✨ Clawpy version 3.12 is live

Outcomes with fewer prompts|

Seven memory layers. Ebbinghaus-inspired decay. Temporal knowledge graphs. A 5-tier wisdom cascade. Budget-gated heartbeat. Self-generating workflows. Deep engineering for autonomous intelligence.

Clawpy Swarm Configurator hierarchy with Alfred assistant panel

Compatible with leading foundation models

GEMINICLAUDEOPENAIQWENDEEPSEEKZ.AIKIMIOPEN SOURCE MODELS

Architectural Innovation at Every Layer

Not a wrapper. Not a chatbot. A deeply engineered cognitive operating system for autonomous agent swarms.

🧠

7-Layer Cognitive Memory

Markdown flat files → SQLite event ledger → ChromaDB vectors → Auto-capture/recall → Cross-session context engine → Temporal knowledge graph → PARA canonical knowledge. Each layer has its own storage engine, retention policy, and query API. Ebbinghaus-inspired forgetting curve extends memory life by 20% per recall.

score = λ × semantic + (1−λ) × exp(−age/τ)
→ Read the deep dive →
👑

Swarm Hierarchy

5-tier operational hierarchy: CEO → Senior Leaders → Leaders → Workers → Junior Workers. Role-driven reporting lines with auto-generated org maps and hierarchy-aware task escalation.

Workspace → SwarmSpace → Team → Agent
→ See the architecture →
🔐

7-Layer Security Stack

Immutable safety core → Memory injection guard → SHA-256 intent cipher → Two-tier Guardian scanner → Action gate with session-scoped approvals → 26-rule skill scanner → Docker DooD sandbox. Defense in depth — not a single fence.

Layer 1 → 2 → 3 → 4 → 5 → 6 → 7
→ Full security deep dive →
💰

Autonomous Budget Governance

Per-agent token budgets with soft-warn at 80% and hard-stop at 100%. LLM-gated heartbeat evaluates before acting. Cost-capped validation loops prevent runaway retries.

if observed ≥ limit → hard_stop → agent paused
→ Budget architecture →
🔄

Self-Learning Architecture

Agents detect their own repeated tool-call patterns and auto-generate deterministic workflows, replacing LLM reasoning with cheap execution. The Introspection Loop evaluates every N tool calls. The Adaptation Engine promotes learnings into prompt fragments, routing hints, and validator policy tweaks. The Wisdom Cascade flows knowledge bidirectionally through the hierarchy.

detect → score → corroborate → promote → evolve
→ Full adaptation pipeline →

Meet Your Specialized Swarm

Assemble the exact combination of experts needed for your project.

📐

The Architect

Designs system specs and handles high-level structural planning.

🔨

The Builder

Translates specs into clean, functional code via reasoning models.

⚙️

The Implementor

Executes deep tasks, deployment scripts, and drives the terminal.

🔍

The Auditor

Reviews pull requests, memory logs, and validates boundaries.

Powerful Under the Hood. Simple on the Surface.

Full transparency and total control — from supervised approval to full autonomy — with a single setting.

🎯

Intelligent Model Routing

Leaders get expensive models for strategic reasoning. Juniors get cheap models for repetitive work. Automatic, based on archetype — not manual configuration.

CEO → Claude Opus  💰💰💰
Worker → Moonshot K2.5  💰
Junior → Claude Haiku  💰
🎚️

4 Autonomy Presets

Go from fully supervised to full Mad Dog with a single switch. Scope autonomy per workspace, per agent, or per task — 6 scoping levels.

Safe → Execution → Governance → Maximum
Scope: Global | Workspace | Agent | Task
📈

Self-Learning Pipeline

Agents introspect their own performance, detect repeated patterns, and autonomously generate deterministic workflows — replacing expensive LLM reasoning with cheap execution.

introspect → adapt → offload → evolve
⬆️

Automatic Escalation

Workers that stall are force-paused and escalated to their lead. Leads escalate to the CEO. Problems find the right authority automatically — no babysitting.

Worker stalls → Lead notified → CEO alerted
Budget exceeded → Agent paused → Owner

How Clawpy Compares

Not all agent frameworks are built equal. Here's what ships out of the box.

FeatureClawpyOpenClawHermesAgent ZeroPaperclip
Dual contact points
Alfred + Lucius
Single
Single
Single
⚠️CEO only
Archetype classification
3-stage auto
Manual
None
None
⚠️User roles
Autonomy presets
4 presets
Per-scope autonomy
6 levels
Tiered model routing
Automatic
⚠️Manual
⚠️Single
⚠️Per-agent
Hierarchical delegation
Tier-enforced
⚠️Basic
⚠️Basic
Org chart
Escalation chains
Per-agent
⚠️Reporting
Budget enforcement
Auto-pause
⚠️Guidance
Auto-throttle
Stall detection
Auto-kill
⚠️Heartbeat
Self-learning
3 systems
⚠️Basic
7-layer memory
⚠️Basic
⚠️Basic
⚠️
⚠️Basic
Knowledge graph
Time-decay
⚠️Basic
Security layers
7 layers
⚠️Sandbox
⚠️Basic
⚠️Budget
Intent cipher
SHA-256
Memory injection guard
11 patterns
⚠️Basic
Skill security scanner
26 rules
⚠️Basic
Command approval gate
7 categories
⚠️Basic
⚠️Board approval
Dashboard GUI
Full
⚠️Logs
CLI
⚠️Basic
React