Open Source Agent Tools: October 23, 2026

1. Halv Runs Claude Code, Codex, Kimi and GLM in One Desktop Workspace and Meters Token Savings

Halv. Halv is a desktop workspace for macOS, Windows and Linux that runs Claude Code, Codex, Kimi and GLM side by side (up to 32 live panes) on the subscriptions users already pay for. It bundles a “Crux” code index for callers, references and impact analysis, local output filtering and context compression (RTK and Headroom), an OS sandbox with per-project approvals, turn diffs with revert, and a live savings meter. In an interim SWE-rebench test of 42 paired runs, both workflows passed 25 of 42 while recorded model cost fell from $337.50 to $144.67 (57.1% lower cost per correct answer), even though total tokens rose 248.9%; the project notes Halv cost more on four of 14 tasks. License: Apache 2.0 for the desktop app and local engine; an optional Halv Cloud service ($10/month) adds task routing and a savings dashboard and is not open source. GitHub · Product page

2. Adyen Engineers Release agent-marketplace to Sync Shared Agent Resources Across Eight Coding Tools

agent-marketplace (Adyen). Adyen engineers published agent-marketplace, a CLI that fetches shared agent resources (instructions, workflows and tool connections), converts them into each tool’s format, and writes them where eight tools read them, including Claude Code, Codex, Cursor and Gemini CLI. A single config file lists resources and target tools; agent-marketplace sync installs pinned versions idempotently, upgrades are a separate step, and settings merge across organization, machine, repository and personal layers. Adyen states it is not an Adyen product or service. License: Apache 2.0. GitHub · Announcement

3. Forked Brings Git-Style Branching and Effectively-Once Side Effects to Agent State

forked. forked versions an agent’s structured state (messages, memory, plans, tool results) as a content-addressed key-to-JSON tree with branches, commits, byte-exact rewinds and hash-based diffs, plus a side-effect log that replays stored results for recorded idempotency keys so emails, payments and API calls are not repeated on retry. The project reports 1,000 forks in about 1.4 ms in memory and about 276 ms on durable SQLite, with no state copied. It ships as an embeddable Rust library, a self-hosted server (one binary plus SQLite, or Docker) and a managed cloud on Cloudflare, with a TypeScript client and a drop-in LangGraph checkpointer. License: Apache 2.0. GitHub

4. Caudra Ships a Rust Terminal Coding Agent With Steerable Subagents in One Binary

Caudra. Caudra is a TUI coding agent written in Rust that began as a fork of Maki and packages steerable subagents, built-in token savings via command-aware output filters, a workbench, and a revertible history in a single native binary for macOS, Linux and Windows. Sessions open in Plan mode and only write a plan until the user approves, and a session can be left working toward a goal unattended. It works with 17 model providers or a local model and installs via a shell script, PowerShell or Homebrew. License: Apache 2.0. GitHub

5. Toli Connects Claude Code, Codex and Other Agents Across Machines Over A2A

toli. toli is a local daemon that lets agents in different harnesses (Claude Code, Codex, OpenCode, OpenClaw or custom ones) delegate tasks, ask questions and return results to each other over the open A2A protocol, on one machine or across the internet with a teammate’s agents. Each agent gets a name and token and reaches the daemon over MCP on localhost; only the user can add peers, with pairing exchanging a token in each direction, and incoming tasks wake the target agent. Tasks can be monitored in a TUI or CLI, and any MCP-capable agent, such as a personal assistant, can join. License: Apache 2.0. GitHub

6. Runspace Runs Agent Commands in Disposable GitHub Codespaces and Returns a Single Patch

runspace. runspace (v0.1.0) is a CLI and MCP server that executes commands in throwaway GitHub Codespaces, returns the output plus one patch of all changes since creation (commits, edits and new files), and then deletes the sandbox. It exposes seven commands (create, exec, diff, cp, run, ls, delete) through both front ends, connects over SSH through a Dev Tunnel with an in-memory key, and touches only codespaces named runspace. Codespaces usage is billed to the creating GitHub account, and the project lists limits including no egress control. License: MIT. GitHub

7. CATT 0.2.1 Tests MCP Servers and APIs Against Surrogate Consumer Assistants in CI

CATT (Coverage Cat). The Consumer Agent Testing Tool runs a scripted or model-driven agent against an MCP server, MCP app, website or HTTP API, checks the service-side outcome and the agent’s answer, and keeps sealed evidence for comparing builds. It emulates consumer assistants Muse and Instinct as surrogate hosts, covering routing, memory, payments, browser and HTTP tasks and scheduling, and includes scaffolding and connector-readiness checks; the maintainers state a passing test does not certify compatibility and the project is unaffiliated with those assistants. Version 0.2.1 was posted on October 22. License: MIT. GitHub

8. agent-usage-guard Holds Claude Code Subagent Fan-Out Before It Burns the Usage Window

agent-usage-guard. agent-usage-guard is a Claude Code mod that reads the plan’s live 5-hour and weekly usage percentages and intervenes before an agent starts, a prompt is sent or a tool call runs, capping subagent fan-out, heavy-context churn and retry loops across every Claude Code session on the machine. It asks the user through Claude Code’s own dialog, or refuses with an explanation when no one is present, and fails open if its own hooks break. The author built it after one max-effort prompt spawned 394 subagent calls (341 concurrent) and pushed 84M context tokens in ten minutes. It stores only counts, requires Claude Code v2.1.287 or later, and makes no network calls. License: MIT. GitHub