────────────────────────────────────────────
Telegram Slack Twitter / X Matrix
WhatsApp Microsoft Teams Mastodon Mattermost
Signal Google Chat Bluesky IRC
LINE Webex Reddit XMPP
Viber Feishu / Lark LinkedIn Custom HTTP
Facebook Messenger Zoom (webhook) Twitch
─────────────────────────────────────────────────────────────────────────
40 adapters total · unified bridge · per-channel RBAC · graceful shutdown
| Standard | Mode | What it enables |
|---|---|---|
OpenAI API /v1/chat/completions |
Server | Drop clawREFORM behind any OpenAI client, IDE, or app |
| MCP (JSON-RPC 2.0) | Both | Connect to 23+ tools as client; expose agent tools as server |
| Google A2A | Both | Delegate tasks to external A2A agents; accept inbound tasks |
| OpenClaw YAML/JSON5 | Importer | Migrate existing OpenClaw agents with zero data loss |
Prometheus /api/metrics |
Export | Plug into Grafana, Datadog, or any monitoring stack |
| Tailscale Mesh | Network | Secure P2P agent networking across all your devices |
clawreform/
├── crates/ # 13 Rust workspace crates
│ ├── clawreform-types # Domain types, config schema, trait definitions
│ ├── clawreform-memory # SQLite memory · semantic recall · vector embeddings
│ ├── clawreform-runtime # LLM drivers · WASM sandbox · tool runner (41 tools)
│ ├── clawreform-kernel # Orchestrator · Self-Rewrite Engine · event bus
│ ├── clawreform-api # axum HTTP/WS/SSE · 100+ endpoints · OpenAI compat
│ ├── clawreform-cli # 14+ subcommands · TUI · shell completions
│ ├── clawreform-channels # 40 channel adapters · unified bridge · RBAC
│ ├── clawreform-wire # P2P wire protocol · HMAC-SHA256 auth · OFP
│ ├── clawreform-skills # 60 bundled skills · ClawHub marketplace · SHA256 verify
│ ├── clawreform-hands # 7 Hands packages for autonomous action
│ ├── clawreform-desktop # Tauri 2.0 native app · system tray
│ ├── clawreform-migrate # OpenClaw migration engine (YAML/JSON5 → TOML)
│ └── clawreform-extensions # Extension / plugin API
├── agents/ # 30 pre-built AgentDNA template stacks
├── sdk/ # Python + JavaScript/TypeScript SDKs
├── docs/ # Full documentation (18 guides)
└── pitch-deck/ # PRD · landing page · investor materials
| Metric | Value | Notes |
|---|---|---|
| Test suite | 1744+ tests | 0 failures, 0 clippy warnings, CI-enforced |
| Codebase | 146K SLOC Rust | No Python bloat, no Node sprawl |
| Binary | ~8 MB | LTO + O3 + symbols stripped |
| API cold start | <200ms | Tokio async, no GC pauses |
| API p99 latency | <5ms | axum on Hyper, zero-copy where possible |
| Context budget | 1M tokens/hr | Per-agent, configurable |
| Compaction trigger | 70% capacity | LLM-summarised, keeps last 20 messages |
| Tool result cap | 15K chars | Prevents context bloat in tool-heavy agents |
| Config hot-reload | 30s polling | No restart required for any config change |
| Platforms | Linux · macOS · Windows | Pre-built release binaries |
| Decision | Rationale |
|---|---|
| Rust, not Python | Every other agent OS is Python. Python has a GC, a GIL, and runtime dependencies. Rust has none of those. clawREFORM ships a single static binary that runs everywhere — no venv, no pip, no version hell. |
| AgentDNA files, not prompt strings | Prompt strings aren't testable, versionable, or reviewable. Markdown files are. git diff your agent's soul. grep its memory. sed its personality. Software engineering practices, for agents. |
| Memory Ladder, not flat context | LLMs have finite context windows. Flat RAG doesn't prioritise. A 4-tier hierarchy with confidence decay ensures important facts survive and noise fades — mimicking how expert cognition actually works. |
| WASM sandbox, not subprocess exec | subprocess.run() is a footgun. WASM's fuel+epoch metering gives deterministic, sandboxed tool execution with hard resource limits — without spawning an OS process per tool call. |
| SQLite, not Postgres | Agents need memory that survives laptop hibernation, not a database server that goes down at 3am. SQLite is embedded, zero-ops, and transactional. |
| A2A + MCP both | Don't bet on one interoperability standard. Implement both and let agents talk to anything in the emerging ecosystem. |
| Self-rewrite in Rust | If the system can improve itself, it should improve the actual binary — not just its prompt. A Rust self-rewrite pipeline with build/test/clippy validation is the honest version of "self-improvement". |
| Getting Started | Install, init, first agent in 5 minutes |
| AgentDNA™ System | Designing agents with file stacks |
| Architecture | 13-crate deep dive |
| Providers | All 27 LLM providers + smart routing |
| Security | Full security model reference |
| API Reference | 100+ REST/WS/SSE endpoints |
| Channel Adapters | 40 channels configuration |
| Workflows | Fan-out, conditionals, cron loops |
| CLI Reference | All 14+ subcommands |
| Configuration | Every config field documented |
- Skool — AutoClaw Swarm — the community for aegntic.ai ecosystem builders
- Twitter / X — release notes, ecosystem updates
- GitHub Discussions — architecture RFCs, Q&A
- CONTRIBUTING.md — how to get your PR in
Dual-licensed under MIT or Apache 2.0 — choose whichever works for your project.
Built by aegntic.ai research for ae.ltd
The agent operating system the ecosystem has been waiting for.