前面做过一期评论,当时的给我的体验确实不怎么好,一个月过去了,看看这个项目还有没希望,有没有必要二刷。
导语
2026 年 4 月 21 日,乐鑫正式开源边缘 AI Agent 框架 ESP-Claw(龙虾),距今近三个月。打开官方仓库espressif/esp-claw,社区数据持续走高:1799 个 Star、370 个 Fork、47 位贡献者、70 + 合并 PR、累计 853 次代码提交。
短短三个月,项目完成三大关键里程碑:
-
4 月快速搭建原型,官宣开源;
-
6 月 12 日首个正式版 v0.1.0 发布,纳入 20 份社区 PR;
-
同步启动
rust-migration分支,启动 C 核心向 Rust 完整重构,截至 7 月 10 日迁移分支已有 751 次提交、8 大 Rust 模块。
一、ESP-Claw 乐鑫边缘 Chat Coding 智能体框架
1. 核心定义
ESP-Claw 是乐鑫面向 ESP32 全系列芯片打造的边缘 AI Agent 框架,核心创新为「Chat Coding 聊天造物」:用户通过微信、QQ、飞书、Telegram 等 IM 软件和设备对话,设备自动调用 LLM 解析自然语言,生成可直接运行的 Lua 脚本,无需编译烧录,就能自定义硬件控制逻辑。
传统 IoT 设备只能提前写死功能,而 ESP-Claw 把物联网设备从 “被动执行终端” 升级为本地自主决策智能节点,支持断网本地规则自治、多设备 MCP 互联互通。
2. 核心基本面
| 项目信息 | 详细参数 |
|---|---|
| 开源仓库 | espressif/esp-claw,Apache-2.0 协议 |
| 技术占比 | C 语言 86.1%、Lua5.5%、TypeScript5.4% |
| 编译构建 | ESP-IDF 5.5.4 / 6 + CMake |
| 支持芯片 | ESP32-S3 / C5 / P4 / C6;最低 8MB Flash+8MB PSRAM,推荐 16MB Flash+8MB PSRAM |
| LLM 兼容 | OpenAI、Qwen、Claude、DeepSeek、Anthropic,支持自定义私有接口 |
| 交互渠道 | Telegram/QQ/ 微信 / 飞书 / Web 可视化后台,可拓展 |
| GitHub 数据 | Star1799|贡献者 47|PR70+|总提交 853 次 |
3. 四层核心架构
项目采用分层解耦架构,各模块可独立裁剪、按需拓展:
二、六大阶段完整演进时间线
基于 853 次提交完整梳理,项目从 2026.04.03 首个初始化提交,到 7 月 10 日形成清晰六阶段迭代轨迹,日均提交 7.5 次,单日最高 33 次。
阶段一:项目初创期 04.03–04.10
4 月 3 日以初始化提交开启项目,一周内完成核心骨架搭建:
-
落地初代 Agent Loop 运行框架;
-
拆分 Lua 独立组件、对接飞书与 Anthropic 大模型;
-
开发任务调度器、会话持久化、硬件能力路由;
-
任务栈迁移 PSRAM,优化嵌入式内存布局。
阶段二:功能爆发期 04.11–04.30
4 月 21 日官方正式开源发布,是功能密度最高的周期,完成首次大型架构拆分:
-
新增板级管理器 Board Manager、V2 结构化长期记忆系统;
-
重构 Lua 脚本路径、重构整机事件输出规范;
-
单体
claw_core首次拆分为多内部模块,从原型走向工程化; -
搭建官方文档站,配套 GitHub Pages 自动化 CI 流水线。
阶段三:架构优化 + 社区涌入期 05.01–05.31
重心从 “快速堆功能” 转向架构标准化,大量第三方开发者提交板卡适配 PR:
-
5 月 7 日重构 Lua 硬件驱动,统一初始化 / 销毁接口,外设拓展标准化;
-
新增 LVGL 图形 UI Lua 模块,支持屏幕可视化开发;
-
上线 Agent Manager,支持多智能体并发、子任务追踪;
-
社区贡献 15 + 款开发板适配:M5Stack、LilyGO、DFRobot 等。
阶段四:稳定打磨,v0.1.0 正式发布 06.01–06.12
全量修复安全、内存、网络问题,6 月 12 日 v0.1.0 正式上线,合并 20 条社区 PR:
稳定性修复
WiFi 断线自动重连、开放 WiFi 网络兼容、LLM 乱码 UTF-8 净化、字符边界安全处理、内存泄漏与野指针修复;
全新能力
Lua 异步执行机制、本地 Plan Mode 技能、视觉检测能力、MCP 服务节点、Web 在线刷机工作流。
阶段五:Rust 核心迁移分支启动 06.09–持续迭代
与正式版发布同步创建rust-migration实验分支,目标将 C 语言核心claw_core完整重写为 Rust,规避嵌入式内存安全问题。
-
独立提交 151 次,搭建完整 8 大 Rust Crate 体系;
-
7 月 2 日增加异步并发运行时;7 月 6 日整套框架可完整运行;
-
通过
claw-cabi兼容层实现 Rust 与原有 C 代码双向调用,支持渐进式迁移,无需一次性重构全项目。
Rust Crate 体系架构
| Crate 模块 | 核心职责 |
|---|---|
| claw-core | 智能体运行时、消息通道、文件 / HTTP 底层抽象 |
| claw-agent | 智能体生命周期、对话编排、多子 Agent 管理 |
| claw-skill / claw-tool | 硬件技能注册、工具调用执行框架 |
| claw-capability | 硬件能力权限、外设注册管理 |
| claw-sandbox | Lua 脚本安全沙箱隔离 |
| claw-memory | 本地长期记忆存储 |
| claw-cabi | C 语言 ABI 兼容桥接层(迁移核心) |
阶段六:生态持续拓展 06.23–07.10
主分支 master 同步迭代,聚焦硬件适配与开发体验优化:
-
anyboard通用板卡方案,无需定制底层 BSP; -
Lua LVGL 网页模拟器,浏览器实时预览屏幕界面;
-
新增微雪、Seeed、Rockbase 等多款开发板支持;
-
开放 ESP-NOW、CAN/TWAI、MQTT 通信协议 PR;
-
文档站升级 Astro v7,完善新手分层教程。
三、仓库分支结构与提交数据统计
1. 三大分支现状
| 分支名称 | 总提交数 | 核心说明 |
|---|---|---|
| master | 702 | 主开发分支,日常迭代、版本发布主线 |
| product/buddy | 669 | 6 月 22 日 master 快照,回退 Buddy Ditto 实验功能,无独立新增提交 |
| rust-migration | 751 | Rust 重构专属分支,151 份独有提交,当前最活跃特性分支 |
2. 社区 PR 主题分布条形图
社区外部贡献 60% 集中在开发板硬件适配,剩余为 Web 交互、LLM 优化、IM 平台修复、Lua 驱动拓展,反映开发者核心诉求:让框架适配更多硬件设备(都来抱大腿的,小白玩得动吗)。
四、五次决定性架构重构
-
4.29 claw_core 首次拆分 初代单体核心包揽所有逻辑,代码臃肿难以维护。拆分后建立模块化边界,是项目从原型转向工程化的分水岭。
-
5.07 Lua 硬件驱动标准化 零散脚本统一为标准化驱动接口,I2C、UART、ADC、USB 等外设快速接入,大幅降低硬件拓展成本。
-
5.27 Agent Manager 多实例重构 支持同时运行多个智能体,子任务独立追踪进度,为多设备协同、并发控制打下底层基础。
-
6 月初 v0.1.0 稳定性重构 聚焦内存安全、字符编码、网络容错,完成从 “功能能用” 到 “稳定可靠” 的转变。
-
6.09 Rust 核心迁移工程启动 项目最具前瞻性调整:用 Rust 解决 C 语言内存溢出、野指针、并发崩溃等嵌入式痛点,依靠 CABI 桥接实现平滑过渡。

五、测评五大核心痛点复盘
2026 年 6 月 3 日 本号发表的《野评 ESP-Claw》指出五大结构性硬伤,结合 7 月最新代码,逐条对比改善情况:
痛点 1:LLM 生成 Lua 链路延迟高、脚本容错差 ⭐⭐⭐⭐
原有问题:IM-LLM - 脚本 - 硬件完整链路最高延迟 7 秒,LLM 代码易出现引脚、内存错误,无校验机制。 ✅ 已优化:
-
底层执行优先级固化:本地路由规则优先执行,高频指令百毫秒响应;
-
Lua 脚本写入前沙盒预校验,拦截非法硬件调用;
-
成功逻辑自动缓存规则,同类指令无需重复调用大模型;
-
v0.1.0 新增 Lua 异步任务 API,并行执行不阻塞设备。 ⚠️ 仍存短板:全新复杂自定义逻辑仍需完整 LLM 链路,网络波动延迟无法根除;嵌入式 Lua 栈限制超长脚本运行。
痛点 2:硬件资源门槛高,8MB 低配设备体验差 ⭐⭐⭐
原有问题:8MB Flash 硬件开启全功能极易内存溢出,核心任务栈占用大量 RAM。 ✅ 已优化:
-
全组件 Kconfig 模块化裁剪,极简配置最低仅需 2MB Flash;
-
PSRAM 虚拟文件系统 RAMfs,减少 Flash 读写损耗;
-
新增内存监控 CLI,实时查看堆、缓存占用。 ⚠️ 仍存短板:完整功能套件硬性要求 16MB Flash+8MB PSRAM,低配硬件只能运行精简子集。
痛点 3:过度依赖云端 LLM,本地事件路由能力薄弱 ⭐⭐⭐⭐⭐(改善最彻底)
原有问题:简单开关、定时等基础指令也要调用大模型,断网设备直接失效。 ✅ 底层逻辑彻底重构:
-
系统强制本地规则优先,无匹配逻辑才请求 LLM;
-
LLM 生成逻辑自动存入本地规则库,支持网页手动编辑;
-
官方宣传重心转向「断网自治、本地毫秒响应」,解决 IoT 致命缺陷。
痛点 4:高阶能力默认开启,新手使用门槛高 ⭐⭐⭐⭐
原有问题:MCP 多设备协同、长期记忆等高阶功能默认启用,占用大量资源,劝退新手。 ✅ 分层解耦优化:
-
SDK 默认仅开启基础聊天、Lua、事件路由,高阶功能手动开启;
-
记忆模式拆分轻量版 / 完整版,普通用户默认轻量化;
-
官方文档拆分新手入门、高级多设备协同两套教程。
痛点 5:国内 IM 适配不完善,量产商用能力缺失 ⭐⭐⭐
原有问题:微信 / QQ 适配简陋,缺少可视化调试、批量 OTA、设备权限管理等商用能力。 ✅ 优化内容:
-
完善微信二维码登录、QQ 消息格式选择、图片附件收发;
-
上线 Web 可视化后台,支持脚本在线编辑、设备状态监控;
-
新增 NVS 密钥加密、增量 OTA、WiFi 开放网络兼容。 ⚠️ 短板:无批量设备管理、分级权限、远程批量固件推送,目前仅适合创客实验,不适合量产产品。
六、其他开发者实测真实反馈
1. ESP32-S3-BOX-3 全流程实测(6.8)
用户完整跑通配网、DeepSeek 接入、微信对话、温湿度采集、屏幕显示流程,问题修复效率极高:
| 问题 | 修复状态 | 说明 |
|---|---|---|
| WiFi 中文 SSID 乱码 | 已修复 | v0.1.0 合入 UTF-8 边界安全处理 PR |
| WiFi 断线无法重连 | 已修复 | 7 月 8 日提交永久重连逻辑 |
| 屏幕排版不美观 | 部分缓解 | LVGL Lua + 网页模拟器完善框架,代码质量依赖大模型 |
| 温湿度读取偏差 | 未修复 | 硬件 I2C 时序板级适配问题 |
2. DFRobot P4/K10 开发板实测
-
配置文件烧录丢失:部分缓解,网页烧录支持保留配置;
-
第三方屏幕显示效果差:基础框架完善,缺少定制化调优;
-
无原生语音 STT/TTS:硬件通路搭建完成,语音识别依赖云端 API。
实测总结
网络、编码类基础问题修复速度极快;硬件适配、传感器精度、屏幕显示等硬件相关问题,仍存在优化空间;Rust 重构将长期解决底层内存、并发稳定性痛点。
七、项目三大长期发展路线推演
路线 1:Rust 核心全面替换 C,解决嵌入式底层痛点
rust-migration分支是当前最高优先级技术方向,依靠claw-cabi实现渐进式迁移,长期收益:
-
Rust 内存安全机制,彻底杜绝野指针、内存泄漏、缓冲区溢出;
-
原生异步并发模型,多智能体、多外设并行运行更稳定;
-
标准化 Crate 包管理,模块复用、版本维护更简单。
路线 2:MCP 双向协议成为乐鑫统一设备互联标准
MCP 是 ESP-Claw 核心技术壁垒,设备同时作为服务端与客户端,实现跨板卡协同。后续计划下沉至 ESP-Brookesia、ESP-SR 等乐鑫官方框架,统一全系列 AI 硬件交互标准。
路线 3:离线本地自治优先,布局端侧 TinyLLM 适配
整套架构底层设计优先保障断网可用,本地规则、记忆、硬件控制完全不依赖网络。仓库已预留 TinyLLM 适配接口,社区高呼声需求为局域网 Ollama 本地大模型对接,未来弱化云端依赖。
八、ESP-Claw 真实定位
项目优势
-
前沿工程落地:将 AI 智能体、动态 Lua 脚本、MCP 设备互联、多模态 IM 交互完整落地 ESP32 嵌入式芯片,行业首创;
-
迭代效率极高:三个月完成原型、正式版、Rust 重构三线并行,日均 7.5 次提交,团队投入充足;
-
社区响应迅速:用户反馈网络、编码类问题最快 3 天完成修复合并,外部开发者持续贡献硬件适配;
-
路线及时修正:第三方测评提出的架构缺陷全部纳入底层重构,本地自治成为核心设计原则。
现存短板与风险
-
Chat Coding 固有延迟:全新复杂逻辑依赖云端大模型,芯片算力限制无法彻底消除延迟;
-
硬件门槛硬性约束:低配 Flash 设备只能使用阉割功能;
-
商用量产能力不足:缺少批量设备管控、权限体系,定位创客实验框架而非消费级产品;
-
Rust 双线维护成本:C 主分支 + Rust 迁移分支并行开发,长期维护压力大,核心外部贡献较少。
我的评价基本没有改变
ESP-Claw 并非面向普通小白的成品工具,而是乐鑫面向下一代边缘 AI 智能体的技术预研与实验框架。它重新定义了 ESP32 物联网设备的交互形态,把自然语言可编程、本地自主决策落地到低成本 MCU,技术方向具备极高参考价值,但现阶段不适合直接商用量产(或许我们可以用长远发展的眼光来看待这个问题,持有开放的态度,用户拥有自由的选择权,我将持续关注这个项目的后续发展)。
你有体验过 ESP-Claw?适配了哪款开发板?
在调试过程中遇到过哪些底层源码痛点?
欢迎在评论区分享你的实操经历,聊聊你对 Rust 重构、MCP 多设备协同的期待!
参考来源
[1] 乐鑫科技官网,乐鑫龙虾 ESP-Claw 来了!聊天就能开发智能设备(2026-04-21)
https://www.espressif.com/zh-hans/news/ESP_CLaw_Release
[2] GitHub,espressif/esp-claw 仓库(含 README、Release Notes、提交记录)
https://github.com/espressif/esp-claw
[3] ESP32Cube,ESP-Claw Open Source: Bringing Agent Runtime to ESP32(技术架构分析)
https://docs.esp32cube.com/esp32cube/cmogwa5xr00000ajrljsv08fv/
[4] CSDN,Tony_Shen,ESP-Claw 与 MimiClaw:ESP32 端侧 AI 智能体的两种实现路径(2026-05-08)
https://blog.csdn.net/Tony_Shen/article/details/160902127
[5] New Claw Times,Espressif Releases ESP-Claw, an OpenClaw-Inspired Agent Framework(2026-05-07)
https://newclawtimes.com/articles/espressif-esp-claw-openclaw-microcontrollers-edge-ai-agents/
[6] Elektor Magazine,ESP-Claw Framework Brings AI Agents to ESP32(2026-05-12)
https://www.elektormagazine.com/news/esp-claw-framework
[7] 51CTO,蓝梦之翼,ESP32-S3 具身智能实战:从 MimiClaw 到 ESPClaw 的全链路自治 Agent 开发(2026-04-30)
https://blog.51cto.com/u_16099304/14578196
[8] 电子工程世界,IC爬虫,ESP-Claw 项目笔记(2026-04-30)
https://m.eeworld.com.cn/bbs_thread-1348944-1-1.html
[9] DF创客社区,云天,基于 FireBeetle 2 ESP32-P4 + ESP-Claw 四驱车(2026-05-30)
https://mc.dfrobot.com.cn/thread-400198-1-1.html
[10] ESP-Claw,官方文档站
https://esp-claw.com
I previously reviewed this, and the experience was indeed not very good. A month has passed — is this project still worth a shot? Should I try again?
Introduction
On April 21, 2026, Espressif officially open-sourced the edge AI agent framework ESP-Claw (Claw), marking nearly three months since its release. The official repository, espressif/esp-claw, shows continuous growth: 1,799 stars, 370 forks, 47 contributors, 70+ merged PRs, and 853 total commits.
Within just three months, the project achieved three key milestones:
April: rapid prototype construction and official open-source announcement;
June 12: first official release (v0.1.0), incorporating 20 community PRs;
Simultaneously: launch of the rust-migration branch, initiating a full rewrite from C to Rust — by July 10, this branch had 751 commits and 8 major Rust modules.
1. ESP-Claw: Espressif’s Edge Chat Coding Agent Framework
1.1 Core Definition
ESP-Claw is an edge AI agent framework built by Espressif for the full ESP32 chip series. Its core innovation is “Chat Coding” — users converse with devices via IM software like WeChat, QQ, Feishu, or Telegram, and the device automatically calls a large language model to parse natural language and generate executable Lua scripts. This enables custom hardware control without compiling or flashing — a paradigm shift from pre-programmed to adaptive, locally autonomous agents.
1.2 Key Metrics
1.3 Four-Layer Architecture
The framework employs a layered, decoupled architecture for modular scalability:
Based on 853 commits, the project’s evolution since April 3 spans six distinct stages, with an average of 7.5 daily commits and a peak of 33 in a single day.
Stage 1: Project Genesis (04.03–04.10)
Initial commit on April 3 kicked off the project. Within one week:
Established the first-generation Agent Loop execution framework;
Modularized Lua components, integrated Feishu and Anthropic API;
Developed task scheduler, session persistence, and hardware capability routing;
Migrated the task stack to PSRAM to optimize memory layout on the ESP32.
Stage 2: Feature Explosion (04.11–04.30)
The official open-source publication on April 21 sparked the most intensive development phase:
Introduced the Board Manager and a structured V2 long-term memory system;
Refactored Lua script paths and standardized device-wide event output;
Split the monolithic claw_core into multiple internal modules, transitioning from prototype to engineered architecture;
Launched the official documentation site with GitHub Pages and automated CI pipelines.
Stage 3: Architectural Optimization + Community Influx (05.01–05.31)
Phase 3: Architectural Optimization + Community Onboarding 05.01–05.31
The focus shifted from "fast stacking features" to architectural standardization, with numerous third-party developers submitting PRs for board compatibility.
On May 7, the Lua hardware driver was refactored to unify initialization/destruction interfaces and standardize peripheral expansion.
A new Lua module for LVGL graphical UI was added, supporting visual screen development.
The Agent Manager was launched, enabling concurrent multi-agent operation and sub-task tracking.
The community contributed 15+ development board adaptations, including M5Stack, LilyGO, and DFRobot.
Phase 4: Stability Polishing, v0.1.0 Official Release (06.01–06.12)
Comprehensive fixes for security, memory, and network issues were made. On June 12, v0.1.0 went live, merging 20 community PRs.
Stability Fixes
Automatic WiFi reconnection, open WiFi network compatibility, LLM UTF-8 purification, character boundary security, and memory leak/wild pointer repairs.
New Capabilities
Lua asynchronous execution, local Plan Mode skills, visual detection, MCP service nodes, and web online flashing workflow.
Phase 5: Rust Core Migration Branch Launched (06.09–Ongoing)
Simultaneously with the official release, the rust-migration experimental branch was created, aiming to rewrite the C-based claw_core into Rust to avoid embedded memory safety issues.
151 independent commits were made, establishing a complete 8-crate Rust system.
On July 2, an asynchronous concurrent runtime was added; on July 6, the full framework became operational.
The claw-cabi compatibility layer enables bidirectional calls between Rust and existing C code, allowing incremental migration without full project refactoring.
Rust Crate Architecture
Phase 6: Ongoing Ecosystem Expansion (06.23–07.10)
The master branch continued iterating, focusing on hardware adaptation and developer experience.
The anyboard solution was introduced for generic board support without custom BSP.
A Lua LVGL web simulator was added for real-time browser visualization.
Support for additional boards like Waveshare, Seeed, and Rockbase was expanded.
Communication protocols ESP-NOW, CAN/TWAI, and MQTT were opened for PRs.
The documentation site was upgraded to Astro v7 with enhanced tutorials.
III. Repository Branch Structure and Commit Data Statistics
1. 三大分支现状
2. Community PR Topic Distribution Bar Chart
Community external contributions (60%) are concentrated on development board hardware adaptation, with the remainder focused on web interaction, LLM optimization, IM platform repairs, and Lua driver expansion. This reflects developers' core demand: making the framework compatible with more hardware devices ("everyone wants to latch on — can beginners handle it?").
IV. Five Pivotal Architectural Reconstructions
4.29 claw_core First Split The initial monolithic core package handled all logic, resulting in bloated and unmaintainable code. The split established modular boundaries, marking the transition from prototype to engineered project.
5.07 Standardization of hardware drivers with Lua: Consolidating scattered scripts into standardized driver interfaces for fast I2C, UART, ADC, and USB peripheral integration, significantly lowering hardware expansion costs.
5.27 Reconstruction of multi-instance Agent Manager: Supporting simultaneous execution of multiple agents with independent progress tracking for subtasks, laying the foundation for multi-device collaboration and concurrent control.
Early June v0.1.0 stability reconstruction: Focusing on memory safety, character encoding, and network fault tolerance to transition from "usable features" to "reliable stability."
6.09 Start of Rust core migration project: The most forward-looking adjustment, using Rust to solve C language issues like memory overflow, wild pointers, and concurrent crashes in embedded systems, with smooth transition via CABI bridge.
V. Review of Five Core Pain Points
On June 3, 2026, this account published "Wild Review of ESP-Claw," identifying five structural flaws. Here is a comparison of improvements against the latest code in July:
Wild Review of ESP-Claw
Pain Point 1: High latency in LLM-generated Lua chain, poor script fault tolerance ⭐⭐⭐⭐
Original issue: The full chain from IM-LLM to script to hardware had a maximum delay of 7 seconds; LLM code was prone to pin and memory errors without verification. ✅ Improvements:
Local routing rules are now prioritized at the bottom layer, enabling sub-second responses for high-frequency commands;
Lua scripts undergo sandbox pre-verification before writing to block illegal hardware calls;
Successful logic is automatically cached as rules, allowing similar commands to execute without repeated model calls;
v0.1.0 introduces asynchronous Lua task API for parallel execution without device blocking. ⚠️ Remaining gap: Completely new custom logic still requires the full LLM chain, and network fluctuations can cause delays; the embedded Lua stack limits the execution of very long scripts.
Pain Point 2: High hardware resource threshold, poor experience on 8MB low-spec devices ⭐⭐⭐
Original issue: Enabling full functionality on 8MB Flash hardware frequently caused memory overflow, as the core task stack consumed too much RAM. ✅ Improvements:
Full modularization with Kconfig allows minimal configurations to run on as little as 2MB Flash;
The PSRAM virtual file system (RAMfs) reduces Flash read/write wear;
A new memory monitoring CLI provides real-time visibility into heap and cache usage. ⚠️ Remaining gap: For full feature sets, the system still requires 16MB Flash + 8MB PSRAM, meaning low-spec hardware can only run a reduced subset.
Pain Point 3: Over-reliance on cloud LLM, weak local event routing capability ⭐⭐⭐⭐⭐ (Most thoroughly improved)
Original issue: Even simple switch or timer commands required calling the large model, rendering devices useless without internet. ✅ Thorough overhaul:
The system now forces local rule priority, only requesting the LLM if no matching rule is found;
LLM-generated logic is saved to a local rule library and can be manually edited via web interface;
Official documentation now emphasizes "off-network autonomy" and "local millisecond response," solving a critical IoT flaw.
Pain Point 4: Advanced features enabled by default, high entry barrier for novices ⭐⭐⭐⭐
Original issue: Features like MCP multi-device collaboration and long-term memory were enabled by default, consuming significant resources and discouraging beginners. ✅ Decoupling and optimization:
The SDK defaults to only basic chat, Lua, and event routing; advanced features must be manually enabled;
Memory modes are split into lightweight and full versions, with lightweight being the default for regular users;
Tutorials are now segmented into beginner and advanced sections.
Pain Point 5: Incomplete domestic IM adaptation, lack of commercialization capabilities ⭐⭐⭐
Original issue: WeChat/QQ adaptation was rudimentary, missing visualization, batch OTA, and device permission management. ✅ Improvements:
Enhanced WeChat QR code login, QQ message format selection, and image attachment transmission/reception functions.
Launched a web-based visualization backend that supports online script editing and device status monitoring.
Added NVS key encryption, incremental OTA, and compatibility with open WiFi networks. ⚠️ Shortcoming: Lacks bulk device management, tiered permissions, and remote bulk firmware pushing, making it currently suitable only for hacker experiments and not for mass-produced products.
VI. Other Developers' Actual Test Feedback
1. ESP32-S3-BOX-3 Full Process Test (6.8)
The user successfully completed the entire workflow including network configuration, DeepSeek integration, WeChat conversation, temperature/humidity acquisition, and screen display. The problem-solving efficiency was extremely high:
2. DFRobot P4/K10 Development Board Test
Configuration file burn-in loss: Partially resolved, web burn-in now supports preserving configurations.
Poor display effect of third-party screens: The basic framework is complete, but lacks customized optimization.
No native voice STT/TTS: The hardware channel is established, but voice recognition relies on cloud API.
Test Summary
Basic network and coding issues are fixed very quickly; hardware-related issues such as hardware adaptation, sensor precision, and screen display still have room for optimization. The Rust refactoring will permanently solve underlying memory and concurrency stability problems.
VII. Three Long-Term Development Roadmaps
Roadmap 1: Complete Replacement of C Core with Rust to Solve Embedded Bottom-Layer Pain Points
The rust-migration branch is the top priority technical direction, using claw-cabi for gradual migration. The long-term benefits include:
Rust's memory safety mechanism completely prevents wild pointers, memory leaks, and buffer overflows.
The native asynchronous concurrency model ensures more stable parallel operation of multiple agents and peripherals.
Standardized Crate package management simplifies module reuse and version maintenance.
Roadmap 2: MCP Bidirectional Protocol Becomes Espressif's Unified Device Connectivity Standard
MCP is the core technical barrier of ESP-Claw, allowing devices to act as both servers and clients for cross-board collaboration. The future plan is to integrate it into official Espressif frameworks like ESP-Brookesia and ESP-SR to unify the interaction standards for all AI hardware series.
Roadmap 3: Prioritization of Offline Local Autonomy and Layout of End-Side TinyLLM Adaptation
The entire architecture is designed to ensure usability without the internet, with local rules, memory, and hardware control being completely independent of the network. The repository has reserved TinyLLM adaptation interfaces, and there is strong community demand for connecting local Ollama models within a local area network, aiming to reduce dependence on cloud services in the future.
VIII. True Positioning of ESP-Claw
Project Advantages
Cutting-edge engineering implementation: Integrates AI agents, dynamic Lua scripts, MCP device connectivity, and multimodal IM interaction on the ESP32 embedded chip, setting a precedent in the industry.
High iteration efficiency: Completed three major phases—prototype, official release, and Rust refactoring—in three months, with an average of 7.5 commits per day, indicating strong team dedication.
Prompt community response: User-reported network and coding issues are fixed and merged within as little as 3 days, and external developers continue to contribute to hardware adaptation.
Timely course correction: Architectural flaws identified by third-party evaluations are all included in the bottom-layer refactoring, with local autonomy becoming the core design principle.
Existing Shortcomings and Risks
Inherent latency in Chat Coding: Novel complex logic relies on cloud-based large models, and the limited computational power of the chip prevents full elimination of latency.
Hardware threshold constraints: Low-configuration Flash devices can only use truncated features;
Insufficient commercial mass-production capability: lacks batch device management and permission systems; positioned as a maker experimental framework rather than a consumer-grade product;
Rust dual-line maintenance cost: parallel development of the C main branch and Rust migration branch; long-term maintenance pressure is high, with limited core external contributions.
My evaluation remains largely unchanged.
ESP-Claw is not a finished tool for ordinary beginners; it is Espressif’s technical pre-research and experimental framework for the next-generation edge AI agents. It redefines the interaction form of ESP32 IoT devices, bringing natural language programmability and local autonomous decision-making to low-cost MCUs. The technical direction has high reference value but is not yet suitable for direct commercial mass-production (perhaps we can view this with a long-term development perspective, maintain an open attitude, and users have freedom of choice; I will continue to follow the project’s progress).
Have you experienced ESP-Claw? Which development board did you adapt it to?
What underlying source code pain points did you encounter during debugging?
Welcome to share your hands-on experience in the comments, discussing your expectations for Rust refactoring and MCP multi-device collaboration!
Reference sources
[1] Espressif official website, Espressif ESP-Claw is here! Develop smart devices just by chatting (2026-04-21)
https://www.espressif.com/zh-hans/news/ESP_CLaw_Release
[2] GitHub, espressif/esp-claw repository (including README, Release Notes, commit records)
https://github.com/espressif/esp-claw
[3] ESP32Cube, ESP-Claw Open Source: Bringing Agent Runtime to ESP32 (technical architecture analysis)
https://docs.esp32cube.com/esp32cube/cmogwa5xr00000ajrljsv08fv/
[4] CSDN, Tony_Shen, ESP-Claw and MimiClaw: Two implementation paths for ESP32 edge-side AI agents (2026-05-08)
https://blog.csdn.net/Tony_Shen/article/details/160902127
[5] New Claw Times, Espressif Releases ESP-Claw, an OpenClaw-Inspired Agent Framework (2026-05-07)
https://newclawtimes.com/articles/espressif-esp-claw-openclaw-microcontrollers-edge-ai-agents/
[6] Elektor Magazine, ESP-Claw Framework Brings AI Agents to ESP32 (2026-05-12)
https://www.elektormagazine.com/news/esp-claw-framework
[7] 51CTO, Blue Dream Wing, ESP32-S3 embodied intelligence practice: Full-link autonomous agent development from MimiClaw to ESPClaw (2026-04-30)
https://blog.51cto.com/u_16099304/14578196
[8] Electronic Engineering World, IC Crawler, ESP-Claw project notes (2026-04-30)
https://m.eeworld.com.cn/bbs_thread-1348944-1-1.html
[9] DF Maker Community, Cloud Sky, Based on FireBeetle 2 ESP32-P4 + ESP-Claw four-wheel drive car (2026-05-30)
https://mc.dfrobot.com.cn/thread-400198-1-1.html
[10] ESP-Claw, official documentation site
https://esp-claw.com
