Silo's home page now fits its whole looping picture of a task in a laptop's window, at any width from 1024 pixels, with the headline and pitch above it and no buttons. Each picture in the loop now fades into the next.

Practical, open-source software on emerging tech.
Practical, open-source software on emerging tech.
Activity
- 1,732
- commits
- 125
- PRs merged · 4 from contributors
- 13
- issues closed
- 57
- releases
- in the last 30 days
Last 30 days · 4 PRs from contributorsamontlabsUpdated
Wed 18 Nov
In LCU, computer-use approvals in T3 Code are approval requests again, so only you can answer them, not an agent. Silo's home page now plays a looping picture of a task in seven steps. Fjord's inbox list now keeps loading as you scroll to its end.
Silo started a rebuild in the open, one feature at a time; a new public page shows its plan and takes comments through GitHub. lastdrop's main Claude account now has its own sign-in, so logging into Claude Code in a terminal no longer changes it.
Fjord's Router now keeps each version of an explainer, one per follow-up, and no longer stalls while its window is resized. atelier now runs a project's tests and its other checks at the same time. lastdrop fixed long agent runs failing on their next turn.
Fjord's Router can now send pictures and files with your messages, and can ask an agent to make a playable explainer of a thread's answer. LCU's computer-use approvals in T3 Code now offer Allow this conversation and Always allow.
atelier, which lets several coding agents build one project in parallel and merge without waiting on CI, got a public README and an MIT licence, and Silo started using it. lastdrop now retries brief server errors, then hands the work to another account.
Fjord's Router got a new page for T3 Code threads that shows each thread's subagents, background commands and waits. In lastdrop, a subagent that hits a usage limit partway through now hands its work to another account instead of failing.
harness-bench compared Claude Code with nanocode, an open-source agent, on the same model: on two sets of coding tasks, nanocode solved about as many at similar or lower cost. Silo added one page listing what agents ask for and copies waiting for review.
Silo made sure exported computers never include their memory, and brought computer use back on older virtual machines. Fjord's Router summary now keeps finished threads until you clear them and shows whether linked pull requests merged or closed.
lastdrop now starts each T3 Code thread on the Claude account with the most usage left, and moves it when that account runs out. Silo's container computers got a desktop, checkpoints, copies, a terminal and export. LCU fixed computer use missing in Claude Code.
Silo added container computers on Linux, an opt-in kind labeled less isolated than its virtual machines, and can now run GitHub Actions jobs, each in a fresh computer. Fjord's Router now shows where a message will go before it is sent.
Fjord's Router now works with T3 Code instead of Claude Code: it sends your words to T3 threads and brings their agents' questions to you. Silo can set up a Mac or Linux machine to host computers for another device, and a computer's desktop can now be turned off.
Silo can now run agents' computers on a Linux machine and send commands and desktop control to computers on another device. Its app got screens to review an agent's changes. Fjord's Router can now send one message to several threads at once.
Silo can now show what an agent changed in a copy of its computer, keep those changes in the original, and undo that. Fjord's Router window became a running summary of every project, where questions waiting for you are answered by number or by typing.
Fjord's Router now picks which Claude Code project or thread on claude.ai each message is for and sends it there, and can offer to create a new project. Silo released 0.12.0, which adds macOS computers next to Linux ones.
Silo, which runs each AI agent in its own virtual machine, switched agents' desktop control from the ChatGPT app and LCU to an open-source tool, and added tests where an agent tries to break out. lastdrop stopped macOS asking for access to Music and Documents.
LCU's build for Windows no longer fails at the very end when a temporary folder can't be deleted because another program is briefly holding a file in it.
LCU's Windows installer, not yet officially supported, works again with current ChatGPT app versions, where it had stopped. lastdrop fixed a full agents list making every agent fail, and its panel now marks the account subagents will really go to next.
lastdrop got a website and a launch video. Its panel now lists each agent's backup models under it, and its account limit settings now cover Codex, Z.ai and OpenRouter too.
lastdrop added a Stats tab with token use and what it would cost at API prices, by day, project, account and model. An agent can now have backup models that run if the first can't start, and the app now updates itself.
lastdrop now takes several accounts for every provider and runs agents on Codex, Z.ai and OpenRouter on the one with the most usage left. LCU's Claude Code setup stopped saving a path tied to one machine, so a skills folder kept in a repository stays the same on every machine.
lastdrop now moves a subagent that hits an account's usage limit to another Claude account with room, where it carries on. LCU stopped Claude Code from showing an error at the end of every turn when LCU wasn't connected.
lastdrop switched to running Claude Code's subagents on other Claude accounts that still have room, instead of opening a Claude window per account. It also let people set up their own agents on Codex, Antigravity, Z.ai or OpenRouter models, which Claude calls like its own.
lastdrop began as a menu bar app that shows how much usage is left on Claude, Codex, Antigravity, Z.ai and OpenRouter accounts. It also learned to open a separate Claude app window for each Claude account, and to show the limit resets each account can still use.
Fjord's Router now shows its answer as it is written, with its thinking and tool calls a click away. LCU fixed computer use for teammates in Claude Code agent teams, where every call had failed, and Junie on macOS now gets LCU's instructions from its first prompt.
LCU added experimental support for Junie, where the agent asks for each app and the person approves it with Touch ID or a password. On Windows, LCU now installs with the current ChatGPT Store app, including when that app is on another drive.
Fjord now delivers every message the same way, through another chat the Claude app runs, whether the target is running or asleep, and shows each one going from sending to delivered. LCU's update notices now reach the person only, not the agent's model.
Fjord can now wake a Claude Code chat that the Claude app isn't running, so a message sent to it still goes through. Retrying a failed send now resends it as it was.
LCU added experimental support for DeepSeek Harness and stopped Hermes sessions from growing their prompt by about 3 KB every turn. In Fjord, each session the Router follows now shows a line on what it is about overall and one on what its latest turn did.
LCU fixed approvals in the Claude app, where answering the card for a new app could leave computer use waiting about 30 minutes, mostly in auto permission mode. Fjord's Router now checks more messages for parts meant for different sessions.
Silo gives your AI agents their own isolated computers that you can branch, diff and fold back like a Git repo.
- Run each agent in its own machine with a terminal, desktop and computer use, away from your host
- Fork a machine at any checkpoint to explore in parallel, let an agent run wild or share with a teammate
- Review the state diff of any fork and fold the changes you want back into your main machine
Stage
- App
- macOS
- Prototype
- WIP
- Alpha
- Released
Last 30 days
- 872commits
- 97merged
- 1release
Since Aug 2026
- 3,426commits
- 25releases
- v0.12.04 Nov
On GitHub
- 9stars
- 1fork
- MITlicence
Built with
- JavaScript52%
- TypeScript36%
- CSS7%
Fjord sits between your T3 Code agents and GitHub, and brings you only what you need to know or decide.
- Type once: the Router delivers your message to the right thread, and brings back every question and plan that needs you
- Sort your GitHub inbox and CI by rules you write as sentences
- Hand a failing check to the thread that can fix it
Works with T3 Code
Stage
- App
- macOS
- Prototype
- WIP
- Alpha
- Released
Last 30 days
- 510commits
- 0merged
- 5releases
Since Oct 2026
- 1,903commits
- 25releases
- v0.13.017 Nov
On GitHub
- 0stars
- 1fork
- MITlicence
Built with
- Swift75%
- Python24%
lastdrop turns any AI subscription into a native subagent for Claude Code and T3 Code.
- Any subscription or API key becomes an agent Claude calls like its own, with its model, effort and fallbacks
- Balances the work across your accounts, on the one with the most usage left
- Adds auto mode where it is missing (Z.ai, Antigravity, OpenRouter): a judge checks every command
Works with Claude Code, T3 Code
Stage
- App
- macOS
- Prototype
- WIP
- Alpha
- Released
Last 30 days
- 203commits
- 1merged
- 36releases
Since Oct 2026
- 176commits
- 36releases
- v0.8.1817 Nov
On GitHub
- 3stars
- 1fork
- MITlicence
Built with
- Swift80%
- TypeScript13%
- HTML4%
atelier lets several coding agents build one repository in parallel and merges only what passed.
- Gives each task its own git worktree; a task can build on one that is not merged yet
- Merges approved branches as one batch once the full check passes on your own Mac and Linux machines
- Writes the same rules and task prompts for every agent
Stage
- CLI
- macOS
- Linux
- Prototype
- WIP
- Alpha
- Released
Last 30 days
- 72commits
- 0merged
- 1release
Since Nov 2026
- 72commits
- 1release
- v0.2.215 Nov
On GitHub
- 0stars
- 0forks
- MITlicence
Built with
- Rust100%
LCU gives any AI agent the computer use built into Codex, without the Codex app or its sign-in.
- Read windows, click, type and take screenshots in desktop apps
- Approve or deny each app the first time an agent tries to use it
- Read and control Chrome tabs through the official extension
Stage
- CLI
- macOS
- Linux
- Windows
- Prototype
- WIP
- Alpha
- Released
Last 30 days
- 63commits
- 27merged
- 14releases
Since Sept 2026
- 521commits
- 51releases
- v0.13.917 Nov
On GitHub
- 2,439stars
- 168forks
- MITlicence
Built with
- JavaScript67%
- Python25%
- TypeScript4%
Hartlib is one research chat over everything you subscribe to, for people who read professional briefings.
- Answers plain-language questions with each claim linked to its page
- Explains every answer with visuals shaped to how you understand best
- Reads dense briefings, letters and official records for you
Stage
- Website
- Prototype
- WIP
- Alpha
- Released
Last 30 days
- 1commits
- 0merged
- 0releases
Since Jun 2026
- 346commits
- 0releases
On GitHub
- 0stars
- 1fork
- MITlicence
Built with
- TypeScript82%
- PLpgSQL16%
Compares coding agent harnesses on the same model. Claude Code vs nanocode on Terminal-Bench 2.1 and SWE-rebench.
Sic checks the facts in French political debates against official sources, live or on replay.
- Watches the debate, identifies each speaker and what they claim
- Picks out checkable claims and skips opinions and promises
- Gives no verdict without a cited source and a precise link
Stage
- Website
- Prototype
- WIP
- Alpha
- Released
Last 30 days
- 1commits
- 0merged
- 0releases
Since Oct 2026
- 7commits
- 0releases
On GitHub
- 0stars
- 0forks
- MITlicence
Built with
- Python67%
- TypeScript32%