NextClaw
An open-source, extensible, local-first self-hosted agent workspace that takes you from one request to a usable result without losing work in between.
Evidence: README states config, conversations, and credentials stay local and warns the public trial gateway should not carry sensitive data — decent data-flow transparency; SECURITY.md candidly discloses a 2026 malicious @nextclaw/bird marketplace skill with removal, scanning, republish-rescan, and mirror reconciliation — unusually transparent incident handling. Deductions: no fine-grained least-privilege model documented despite broad capabilities (local files, browser, messaging channels); user-confirmation flows and external-effect controls (remote channel delivery, cron) are not shown in the provided files; rollback appears only as an unfinished 1.0 plan item.
Evidence: the aigen package has vitest unit tests (CLI tree, stable JSON error codes, provider response parsing); the repo carries extensive governance/lint/smoke scripts and release-preflight workflows, structurally self-consistent. Deductions: visible tests only cover the image-generation submodule; no test evidence for core agent runtime behavior; long serial build/lint/tsc chains are fragile to external source failures; failure-message quality is evidenced for one package only.
Evidence: README clearly defines audience (long-term personal AI partner) and scenarios (research, data, documents, local files, scheduling); multi-runtime (Native/Codex/Claude Code/OpenCode/Hermes) and multi-platform (macOS/Windows/Linux/Docker/cloud) fit well documented; provider and channel lists detailed. Deductions: 1.0 acceptance incomplete and capability boundaries mostly deferred to linked gap-analysis docs; trigger precision (cron/channel matching) has no file evidence.
Evidence: full MIT LICENSE consistent with README badge, upstream acknowledgements (OpenClaw, NanoBot) give clear attribution; install notes cover desktop/npm/Docker including port and reverse-proxy cautions; planning docs (ROADMAP, PRD, gap analysis) are public and honestly state 1.0 is unfinished with no release date; changesets and release-preflight workflows exist. Deductions: no CHANGELOG file in evidence — release notes live off-repo; maintenance responsibility rests on a single unverified publisher with no multi-maintainer evidence.
Evidence: clear positioning (local-first, multi-runtime, machine-readable CLI output), free trial lowers entry cost, resource-usage docs referenced, concrete use cases. Deductions: effectiveness and correctness are asserted, not statically verifiable; the security cost of the public gateway is covered by a single warning sentence rather than substantive guidance.
Evidence: README claims (install methods, port, Node 22, test commands) are corroborable against package., workflows, and test files; planned vs current scope is explicitly separated. Deductions: external sources (npm metrics, docs site, Docker install script) fall outside the provided files so cross-source corroboration is limited; performance and 'usable result' claims cannot be verified statically.
- Static review with low confidence: nothing was installed or executed; effectiveness claims rest on file evidence only.
- This repository disclosed a malicious marketplace skill (@nextclaw/bird) containing remote-execution instructions, live 2026-02-27 to 2026-08-11. If installed, verify your SKILL.md SHA-256 and follow SECURITY.md remediation steps.
- The free trial routes through a public gateway; the README itself warns against sending sensitive or confidential data — do not send private content before attaching your own provider.
- The agent can operate local files, browsers, multiple messaging channels, and scheduled jobs — a large permission surface; deploy in an isolated environment and vet channel and skill sources.
- 1.0 acceptance is incomplete and capability boundaries live in external links; confirm the current feature scope before relying on it.
- Publisher identity is unverified and maintenance rests on a single party; continuity of updates and security patches is uncertain.
What does this agent do, and when should you use it?
NextClaw is an open-source, self-hosted agent workspace positioned as a long-term personal AI partner, deployable on macOS, Windows, Linux, Docker, and cloud VMs. It ships as a desktop app, an npm package, and a first-class nextclaw CLI, serving a local web workspace by default at 127.0.0.1:55667. Core capabilities include task workspaces that keep conversations, local files, web research, and generated documents in one context; multi-runtime execution with Native, Codex, Claude Code, OpenCode, and Hermes; Dedicated Agents with their own role, memory, skills, and runtime; scheduled tasks; messaging channels such as Weixin, Feishu/Lark, QQ, and Telegram; and persistent Panel Apps you build with the agent. Models include a built-in free-trial gateway plus OpenRouter, OpenAI, Anthropic, Gemini, DeepSeek, and custom OpenAI-compatible endpoints. The project is working toward a 1.0 that understands users, takes initiative, and improves itself, but full acceptance testing is incomplete and no release date is set.
You issue a request in the local workspace — a report, data analysis, file operation, small app, or recurring task — and NextClaw organizes the tools and steps behind it. Concretely it can: collect pages, notes, and references into a brief, source list, or comparison table; gather data from websites, CSV files, or spreadsheets, clean it, draw charts, and write conclusions; shape source material into reports, articles, proposals, or weekly updates; inspect, rename, extract, classify, and summarize local documents; turn repeated jobs into scripts, local apps (Panel Apps), or reusable workflows; and receive requests from chat channels, run scheduled briefs, and deliver results to an AI Inbox. Each task can run on Native, Codex, Claude Code, OpenCode, or Hermes at start time while keeping the Agent's identity, workspace, memory, and skills. The workspace opens code, Markdown, HTML, Word, Excel, and PowerPoint beside the conversation, supports image generation, a skill market (MCP servers, CLI tools, browser control), and configurable model providers.
- A researcher who collects web pages, notes, and references and needs them turned into a brief, source list, or comparison table
- A data worker who needs to pull data from CSVs or spreadsheets, clean it, plot charts, and write conclusions
- Users processing many local documents who want to inspect, rename, classify, and summarize files without losing task context
- Developers or power users who want to turn recurring jobs into scripts, dashboards, or reusable workflows
- Teams on Weixin, Feishu, or Telegram who want to send tasks from chat, receive scheduled briefs, and get results back in the right channel
What are this agent's strengths and limitations?
- Local-first and self-hosted: configuration, conversations, and credentials stay in an environment you control, with desktop, server, NAS, and Docker deployment paths
- Multi-runtime architecture: switch between Native, Codex, Claude Code, OpenCode, and Hermes while keeping the Agent's identity, workspace, memory, and skills
- Complete task context: conversations, local files, research, and generated documents stay in one task, with in-place preview including Office formats
- Broad connectivity: a dozen messaging channels and many model providers, plus custom OpenAI-compatible endpoints
- First-class CLI: nextclaw commands support scripts, CI jobs, and other Agents with machine-readable output
- The 1.0 goals (proactive analysis, continuous self-improvement) are planned scope; acceptance testing is incomplete with no release date
- The built-in free trial runs through a public gateway with changing limits and models — unsuitable for sensitive or confidential data
- Remote deployments default to plain HTTP on port 55667; you must set up Nginx/Caddy HTTPS termination for production
- Memory usage increases during active work, so you should consult the official resource-usage documentation before sizing deployments
- Long-term stable use requires connecting your own model provider; the free gateway is trial only
How do you install or deploy this agent?
Desktop app: download the latest stable release for macOS, Windows, or Linux from https://nextclaw.io/en/download/. Via npm: install Node.js LTS first, then run:
npm install -g nextclaw
nextclaw startOpen http://127.0.0.1:55667 and start with the built-in free-trial model; no API key required. Docker (for long-running servers or cloud VMs):
curl -fsSL https://nextclaw.io/install-docker.sh | bashPort 55667 on remote hosts serves plain HTTP; terminate HTTPS with Nginx or Caddy for regular access. From source: run pnpm install && pnpm dev start at the repo root, optionally split with pnpm dev:backend / pnpm dev:frontend. The default data directory is ~/.nextclaw; set NEXTCLAW_HOME for isolation.
How do you use this agent?
After starting, issue natural-language requests in the web workspace at 127.0.0.1:55667 — reports, data analysis, or file operations. Choose a runtime (Native, Codex, Claude Code, OpenCode, or Hermes) when starting a task. Inspect and edit generated files (code, Markdown, Office documents, images) beside the conversation, and keep generated pages as Panel Apps to improve later. In settings, connect messaging channels (Weixin, Feishu/Lark, QQ, Telegram, Slack, etc.) so Agents can receive and reply to messages, configure scheduled jobs for recurring briefs or checks, and route results to the AI Inbox. CLI users can run core operations via nextclaw commands with machine-readable output for scripts, CI, or other Agents. Stop the service with nextclaw stop. The built-in free trial uses a public gateway — avoid sending sensitive data; connect your own provider for production use.
How does this agent compare with similar options?
NextClaw's acknowledgements note that its early exploration was inspired by OpenClaw (a full-stack AI assistant) and NanoBot (a small, extensible agent framework), useful as reference projects in the same space.