vibehacker
News
GitHub (chengmatt416/umbra) ·

Umbra is an Android agent that operates your apps in a hidden display while you keep using the phone

Shown on HN Nov 4, it reads the screen, taps and types in an invisible background display, and also builds and tests code in an unprivileged on-device Debian environment, using your own Anthropic, OpenAI, Google, OpenRouter or OpenAI-compatible models. Payments, messages, deletions, installs and passwords always wait for your approval, and a single setting stops everything and revokes its access.

More news

View all

Engraphis is an Apache-2.0, local-first memory for coding agents that cites its sources or abstains

pip install "engraphis[all]" gives agents an offline SQLite memory scoped by workspace, repo and session, mixing vector, lexical, graph and code aware recall packed to a hard token budget, with an MCP connection and a local dashboard at port 8700. It keeps superseded facts as history instead of overwriting them, and needs no cloud account or API key for local use…

Droneship runs Claude Code or Codex in a gVisor cloud sandbox where the agent never sees your keys

Shown on HN Nov 4, droneship run claude clones your repo into its own sandbox with Python, Node and Go preinstalled, while a gateway outside it allowlists hosts and individual MCP tools and injects credentials into allowed calls, logging every file, command and tool call. It's free in early access for up to 5 sandboxes and 30 sandbox hours a month, with Copilot, Gemini and Cursor support listed as coming soon…

Nomos is an Apache-2.0 Go harness that blocks coding agents from deleting tests or stubbing code to go green

Shown on HN Nov 3, the single 28MB binary sits between your repo and Claude Code, Cursor or OpenCode, locks main read only with Git hooks, runs each task in its own worktree, and only merges after nomos verify passes 42 AST gates (deleted assertions, skipped tests, fake shims, leaked secrets) plus your own tests. It runs locally with no account or telemetry, and the maker says it's overkill for weekend prototypes…

Elpis v0.5.0 runs checked multi-agent Codex workflows on Linux, with you choosing what each agent sees

Shown on HN Nov 3, the open source Linux terminal and desktop harness built on Codex CLI lists every instruction file and memory with its token cost so you include or drop each one, and runs sequential, parallel, graph, recursive or voting workflows where each agent works in its own folder or Git worktree, gets checked, and a parent agent reviews the combined result before it lands. There are no macOS or Windows builds, and outside Git the folder boundaries are only instructions…

Spotted something we missed? Start a thread.