Open Source Agent Tools: October 24, 2026

1. Task-view Turns a Coding Agent’s Change Into a Reviewable Map Instead of a Flat Diff

Task-view. Task-view shipped v1.0.0 and v1.0.1 on October 23 and was posted to Show HN on October 24. When an agent finishes a task, it describes the functions, components, endpoints and tests it touched, and task-view renders that as one offline HTML page: a graph of the change, its user-visible flows step by step with attached tests, and each piece’s diff next to the agent’s stated reason. Every file, line and commit the agent cites is checked against Git, and changed lines the agent did not describe are flagged. It works with any language and with any agent that can follow written instructions, such as Claude Code and Codex. Licensed MIT. Source

2. Token Beggar Routes Claude Code Subtasks to Cheaper Models and Keeps Hard Calls on the Main Model

Token Beggar. Token Beggar is a Claude Code plugin that, before the main model takes a job, checks whether a cheaper model can do it equally well. It sends large tasks to one cheaper worker, then has the main chat review the full diff and run the tests, while design, root causes and the final answer stay with the user’s chosen model. The author reports that on 18 real bug fixes it used 1.42x less quota with about the same share of hidden tests passing (69% versus 70%). Version 0.3.0 was released October 23 alongside its Show HN post. It is Python with zero dependencies and is licensed MIT. Source

3. Lcc Compacts Claude Code Sessions Verbatim and Keeps Removed Tool Output One Recall Away

Lcc. Lcc (Local Context Compiler), posted to Show HN on October 23, is an alternative to Claude Code’s /compact that drops old tool outputs from context while keeping every user and assistant message byte for byte, with removed content retrievable through a recall command and a receipt for every cut. The author reports a median 21.8% token cut across 299 long sessions, and a pre-registered test on 172 requests where answers from the stashed context plus recall matched full context on file recall with 36.6% fewer tokens. It ships as a CLI, MCP server and Claude Code plugin via pip install local-context-compiler. Licensed MIT. Source

4. Slate Term Is a Native macOS Terminal Whose Panels Your Agent Can Build

Slate Term. Slate Term, posted to Show HN on October 23 after its v0.9.0 release on October 21, is a native macOS terminal built on libghostty that users can extend with custom HTML, CSS and JavaScript panels, commands and layouts, or ask a coding agent to build them. A floating indicator shows which agent is working in each pane and why, and saved layouts can be launched from Spotlight or the Dock as their own apps. It requires macOS 27 or later on Apple silicon. Licensed GPL-3.0. Source

5. Kucedr Relaunches the Friday Desktop Agent Under a New Name

Kucedr. Kucedr, previously called Friday, was relaunched on Show HN on October 23 as a local desktop AI assistant for macOS, Linux and Windows. It combines chat, agent, voice and media modes, lets users pick their own models, and extends through apps, MCP servers and skills, with more than 20 built-in tools for files, media and memory. It also supports scheduled tasks and A2A agents. Licensed MIT. Source

6. TenseMem Keeps One Current Fact per Slot for Agent Memory in a Single SQLite File

TenseMem. TenseMem, posted to Show HN on October 22 and October 23, is an embedded memory library for agents in which the LLM proposes typed facts and a deterministic resolver decides which value is current, so “I moved to Bangalore” supersedes “I live in Delhi” while history is preserved. Every read reports freshness (as_of_seq, pending, is_current), recalled context is fenced so the engine does not learn from its own output, and correct() reverts a wrongly applied fact. It runs in-process on one SQLite file with adapters for OpenAI, Anthropic, Ollama, LiteLLM and Gemini. Licensed Apache-2.0. Source

7. Mnemolith Flags Coding-Agent Memories as Stale When the Files Behind Them Change

Mnemolith. Mnemolith, posted to Show HN on October 22, is an MCP server that gives Claude Code, Cursor and Codex one shared memory per git repo. Memories can be linked to the files they depend on, and Mnemolith stores each file’s git blob hash so a memory is set aside for review once those contents change. A local dashboard shows each memory’s origin and history, contradictions are judged by the agent rather than a similarity threshold, and current memories can sync into AGENTS.md or CLAUDE.md for PR review. It runs locally on a custom Java engine with local embeddings, is at v0.2.1, and is licensed MIT. Source

8. XTrace Tracks Agent Time, API-Equivalent Spend and Usage Limits for Claude Code, Codex and Cursor

XTrace. XTrace, posted to Show HN on October 22 with its v0.1.8 release the same day, is an open-source macOS app that reads the Claude Code, Codex and Cursor history already on a Mac. It shows agent activity against estimated human time, API-equivalent spend by day and session, concurrency, and Claude and Codex account usage with a forecast to the next reset. A session timeline marks prompts and context compactions, and customizable scoring skills can be run against a project. Licensed Apache-2.0. Source