Open Source Agent Tools: October 19, 2026

1. WhatsApp MCP Lets Agents Search a Personal WhatsApp Account From a Cloudflare Worker

WhatsApp MCP. WhatsApp MCP, created and posted to Show HN on October 18, is a read-only MCP server that runs in the user’s own Cloudflare account and links to WhatsApp as a companion device, the same way WhatsApp Web does. Any MCP client (the README shows Claude Code via claude mcp add --transport http) can then search and read chats and store names for phone numbers, but it cannot send messages. It connects only on demand and once a day to collect queued messages, which are stored in SQLite inside a Durable Object, and the author recommends the Workers Paid plan because the initial history import exceeds the default CPU limit. It uses the unofficial Baileys library, which the README notes is against WhatsApp’s terms, and is MIT-licensed. Source

2. Rubi Adds Early Compaction and a Project Knowledge Graph to Coding Agents

Rubi. Rubi, first published on October 17 and posted to Show HN on October 18 alongside its v0.4.0 release, is a context manager for Claude Code, Codex, Gemini CLI, Cursor, Copilot, and local models through Ollama or LM Studio. It compacts sessions early (at 150k tokens by default) with a saved handoff, builds a code map ranked by PageRank over imports, and keeps a temporal knowledge graph of decisions, bugs, and gotchas that the agent writes itself and that is recalled with a hybrid of BM25, Personalized PageRank, and Reciprocal Rank Fusion. The author reports that 71% of cost in 284 analyzed Claude Code transcripts went to re-reading context, and that a mature graph cut steps and re-read context by 37% in an A/B benchmark on Flask and Hono. Everything is stored as markdown in a .rubi/ folder with zero dependencies, and the project is MIT-licensed. Source

3. Agentfork Replays an Agent Run From Any Step and Sweeps Candidate Fixes

Agentfork. Agentfork, posted to Show HN on October 17, records every LLM and tool call in an agent run, replays earlier steps from cache, and runs many variants of a single step several times each so developers get pass rates instead of single anecdotal runs. A sweep command varies the system prompt, model, or tool results at a chosen step against an expected output and prints per-variant pass counts, tokens, latency, and estimated cost. Instrumentation is three additions (a wrapped OpenAI or Anthropic client plus @tool and @trace decorators), LangGraph is supported, and traces live in a local SQLite file with a web UI for comparing runs. It installs from PyPI as agentfork-cli and is MIT-licensed. Source

4. Offscreen Moves Coding Agents’ Mac GUI Work Into Throwaway VMs

Offscreen. Offscreen, posted to Show HN on October 17 shortly after its v0.0.5 release, runs XCUITest suites, the iOS Simulator, and GUI apps launched by coding agents inside fresh macOS VMs (via Tart) so they do not take over the developer’s desktop. Agents, terminals, and scripts share one job queue through a daemon, and each job returns logs, screenshots, .xcresult bundles, and an exit code. An optional host lane runs jobs on the real desktop only while the user is away and requeues them when the user returns. It requires an Apple silicon Mac with macOS 14 and Xcode, installs through Homebrew, is labeled beta, and is MIT-licensed. Source

5. Shipback Gives Agents Undoable SFTP Deploys to Shared Hosting

Shipback. Shipback SFTP, released as v0.1.0 and posted to Show HN on October 18, deploys a folder to cPanel, Plesk, DirectAdmin, or any SFTP host and backs up every file it overwrites on the server before uploading. A --restore command reverts a deploy, refuses to overwrite files edited live on the server unless forced, and --pull --dry-run shows server-side drift since the last deploy. Its MCP server, aimed at Claude Code, Cursor, and other clients, has no option to delete server files or skip backups, only executes the exact plan the agent previewed, and keeps passwords in the OS keychain so they never reach the model. It installs from npm as shipback-sftp and is MIT-licensed. Source

6. Keel Builds One Event-Sourced Context Graph Shared by a Product’s Agents

Keel. Keel, posted to Show HN on October 18 with v0.29.1 following on October 19, is a context layer for agents embedded in a SaaS product: domain events go in through Postgres CDC, webhooks, Kafka, SQS, or typed SDKs, and a bitemporal context graph per tenant comes out. Each vertical is described once in a YAML “cartridge” (ontology, event catalog, rules, memory policies), every fact carries source and validity timestamps, and all agents use the same five tools (including context, history, and search) over MCP, an HTTP API, or SDKs for TypeScript, Python, Go, and PHP. It runs on an in-memory store, SQLite, or MySQL, ships as a container image, is marked alpha, and is licensed under Apache 2.0. Source