fcdb3fc15e
A series of cross-cutting stability fixes that surfaced together
during a long debugging session.
reasoning_content / Claude prefill self-replicating 400:
- ChatController persists typed errors (content starts with '[错误] ')
with status='error', so the failure text stops being re-sent as
multi-turn context — DeepSeek thinking 400 ('reasoning_content
must be passed back') and Claude 400 ('does not support assistant
message prefill') used to recursively re-create themselves every
retry by polluting history.
- BaseAgent.sanitizeForLlm filters status='error' / '[错误] ' prefix
assistant messages from history before LLM dispatch.
- BaseAgent.fetchHistoryMessages defensively drops trailing
AssistantMessages — Claude rejects assistant-tail prompts.
- NodeStreamingChatHelper.dropTrailingAssistant runs the same
defense at every doStreamCall pre-egress, so the in-turn
summarizing→reasoning transition (which leaves an assistant
scaffold at the tail) doesn't trip Claude either.
- AgentGraphBuilder.FallbackPolicy.DEEPSEEK switched (null,true,true)
→ (' ',false,true), aligning with KIMI/OPENAI's tolerant ' '
fallback. The previous 'force explicit 400' design was the
self-replicating loop's prime mover.
narration + tool args truncation:
- ReasoningNode.DEFAULT_MAX_OUTPUT_TOKENS 4096 → 16384. The 4k cap
was decapitating renderDocx tool_call args mid-stream when the
model emitted a long content field on top of thinking content;
the resulting 'invalid JSON' aborted execution silently.
- ReasoningNode appends a hermes-style TOOL_USE_ENFORCEMENT clause
to every system prompt: 'when you say you will perform an action,
call the tool now in the same response — narration is a protocol
violation'. Treats 'now I will generate the docx' (and never
actually calling renderDocx) as a forbidden pattern.
- ToolExecutionExecutor.normalizeToolExecutionError reframes the
JSON-truncated error as actionable instructions: 're-call the
same tool now with shorter content or split into multiple
sequential calls; do NOT describe the result as text'.
side fixes from the same evening:
- ChatController doOnComplete skips completionPublisher.publish
when isError=true, keeping memory extraction off the garbage path.
- ChatController doOnComplete queued-message guard simplified to
'cr.queuedInput() != null', matching the other 4 sites in the
controller. The previous 'isInterruptFollowup || !wasStopped'
guard silently dropped queued messages when the user did
Stop-then-Enqueue (wasStopped=true && interruptType=null), losing
the freshly-typed follow-up message.
- prompts/graph/summarize-system.txt now distinguishes 'single
task' (default; output one cohesive summary) from 'multiple
independent sub-tasks' (use the子任务 N format). Stops the
summarizer from inventing '子任务 1: PRO-027' decomposition for
unitary requests like 'write me a project proposal'.
|
||
|---|---|---|
| assets | ||
| docker/searxng | ||
| mateclaw-plugin-api | ||
| mateclaw-plugin-sample | ||
| mateclaw-server | ||
| mateclaw-ui | ||
| mateclaw-webchat | ||
| .dockerignore | ||
| .env.example | ||
| .gitignore | ||
| docker-compose.yml | ||
| LICENSE | ||
| README_zh.md | ||
| README.md | ||
Most AI tools die when their vendor has a bad day. Most forget you the moment the tab closes. Most give you a chatbox and call it a product.
MateClaw is the whole widget. One deployment. Reasoning, knowledge, memory, tools, channels — built together, not bolted on. And when your primary model goes down, the next one picks up mid-sentence.
Three things that make it different
1 · Your AI doesn't die when a model does
Primary key expired. Vendor returns 401. Network blip. Quota drained.
Other tools hand you a red error card. MateClaw routes to the next healthy provider — DashScope, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, Ollama, LM Studio, MLX, 14+ in total — and the user sees the reply finish. A provider health tracker parks bad vendors in a cooldown window so they don't waste seconds on every turn.
You don't write a retry script. You drag providers into priority order in Settings → Models and watch the health dashboard fill with green dots as requests route around failures in real time.
2 · Knowledge that links itself
Upload a PDF, a batch of markdown, a scraped page — raw material in.
MateClaw's LLM Wiki digests it into structured pages, builds [[links]] between them, and remembers where every sentence came from. Click a citation, see the exact source chunk. Ask a question, the page you get is stitched from the right chunks — with references you can verify.
This is the difference between a warehouse and a library.
3 · One product, five surfaces
| Surface | What it is |
|---|---|
| Web Console | Full admin — agents, models, tools, skills, knowledge, security, cron |
| Desktop | Electron app with a bundled JRE 21. Double-click, run. No Java install |
| Webchat Widget | One <script> tag embed. Drop it on any site |
| IM Channels | DingTalk · Feishu · WeChat Work · Telegram · Discord · QQ |
| Plugin SDK | Java module for third-party capability packs |
Same brain. Same memory. Same tools. Different doors.
$0 · No tokens metered. No seats billed. Your server. Your data. Your keys.
What's in the box
Agent runtime
ReAct for iterative reasoning. Plan-and-Execute for complex multi-step work. Dynamic context pruning, smart truncation, stale-stream cleanup — the boring stuff that makes long conversations actually work.
Knowledge & memory
- LLM Wiki — raw materials digest into linked pages with citations
- Workspace memory —
AGENTS.md,SOUL.md,PROFILE.md,MEMORY.md, daily notes - Memory lifecycle — post-conversation extraction, scheduled consolidation, dreaming workflows
Tools, skills, MCP
Built-in tools for web search, files, memory, date/time. MCP over stdio / SSE / Streamable HTTP. SKILL.md packages from the ClawHub marketplace. A Tool Guard layer with RBAC, approval flows, and path protection — capability needs boundaries.
Multimodal creation
Text-to-speech · Speech-to-text · Image · Music · Video. First-class, not add-ons.
Enterprise-ready
RBAC + JWT. Full audit trail. Flyway-managed schema that auto-heals on upgrade. One JAR to ship. MySQL in production, H2 for dev — nothing to change in your code.
AI is becoming infrastructure
On March 2, 2026, Claude went dark for 4 hours across API, web, and mobile. Three weeks later, another 5 hours. Every company that bet their AI strategy on a single vendor spent those outages staring at red error cards.
This is the same shift databases went through around 2010 and cloud went through around 2018: the winning layer stops being tied to one supplier. 57% of companies now run AI agents in production. None of them want one vendor's bad day to become their bad day.
MateClaw is that layer — built the Spring Boot way.
Why MateClaw
| MateClaw | OpenClaw | Hermes Agent | Claude Code | Cursor | |
|---|---|---|---|---|---|
| Multi-vendor failover | Chain + health tracker + cooldown | Swap providers via config | Orchestration w/ retry | Anthropic only | One model |
| Knowledge digestion | LLM Wiki + page-level citations | Canvas + memory | Skills Hub + memory | — | Code index |
| Multi-user admin | RBAC + approval flow + audit | Config-file first | Single-user CLI | Enterprise tier | Teams plan |
| Surfaces | Web admin + Desktop + Widget + SDK + 6 IM | 25+ chat channels | 15+ channels (CLI-led) | 3 IM preview | IDE only |
| Stack | Java (Spring Boot) | TypeScript | Python | TypeScript | Electron/TS |
| License / Price | Apache 2.0 · Free | MIT · Free | MIT · Free | Proprietary · $20–200/mo | Proprietary · $0–200/mo |
OpenClaw and Hermes Agent are excellent personal AI platforms — pick either if you're running one user on one laptop, building your own agent from CLI, and treating everything as config files to hand-tune. Both have bigger communities than MateClaw today.
MateClaw is the version built for teams. RBAC per agent, per model, per tool. An approval flow that pauses risky actions for review. Full audit trail. A web admin dashboard where one operator manages 50 agents across 14 vendors. Spring Boot inside — drop-in for any Java shop already running production services.
Same "whole widget" philosophy. Different center of gravity.
Quick start
# Backend
cd mateclaw-server
mvn spring-boot:run # http://localhost:18088
# Frontend
cd mateclaw-ui
pnpm install && pnpm dev # http://localhost:5173
Login: admin / admin123
Docker
cp .env.example .env
docker compose up -d # http://localhost:18080
Desktop
Download from GitHub Releases. Bundles JRE 21. No Java install needed.
Architecture
Technical architecture
Project structure
mateclaw/
├── mateclaw-server/ Spring Boot 3.5 backend (Spring AI Alibaba, StateGraph runtime)
├── mateclaw-ui/ Vue 3 + TypeScript admin SPA (built into the server JAR)
├── mateclaw-webchat/ Embeddable chat widget (UMD / ES bundles)
├── mateclaw-plugin-api/ Java SDK for third-party capability plugins
├── mateclaw-plugin-sample/ Reference plugin implementation
├── docker-compose.yml
└── .env.example
Desktop binaries ship via GitHub Releases with a bundled JRE 21 — no Java install needed.
Tech stack
| Layer | Technology |
|---|---|
| Backend | Spring Boot 3.5 · Spring AI Alibaba 1.1 · MyBatis Plus · Flyway |
| Agent | StateGraph runtime · ReAct + Plan-Execute |
| Database | H2 (dev) · MySQL 8.0+ (prod) |
| Auth | Spring Security + JWT |
| Frontend | Vue 3 · TypeScript · Vite · Element Plus · TailwindCSS 4 |
| Desktop | Electron · electron-updater · JRE 21 (bundled) |
| Widget | Vite library mode · UMD + ES bundles |
Documentation
Full docs at claw.mate.vip/docs — setup, architecture, each subsystem, API reference.
Roadmap
Sharper multi-agent collaboration · Smarter model routing · Deeper multimodal understanding · Longer-lived memory · A richer ClawHub.
Contributing
git clone https://github.com/matevip/mateclaw.git
cd mateclaw
cd mateclaw-server && mvn clean compile
cd ../mateclaw-ui && pnpm install && pnpm dev
Why the name
Mate is companion. Claw is capability.
Something that stays with you — and grabs work and moves it.
License
Apache License 2.0. No asterisks.