NexusDesk is a self-hosted agent platform whose first complete application turns customer support conversations into retrieved answers, tool calls, and confirmed tickets on your own infrastructure. An agent takes a request, retrieves context from your knowledge base, decides what to do, calls business tools, asks a human to confirm consequential actions, executes, records everything, and feeds evaluation, all as one visible loop rather than a black-box chat. The Harness around that loop is the engineering control layer: it governs permissions and execution boundaries, manages agent memory, routes human handoffs, and keeps an audit trail, so autonomy stays inside lines you drew. You configure, publish, and roll back agents, manage documents, chunks, indexes, and published knowledge versions, and connect Chat, Embedding, Rerank, OCR, and speech models through versioned profiles. Business APIs plug in as tools with their own credentials, parameters, and permissions, and support staff can take over a conversation, draft a ticket from it, and confirm follow-up work without leaving the flow. Channels and application APIs with server-sent events expose the same capabilities to your other systems, while execution records, model and tool call logs, and evaluation results show what the agent actually did. One PostgreSQL handles storage, job queues, and cross-process messaging, and one Compose file starts the stack. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

Point the gh CLI at your private Githome server and it works identically to github.com, because every REST v3 and GraphQL v4 endpoint returns byte-for-byte compatible JSON shapes, headers, Link pagination, and ETag caching. A single compiled Go binary serves the REST API under /api/v3, GraphQL under /api/graphql, git smart-HTTP transport, and a server-rendered web UI implementing the GitHub Primer design system with full functionality when JavaScript is disabled. The web interface provides repository browsing with Markdown rendering and syntax highlighting, issue tracking, pull requests with unified diffs and inline code review threads, full-text search, user profiles, and a notifications inbox. Authentication supports personal access tokens with prefix-based identification, OAuth device and web flows, and scoped permissions on every response header with only token hashes stored. The pull request system implements the complete lifecycle including async mergeable state computation, squash, merge, and rebase strategies, commit statuses, check runs, and statusCheckRollup fields. Webhooks deliver events with X-Hub-Signature-256 verification, SSRF protection, and at-least-once retry semantics. Storage supports SQLite for single-node deployments or PostgreSQL for production configurations. Performance optimizations include batch loaders replacing N+1 queries, keyset pagination, FTS-backed search, and GraphQL dataloaders. Docker deployment pulls the official multi-arch container image from GitHub Container Registry. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

Your team's documents sit in folders nobody can search, and the answer to a simple question hides on page 40 of a PDF someone saved last year. Pharos ingests PDFs, scans, Word documents, presentations, and spreadsheets into a local knowledge base, then answers questions against that library with citations traced back to the exact page. Hybrid retrieval finds the passages that matter, and every answer carries its evidence so readers can verify the claim in the source. Access control decides visibility per person: an API key resolves to an identity with a tenant and group memberships, retrieval filters hard to only the documents that identity may see, and a document with no permissions attached stays invisible to everyone. Two doors lead into the same library. The HTTP API runs a closed question-answering pipeline that retrieves, grounds, and answers in one call, and coding agents connect through MCP to decide for themselves when to retrieve, how to rewrite a query, and when a question needs multiple hops. A resident daemon keeps the search index and models warm, so every session shares one backend instead of reloading heavy resources. The deployment scales from one machine to replicated application tiers behind a load balancer. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

PandaDOS is a self-hosted control plane that turns Claude Code and Codex into long-running coding agents you direct through an issue queue from any browser, instead of babysitting a terminal. You file an issue with the outcome and acceptance criteria, and PandaDOS clarifies missing requirements, plans the work, picks the agent, and drives implementation, verification, and delivery in stages. Three approval levels, cautious, balanced, and automatic, decide which actions proceed alone; safe reversible steps continue automatically while destructive, irreversible, production, or secret-related decisions always come back to you. The browser workspace shows the real agent conversation, a native terminal, project files, Git history and diffs, and an execution summary, so you can inspect or intervene at any moment from a phone or laptop. Agents run on local or SSH-connected executor machines inside tmux sessions, and work sharing one repository is serialized while independent projects proceed in parallel without conflicting edits. Modules keep one assigned agent and durable project knowledge, so context from finished issues carries into the next task instead of evaporating when a chat ends. Teams can sign in through Feishu, ask the project manager questions in chat, and receive subscription notifications and approval cards on their phones. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

Codex LB Enhanced is a session-aware gateway that pools your ChatGPT accounts so a long Codex session survives quota limits, dropped connections, and paused accounts without losing the conversation. It serves Codex desktop, CLI, and IDE clients through the native Responses protocol and WebSocket transport, and answers any OpenAI-compatible client through the familiar /v1 endpoints. When an account runs out of quota, gets suspended, times out, or drops mid-stream, the gateway classifies the failure, removes that account from the current request, and selects the next one through your load-balancing policy. Verified request history, including interrupted tool-call context, is replayed across the account switch when it is safe, so the client sees a slower answer instead of a dead stream. Every API key carries its own policy: time windows from five hours to monthly, limits by credits, tokens, or cost, and model-specific rules. Usage appears in native Codex quota headers and in a /v1/usage view backed by the same data, so editors and automation agree on what is left. Diagnostics correlate each request with its upstream account, transport, failover stage, and retry count, and a web dashboard manages accounts, keys, usage windows, request logs, and recovery reports. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

BlogFactory is the first open-source agent control plane purpose-built for multi-site content operations, shipping 16 distinct capabilities across four operational pillars — Operate, Grow, Control, and Deliver — in its v0.1.0 release. The platform keeps source evidence, AI-generated drafts, revisions, SEO metadata, human review, preflight checks, and CMS delivery destinations in one workspace while enforcing a server-side authority ceiling that prevents agents from ever publishing live content, accessing credentials, or performing administrative actions. AI agents connect through a Streamable HTTP MCP endpoint with OAuth authentication and site-scoped tool discovery, compatible with Codex, ChatGPT, and any MCP-capable client. Content creation supports single articles from keywords, titles, URLs, PDFs, text, or YouTube sources, batch campaigns, and programmatic content via reusable templates with structured data rows. The Search Growth module connects Google Search Console complete-day data to ranked optimization opportunities, 30-day growth plans, content classification by signal type (growing, low-CTR, almost-ranking, page-two, zero-click), indexing checks, and semantic internal-link suggestions. Brand Voice separates global voice rules, writer profiles, knowledge context, and brand guidelines from article mechanics. Approved drafts are delivered as CMS drafts to WordPress, Ghost, Wix, or Framer through idempotent, optimistic-locking-protected handoffs with explicit destination selection and preflight validation. The stack runs on React/Vite, a Hono/TypeScript API, PostgreSQL, and MinIO via Docker Compose, with providers including OpenRouter, OpenAI, and Google services under a bring-your-own-AI model. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. AGPL-3.0 licensed.

Your application already knows how to send mail through Amazon SES; NitroMTA lets it keep that code, change the endpoint and keys, and deliver from IP addresses you actually control. The engine answers a submission only after the message is committed to disk, so accepted mail survives a crash, a power cut, and even the loss of its host when the disk survives. Capture mode runs the whole engine locally and shows each message in a browser at port 9090 instead of sending it, which makes testing painless before any DNS exists. Out in production, new IPs warm up receiver by receiver from fifty messages a day, rising after clean days and backing off when complaints, bounces, or throttling climb. Each provider gets its own connection limits and cooldowns, hard bounces and spam complaints suppress the address automatically, and outgoing mail is DKIM-signed and sent over TLS with MTA-STS and DANE verified where published. Templates, identities, configuration sets, tenants, dedicated IPs, and SES-shaped events through an SNS subset cover the SES surface your tooling expects, and a CLI explains why any given message is still waiting. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

Search is being answered by AI crawlers first and humans second, yet most site owners cannot name a single bot that read their pages last week; Geo-Monitor answers that from the access logs you already have. A small collector sits beside Nginx or OpenResty, mounts the log read-only, and streams only new lines in batches, resuming after outages and surviving log rotation without ever touching your site's code. The console separates today, the last week, and the last month into identified requests, crawled pages, active crawlers, and trend lines, then drills into every visit with time, path, method, status code, claimed identity, and verification state. Identification and verification are deliberately different things here: a user agent only proves what a request claims to be, while a verified badge requires trusted log addresses and official IP evidence to agree. Twenty built-in identities cover OpenAI, Anthropic, Perplexity, Kimi, and Bytespider crawlers, and you can add, edit, or disable rules as new bots appear. Your sitemap syncs or imports as the content catalog, so coverage views show which pages crawlers actually reached, and collector health exposes heartbeat, backlog, and parsing problems. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

Training, meals, and body changes usually live in three different apps, and LiftCut puts them in one bilingual tracker where an AI agent helps keep the log instead of just drawing charts. Tell it what you ate, what you lifted, or what the scale said in plain language, in Chinese or English, and it prepares the entry from the exact numbers you gave, never inventing nutrients or weights, and saves nothing until you confirm it. Workout plans are editable rather than rigid programs, text import brings existing routines in, logging tracks completion and exercise performance, and personal records and PDF exports mark how far you have come. Food logs roll into nutrition summaries, and dated body measurements draw trend charts beside your training, so cause and effect sit on the same screen. AI training and nutrition proposals read your saved goals, restrictions, measurements, and recent actual loads first, then pass through extraction, strict validation, and an editable preview, and a proposal is rejected if the records it relied on changed in the meantime. Begin in the public no-sign-in demo or local-first guest mode, then create an account and migrate your history when ready, with offline queues holding entries until a connection returns. Providers span DeepSeek, any OpenAI-compatible endpoint, or a local vLLM server, so the intelligence can stay on hardware you control. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

Which agent should answer a question that just arrived in your QQ group, and how much of the conversation should it see? RabiRoute is the self-hosted message gateway that answers those routing questions before your agent ever wakes up. Chat messages from QQ groups and direct chats (through NapCat and OneBot), webhooks, schedules, voice, desktop selections, and device events all arrive at one gateway, and message Routes decide which agent or local program receives each one. Every role keeps its own message history plus references to plans, memories, and skills, while each Route limits how many recent messages travel with a delivery. Roles can also fire agents on interval, time-window, daily, or one-time schedules, or run a configured local script, so recurring checks and reports happen without anyone typing a prompt. Replies to QQ, RabiLink, and other channels leave through a single sending API that validates targets, quoted messages, and sender identity, then records receipts for everything delivered. The Windows desktop side adds selected-text actions, screenshots, and annotations; speech, phone, glasses, wearable, and remote-computer clients connect as experimental integrations, and the Linux Host serves the Manager and browser WebGUI from a Node.js process. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

Small teams do not need another cloud subscription to know who is doing what this week: Work-Calendar runs your whole plan, do, finish, archive loop on one little server you own. The week view lays out every member's tasks across seven day columns with today highlighted, so standups start from the same picture instead of five private lists. Three kinds of tasks cover real work: project tasks, one-off tasks you can assign to yourself or teammates, and fixed recurring tasks on daily, weekday, monthly, or specific-date schedules. Anything unfinished carries over to the next day automatically, and completion records can be corrected by the assignee or sent back with a reason by the owner. Projects get spaces with folders, files up to 20MB, meetings with AI-generated minutes from pasted or dictated transcripts, timelines, and closing reports that an admin approves. Roles for superadmin, admin, and member, plus per-group rules, decide who may create projects directly, who needs approval, and who can use the built-in AI assistant, which drafts task and project changes as reviewable cards rather than acting silently. Inbox notifications, personal diaries, and daily statistics round out the routine, and keyboard shortcuts drive the whole interface. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

Close the laptop, walk away, and your coding agents keep working; CodeRail is the remote control that lets you supervise them from a browser or your phone without handing your machines to anyone's cloud. A small bridge on each computer connects outbound to your server, so there are no open ports and no VPN to babysit, and each machine keeps its own Claude Code and Codex logins because CodeRail never holds your AI keys. Chats show plans, tool calls, models, modes, and background tasks, and when an agent wants to run something sensitive, the approval lands with you wherever you are, including as a phone push notification. Real terminals outlive the browser tab, files can be viewed, edited, uploaded, and downloaded in place, and git work, pull requests included, sits beside the conversation that produced it. Each chat even gets a browser on the computer itself, shared live with everyone in the chat and drivable by the agent. Roles run Owner, Operator, or Viewer per computer, narrowed down to folders and agents, with every change audited, and sessions keep running with every browser closed while history stays readable offline. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.