Open Source Radar — September 20: where the agent actually runs
Today's board is unusually infrastructural. The fresh momentum is not another coding assistant but the layer underneath it: a model small enough to sit on the device, sandboxes that hand an agent a whole computer, a self-hosted control plane for the agents a company runs, and the parser that feeds documents to all of them.
Needle 3 (Python, ~11,800 stars) — Cactus Compute's third generation of its tiny automation model: the whole thing is a single 8–29 MB binary at roughly 2 bits per weight, and it does three jobs on-device — pick and fill tool calls, extract messy text into typed fields, and return sentence embeddings for local search. The trick is a ladder architecture where every depth from 2 to 20 layers is a deployable model, so a 2-layer subnetwork fine-tuned on your own tools runs on hardware the full model can't touch; the team reports fine-tuning lifts every subnetwork by 18 to 36 points, and from four layers up the tuned model beats DeepSeek V4 Flash at 29M parameters. Every response carries a calibrated confidence score, and a request no tool covers returns an empty list instead of a guess. One caveat worth knowing before you ship it: telemetry is on in the binary unless you switch it off.
Cua (Python, MIT, ~24,800 stars) — The open-source stack for computer-use agents, and the pieces are separable: a driver that sends clicks and keystrokes to a chosen window on macOS, Windows or Linux without stealing the cursor, sandboxes that boot Linux, Windows, macOS or Android machines through one API, and MCP and CLI surfaces so an agent you already run can drive them. The evaluation layer is the part most projects skip — the team's own Cua Bench ships expert tasks and its headline number is that the best frontier agent clears just 6 of 25 KiCad circuit tasks. If you are building or testing an agent that has to use real software, this is the plumbing; the published benchmark is the reason to take it seriously.
Coder (Go, AGPL-3.0, ~15,800 stars) — The self-hosted platform for cloud development environments, now aimed squarely at agents: workspaces are defined in Terraform, and Coder Agents runs the agent loop in the control plane on your own infrastructure, so model credentials never sit inside the workspace and every action carries a user identity. That inversion is the pitch — bring any model, keep the keys and the audit trail on your side, and use Agent Relay if developers want to keep the hosted agent experience they already know. The governance pieces that matter most, a prompt-level AI gateway and a process-level agent firewall, sit behind a paid Premium license, so budget accordingly.
Docling (Python, MIT, ~67,000 stars) — IBM Research Zurich's document parser, now an LF AI & Data project, and the quiet dependency under a lot of retrieval pipelines. Recent releases widened what it ingests to email, EPUB, Apple Pages, OpenDocument, XBRL financial reports and video with transcripts and keyframes, added chart understanding that turns a bar or line chart into a table or code, and kept the parts that make it survivable in a regulated shop: OCR for scanned pages, page layout and table structure recovery, local execution for air-gapped setups, and an MCP server so an agent can parse files directly. If your pipeline still loses tables when it reads a PDF, this is the fix.
Worth watching this week: whether the open layer under the agent consolidates on MCP as the one interface, or keeps growing a bespoke driver per project.
Would you rather run your agent in your own data center or hand it to a hosted sandbox? Tell us in the comments.
Sources: Needle (GitHub) · Cua (GitHub) · Coder (GitHub) · Docling (GitHub)
