Open Source Agent Tools: November 1, 2026

1. DeepSeek Open-Sources Super-Agent, a Research Prototype of a Proactive Coding Agent That Works From GitHub and Group Chats

DeepSeek published dsh-superagent-prototype on GitHub on October 31, a proactive coding agent for a single GitHub repository packaged as a bundle for DeepSeek Harness. It listens to GitHub and, optionally, Feishu group chats; when an allowed user asks for work, it opens a work item, runs it in its own agent session and working directory, and answers in the original conversation, for example with a pull request, while a web panel lists what needs a person. The stated design rule is to take as little human attention as possible: the agent acts only when asked or when a rule the team wrote allows it, and speaks only when someone must act or can use a result, with separate documents covering authorization, execution, prompts and deployment. DeepSeek labels it a research prototype rather than an official product, says much of the engineering was done by coding agents, and notes that it differs from the internal version described in its “Exploring Super-agent Design” article, with some of the article’s designs not yet implemented. Bloomberg reported the release on November 1 as DeepSeek’s bid to challenge Claude Code. Setup is meant to be done by asking an existing coding agent to follow the repository’s SETUP.md. License: MIT. Source Bloomberg

2. Prospero Is a Lean C++ Coding Harness With Subagents, Headless Remote Engines and Desktop Control

Prospero, announced by Chris Wellons on his nullprogram.com blog on October 31, is a coding harness written in C++ with a Dear ImGui docking interface that the author says uses about two orders of magnitude fewer resources than harnesses built on web stacks. It ships with no built-in model provider: users point it at llama.cpp, an on-prem model, OpenRouter, or the Anthropic or OpenAI APIs, and Claude subscriptions are not supported. Agents can spawn subagents that message each other, and given SSH or WSL access Prospero deploys a headless copy of its engine to other hosts so agents can work there, with the GUI able to detach and re-attach later. With a vision-capable model it can see and control the desktop (click, type and drag) to test GUI applications without a per-app MCP server, and on Windows it prefers the w64devkit shell while letting users switch shells mid-session. License: Unlicense. Source GitHub

3. Tadpole Is a 3,500-Line From-Scratch Agent Whose Post-Training Lifts Qwen3-4B From 0.419 to 0.774 on Unseen Tasks

Tadpole, posted to Show HN on October 31, is a tool-using agent written in about 3,500 lines of plain Python with no agent framework, plus a stage-by-stage record of how it was built, measured and post-trained. The agent handles files, shell, web search and fetch, sessions and compaction, and runs in a terminal or as an always-on Telegram bot against any OpenAI-compatible API (DeepSeek, OpenAI, OpenRouter, vLLM, LM Studio) or a local Qwen3-4B through Ollama or in-process llama.cpp; it executes shell commands with the user’s permissions and answers only one allowed Telegram user ID. Each build stage adds one mechanism with tests (601, none needing a model) and measured runs, from the first agent loop through SFT, expert iteration and a GRPO pilot. With the agent’s code and prompts frozen, SFT on verified teacher trajectories raised held-out pass@1 on unseen task types from 0.419 to 0.774 and cut false “done” claims from 31 to 1 in 420 runs; the three post-trained variants ended level overall, validation overstated expert iteration, and GRPO lost to expert iteration at equal rollouts. LoRA adapters and data are on Hugging Face under Shakewin. License: MIT. Source Show HN

4. Proxenos Is a Local-First Personal Agent That Tracks Health Care Paperwork and Drafts Messages Without Sending Them

Proxenos, posted to Show HN on November 1, gathers a user’s own messages, email, calendar, notes, documents and activity into a private vault on a Mac, builds a written “self-model” from it, and runs small single-purpose agents (“clones”) on top: a health care timeline with open items, call scripts and a one-page brief for the next appointment, PDF and intake forms filled from records (leaving ID numbers and signatures blank), task tracking, a daily journal, and reply drafts written in the user’s voice. It drafts but never sends, posts or buys by default; an opt-in per-app outbox can send only where past drafts went out nearly unchanged, with a cancel window, and phone calls (through Twilio) happen only when the user presses Call. Models run locally through Ollama (the README cites gemma4:26b on a 32 GB Mac), in which case nothing but call audio leaves the machine, or through an Anthropic API key, with observers sending only the user’s side of conversations. The author describes it as early and used daily by one person: about 56,000 lines of Python, TypeScript and Swift with 720 tests, macOS on Apple Silicon only, and not yet installable without a terminal. It is explicitly not medical advice. License: MIT. Source Show HN

5. CodeConnect Puts Claude Code and Codex Approvals on an iPhone Over Tailscale, With Face ID for Risky Commands

CodeConnect, posted to Show HN on October 31 with releases v0.15.1 on October 30 and v0.15.2 on November 1, is a daemon for macOS and Linux plus a free iPhone app that streams Claude Code and Codex sessions from several computers into one list, ranked by which agents are waiting on the user. Users start sessions with codeconnect claude or codeconnect codex; when an agent asks for approval, the phone shows the full command with a risk level, a read takes one tap, and something like git push --force needs a deliberate hold plus Face ID or the passcode, while a refusal stops the call and tells Claude the user said no. The app can also answer Claude’s questions, review diffs with per-hunk notes, or take over the live terminal. Traffic runs over the user’s own tailnet with no CodeConnect server or account, and by default the daemon listens only on the Tailscale address and loopback. License: MIT for the daemon; the iPhone app is source-available. Source Site Show HN

6. Clawdline Lists Every Claude Code and Codex Session on a Machine and Lets a Root Session Dispatch Work to Other Computers

Clawdline, posted to Show HN on October 31 and released as v2.10.0 through v2.12.0 between October 30 and November 1, is a Go control plane for macOS and Linux that reads Claude Code and Codex sessions already running in iTerm2 or tmux, with no hooks, MCP server or wrapper, and shows each as working, done or waiting on the user in a browser console. A root session can start child sessions locally or on a remote Linux machine and is notified when they finish, sessions can message each other through Clawdline, scheduled tasks can be triggered by webhooks, and an optional board tracks verification, merge and deployment separately from an agent’s own delivery report. Replies can be typed, dictated or sent as annotated screenshots. Local use is free and open source; an optional Clawdline Cloud adds a phone view across machines. The author notes overlap with Herdr, T3 Code and Orca. License: MIT. Source Show HN

7. korea-data-mcp Fixes 60 Traps in Korean Government APIs That Lead Agents to Wrong Answers

korea-data-mcp, posted to Show HN on October 30 and updated through v0.1.6 on November 1, is a single-binary Go MCP server with 20 tools over the official free OpenDART, data.go.kr, law.go.kr and Bank of Korea ECOS APIs: corporate filings, financials, dividends, major shareholders, stock prices, apartment sales and rents, business registration status, holidays, statutes and court rulings. Its pitch is correctness: it handles 60 documented quirks found in live responses, each covered by a test, such as OpenDART’s half-year report placing second-quarter revenue in the field an agent will read as the half-year figure (₩74.6 trillion instead of ₩153.7 trillion for Samsung Electronics in H1 2025), or law.go.kr’s “current” statute text including amendments not yet in force. In the author’s small evaluation, agents answered 10 of 10 questions with one call each using the server, versus about three times the calls and one miss against the raw APIs. It installs as an .mcpb bundle in Claude Desktop, is listed in the official MCP Registry, sends keys only to the government APIs and has no telemetry. License: Apache-2.0. Source Show HN

8. cc-pair Has Claude Code and Codex Review Each Other’s Changes Until the Reviewer Approves

cc-pair, posted to Show HN on October 30, pairs Claude Code and Codex so that one authors a change and the other reviews it, iterating until the reviewer approves or a round limit (three by default) escalates to the human. Roles are symmetric; in auto mode pair start --as claude or --as codex launches the other agent as reviewer each round without a second terminal, and manual mode works on Windows. The agents exchange numbered message files in a shared ~/.pair/<repo>-<hash>/ directory, so the full review thread can be inspected on disk, and a DESIGN.md specifies the protocol. It installs from npm as claude-codex-pair (Node 22 or newer) and needs both CLIs installed and signed in. The author’s premise, that an agent reviewing its own work goes easy on it, is anecdotal. License: Unlicense. Source Show HN