DuDuClaw Feature Highlights
DuDuClaw v1.62.0 | Last updated: 2026-09-04
This directory contains detailed introductions to DuDuClaw’s standout features. Each article explains the design rationale, system behavior, and operational flow — aimed at developers who want to understand how things work without diving into source code.
Feature Index
Section titled “Feature Index”| # | Article | One-liner |
|---|---|---|
| 1 | Prediction-Driven Evolution | 90% of conversations evolve at zero LLM cost |
| 3 | Confidence Router & Local Inference | Smart model selection that saves 80%+ on API bills |
| 4 | File-Based IPC Message Bus | Structured inter-agent delegation with TaskSpec workflows |
| 5 | Security Defense | Four live guards, where each runs, and what none of them covers |
| 7 | Multi-Account Rotation | Cross-provider credential scheduling with failover |
| 8 | Browser Automation and Computer Use | Two fetch tools and an optional browser server the agent picks between, plus computer-use sessions driven through eight computer_* MCP tools; no auto-router |
| 9 | Behavioral Contracts & Red-Team Testing | Machine-enforceable agent boundaries |
| 10 | Cognitive Memory System | Human-inspired memory with forgetting curves |
| 11 | Prompt Budget Enforcement | Estimate the prompt, walk three stages, or refuse the request |
| 12 | Industry Templates & Odoo ERP Bridge | Out-of-the-box business intelligence |
| 13 | Multi-Runtime Agent Execution | Claude / Codex / Antigravity / Grok / OpenAI-compat and more behind one backend (Gemini CLI removed) |
| 14 | Voice Pipeline | STT/TTS over two paths — a fail-closed HTTP pair and a hardcoded Telegram handler |
| 15 | Skill Lifecycle Engine | 6-stage automated skill extraction and management |
| 16 | Session Memory Stack | Pinned instructions + snowball recap + key-fact accumulator |
| 17 | Wiki Knowledge Layer | L0-L3 trust-weighted knowledge with auto-injection |
| 19 | Agent Client Protocol (ACP/A2A) | duduclaw acp speaks Agent Client Protocol v1 for IDE panels (Zed / JetBrains / nvim); duduclaw acp server is the A2A stdio surface |
| 20 | Memory Intelligence | Temporal facts + reflexion loop + batch fetch |
| 23 | Autopilot Rule Engine | Event-driven automation + circuit breaker |
| 24 | Task Board & Activity Feed | Agent-as-teammate task management |
| 25 | Identity Resolution | WikiCache / Notion / Chained providers (RFC-21 §1) |
| 26 | MCP HTTP/SSE Transport | Bearer-authed REST + SSE endpoints (W20) |
| 27 | One-shot PTY invocation | Give a CLI a real terminal; the session pool was removed in 2026-09 |
| 28 | Live Run Forking | Parallel branches + AI judge; filtered copies, retained branch adoption and test timeout cleanup |
| 29 | Evolution Events | Black-box recorder with batch+retry delivery |
| 30 | Custom Dashboard Widgets | AI-guided or raw-HTML dashboard cards in a sandbox |
| 31 | Office Document Suite | Real docx/xlsx/pptx/pdf output with the DELIVER protocol, archive and preview |
| 32 | Expert Packs | Installable AI teams: built-in catalog, LLM-guided authoring, department × rank org placement |
| 33 | OS-Native Perception & Proactive Care | File watch + frontmost sensing → footprint memory, care checks, one-click automations |
| 34 | Autonomous Goal Loop | /goal → MAV acceptance, recorded round state and costs; weekly survival report shows sample limits |
| 35 | Photo → Desktop Pet | Local photo-to-pixel-pet pipeline with a Codex-Pets spritesheet and wander engine |
| 36 | Recording → Skill | Browser/desktop recordings distilled into approval-gated SKILL.md drafts |
| 37 | Delegation Isolation | Org-boundary delegation policy: hierarchy / department / white-list enforcement |
| 38 | Agentic Evolution Engine & Playbook | SOUL.md becomes a read-only persona layer; the playbook is what learns and can be retired rule by rule |
| 39 | Calibrated Forward Model & Held-Out Learning Gate | Every guess gets scored against reality; self-derived lessons stay on the bench until the numbers back them |
| 40 | Notification Governance | Severity levels, quiet-hour deferral, one digest a day, and action-rate metrics for every push |
| 41 | Resident Sensing & Signal Wake-Up | External data streams wake the AI employee only when a rule actually fires |
| 42 | Human Takeover | An admin typing directly in a channel silences the AI for that one conversation until they hand it back |
| 43 | Telegram Mini App Approval Card | Read the full approval and decide inside Telegram, no dashboard switch (spike, off by default) |
| 44 | Working State | One authoritative cross-wake record of each AI employee’s current rules, changed only through audited tool calls |
| 45 | Local Model Marketplace | Pick a model by purpose, see whether it fits this machine, install in one click |
| 46 | Belief Loop | State a prediction about the outside world, get scored against reality, see your own calibration next time |
| 47 | Agent Mail | Per-agent inbox for incoming mail; nothing goes out until a person confirms it |
| 48 | Goal Intent Router | Chat channels notice task delegation and offer to create a goal; never auto-created |
| 49 | Code Security Audit | duduclaw secaudit: static scanners + AI deep audit + adversarial review + sandboxed PoC |
| 50 | DuDuClaw OS Appliance | Bootable appliance image — LAN dashboard onboarding, device page, sysd privilege separation, webhook relay |
| 51 | OS Keyboard Shortcuts | Every DuDuClaw OS shortcut — global compositor bindings, shell UI, first-run setup, lock screen |
| 52 | DuDuClaw OS Desktop Edition | One machine shared by a person and the AI — shadow workspace, human input always wins, explicit hand-back, off-by-default co-driving |
| 53 | Local Models on the Device | Six verified GGUFs, one click to download and one to switch on; hybrid by default, honest about speed |
| 54 | Fine-tuning and Post-training | Curate the dataset here, train it on a GPU elsewhere, import the GGUF/LoRA back — this machine never trains |
| 55 | Data Sources & Native DB Connector | A registry any db_field rule can bind to, an MCP proxy that redacts a customer’s own MCP servers, and a first-party read-only PostgreSQL/MySQL/SQLite connector |
| 56 | Team as Employee (Team-as-Agent) | 規劃／執行／審核／合成 four roles inside one employee, each on its own runtime and model — the switch is on by default but naming a second vendor under [team.roles] is what forms a team, a decomposability gate decides per task, and roles exchange a TaskPacket instead of a transcript |
| 57 | UCCI Calibrated Cascade | Experimental, opt-in local-routing tier — an isotonic-calibrated router decides LocalFast→LocalStrong→cloud escalation from token-margin uncertainty instead of the legacy post-hoc confidence gate; fitted offline, off by default |
| 58 | Night Engine | Idle-window memory tidy-up: four sub-passes (two deterministic, two on the utility model), a per-pass spend cap and a daily circuit breaker, behind two separate opt-in switches; productive passes appear in Activity Feed |
| 59 | Local Proxy | duduclaw proxy — an OpenAI-compatible localhost endpoint so Aider / Cline / Codex can borrow the account pool; Bearer-gated, loopback by default, honest about OAuth seats not being forwardable |
| 60 | Discovery | Approved-workspace exploration, budgeted tree records, verified artifacts and zero-LLM held-out policy comparisons; integration validation in progress |
| 61 | Odoo website live chat | AI employees answer Odoo Live Chat visitors as an operator; audience rules pick the employee per visitor; human takeover and handoff |
| 62 | Calling AI staff with @ |
Hand one message to another AI employee by @name, limited to the employees the sender may call; @ list in WebChat, LINE and Telegram; one-shot call; mention-only groups |
| 63 | Document knowledge base | Collections of company documents with per-collection permissions (users, departments, AI employees, channels), sandboxed extraction, keyword search with page / row locations, read-only knowledge_* MCP tools, duduclaw knowledge eval |
| 64 | Thinking display | Live thinking timer in WebChat for every runtime; masked, shortened thinking summaries for signed-in dashboard users; never sent to messaging apps |
| 65 | Where thinking comes from | Which runtimes report thinking (Claude CLI, Codex, OpenAI-compatible APIs, Anthropic Direct API), what each shows, and when a request changes |
| 66 | Thinking display in messaging apps | Timer and masked summary in Telegram, Discord, Slack, Teams, LINE and the other apps; [thinking] channel_audience / channel_group_display |
| 67 | Assigning tasks from a chat | /交辦 @employee task (/assign) creates a goal task on the sender’s behalf, limited to employees they may assign to; progress back to the chat; /交辦 狀態; “Assigned by me” on the Goals page |
| 68 | Read-only team conversations | Unreleased Enterprise backend: live reporting-line scope, full-text report WebChat, masked external transcripts, activation boundary, mandatory audit and limits; 繁體中文 / 日本語 |
Companion Articles
Section titled “Companion Articles”| Article | One-liner |
|---|---|
| Live Forking: When to Use It | Usage-scenario companion to #28 — when to fork, when not to, and how it differs from duduclaw eval |
| ERP / CRM Support Matrix | One-page coverage table for sales and customer conversations |
Translations
Section titled “Translations”Full Feature Inventory
Section titled “Full Feature Inventory”For a complete list of all features (not just highlights), see feature-inventory.md.