Open Source Agent Tools: October 20, 2026

1. OpenFilm Turns Coding Agents Into Video Editors Using One HTML File Per Film

OpenFilm. OpenFilm, posted to Show HN on October 19 alongside its v0.2.1 and v0.2.2 releases, has a coding agent such as Codex or Claude Code write a video as web pages: a film.html holds the tracks, and each shot is a page that draws any moment through a single frame(t) function using CSS, canvas, SVG, WebGL or three.js. A local browser-based Studio editor previews the film and supports manual edits that are saved back to the files the agent reads, with export to MP4, GIF, stems, subtitles and timelines for Final Cut Pro, Premiere Pro and DaVinci Resolve. It installs as an agent skill (npx skills add openfilm/openfilm) or as a desktop app for macOS and Windows, and generated voice, music and images use the user’s own provider keys. The repository has about 1,400 stars and is MIT-licensed. Source

2. Kompressor Cuts Coding-Agent Bills 22.7% With Cache-Friendly Deterministic Compression

OptScale Kompressor (Hystax). Kompressor, shown on HN on October 19, is a single Rust binary that sits as a proxy between a coding agent and the model and compresses tool output such as build logs, with removed details retrievable on demand. Its design rule is that the same input always compresses to the same bytes, so the provider’s prompt cache keeps hitting; the authors measured cache reads at 93% of billed prompt tokens on agent traffic. On an 89-task Terminal-Bench 2.1 run with Codex CLI on GPT-5.5, they report a 22.7% lower real bill ($83.40 vs $107.89) at 81.1% tasks solved vs 82.2% uncompressed, compared with 8.2% to 11.6% savings for Headroom. kompressor wrap codex starts the proxy and runs the agent through it. MIT license. Source

3. ApolloOffice Runs a Local Pipeline of Workers That Prove a Bug Fix Before Opening a PR

ApolloOffice. ApolloOffice, launched on Show HN on October 19, is a self-hosted service that picks up GitHub issues (or hunts for bugs itself), reproduces them, writes a fix with a regression test and opens a pull request through a fixed pipeline of workers (Scout, Guard, Vet, Hunter, Repro, Fixer, CI, Reviewer, Submit, Watcher). It verifies that the new test fails without the fix and passes with it, runs the project’s test command, and by default holds every PR for human approval. Each worker can use Claude Code or Codex logins, API keys for Claude, OpenAI or Gemini, or a local Ollama model, with Docker isolation, per-repo daily USD budgets and a table of merge rate by model and cost. The author reports 361 PRs submitted with 234 merged on one project. AGPL-3.0 license. Source

4. iphone-use-app Lets a Vision Model Operate the iPhone It Runs On

iphone-use-app. Created and posted to Show HN on October 19, iphone-use-app is a Swift and Rust iOS app in which an LLM controls the same iPhone it runs on, with no Mac at run time and no jailbreak. It mounts a developer disk image and uses a loopback VPN to reach the phone’s developer services, then loops screenshot to model to tool call (tap, swipe, type, open app); runs can start from chat, Siri, a Shortcut or a URL scheme. It needs iOS 26 with Developer Mode, a one-time pairing file made on a Mac, and an OpenRouter key or any OpenAI-compatible vision model with tool calls. It is sideload-only and cannot run alongside another VPN. AGPL-3.0 license. Source

5. Pi-cowboy Orchestrates Parallel Pi, Claude Code, Codex and OpenCode Subagents in Copy-on-Write Worktrees

Pi-cowboy. Pi-cowboy, posted to Show HN on October 18, is an orchestrator extension for the Pi coding agent that splits a job into subtasks and runs background agents in parallel, each in its own Herdr terminal pane and Git worktree, with Claude Code, Codex and OpenCode supported as foreign harnesses. Worktrees are created copy-on-write on APFS, BTRFS or ZFS; on a 153,607-file repository the author measured 75 MiB on disk versus 479 MiB for a plain git worktree add and 27 GiB when untracked files are copied. A default global limit of four concurrent agents queues the rest, and the orchestrator merges each agent’s branch and cleans up when done. Installs with pi install npm:pi-cowboy. MIT license. Source

6. Craig Is a Keyboard-Driven TUI for Running Claude Code, Codex and Cursor Agents Side by Side

Craig. Craig, shown on HN on October 19, is a local-first terminal UI that runs Claude Code, Codex and Cursor agents in real terminal sessions, giving each task its own branch and worktree and supporting workspaces that span several repositories. It shows diffs, syncs pull request and CI status, and exposes a CLI so agents and scripts can create tasks and read task context. There is no account or hosted control plane; it needs Node.js 22+, Git and an authenticated agent CLI, and installs with npm install -g craig-cli. MIT license. Source

7. Senseplane Gives Agents Look, Watch and Ask Tools Over Existing Cameras and Microphones

Senseplane. Senseplane, posted to Show HN on October 19 with v0.2.0 and v0.2.1 released the same day, is a self-hosted perception service that connects existing cameras, microphones and sensors so agents can take snapshots, set watches (“tell me when a package is left at the door”), query past events with clips, and get counts or durations of activities. It reports what it saw with confidence and evidence and leaves decisions to the agent, which connects through an MCP server, REST API, signed webhooks or a TypeScript SDK. It runs via Docker Compose with 7-day replay retention by default and is labeled early-stage. Apache-2.0 license. Source

8. Gojko Adzic Publishes a “Lights-On Software Factory” Boilerplate for Claude Code

Lights-on software factory. Gojko Adzic released a boilerplate on October 20 based on the agent setup behind his product Votito, where he reports the orchestrator cut token spend by about 80% compared with bare Claude Code workflow execution. It packages a container that shares only the workspace directory, a deterministic Bash command regulator that auto-approves or denies commands by project rules, a PostToolUse lint hook that enforces domain-specific ESLint rules on every edited file, and a generator that turns a plan into multi-agent workflow scripts with test gates so inference is not spent on mechanical steps. It is meant as both a starter and a teaching example. MIT license. Source