zed-industries/zedThe README describes Zed as a high-performance, multiplayer code editor from the creators of Atom and Tree-sitter. Two qualifiers this entry previously asserted need their sourcing marked. Rust is not stated in the README, it comes from the GitHub language metric where Rust is primary, corroborated by the Cargo manifests and crates layout. And GPUI, the GPU-accelerated UI framework, is absent from the root README and does appear as a repository topic tag, but it is also a first-party crate in this repository. crates/gpui carries its own README opening GPUI is a hybrid immediate and retained mode, GPU accelerated, UI framework for Rust, its Cargo.toml sets the same description, and it is published standalone on crates.io under Apache-2.0, which is also the Apache-2.0 component this entry alludes to. So the README silence is real and the conclusion drawn from it was wrong, because the project does make the claim, just not in the file a reader checks first. Licensing is confirmed as GPL-3.0-or-later with Apache-2.0 components where marked, which is why the GitHub API reports no single SPDX identifier and the metrics block reads other. Two ownership facts belong here and were absent. The README's own Sponsorship section states that Zed is developed by Zed Industries, Inc., a for-profit company, and that GitHub Sponsors money goes to Zed Industries as general company revenue. A hosted commercial tier sits in front of the GPL editor, zed.dev/pricing listing Personal at zero dollars capped at 2,000 accepted edit predictions, Pro at 10 dollars a month, and Business at 30 dollars per seat per month, the last omitted by this entry until the 29 August 2026 audit and the one that carries the most weight here, because a per-seat team tier is the commercial form of the collaboration positioning. The remaining first-party AGPL surface was relicensed to GPL on 28 May 2026.
nexu-io/open-designOpenDesign is a local-first desktop app that turns a coding agent already installed on the machine into a design engine, producing prototypes, dashboards, decks, images and video as real files with HTML, PDF, PPTX and MP4 export. It drives DeepSeek Harness, Claude Code, OpenClaw, Codex, Cursor, OpenCode, Qwen, Copilot, Amp, Hermes, Kimi and Antigravity, or any OpenAI-compatible endpoint via BYOK, and keeps a team's design system as a DESIGN.md the agents read and write, with 151 design-system packages shipped in the repository. A correction to the count this entry carried. It said more than twenty OTHER local CLI executables on top of the named list, which overstates the project's own figure in both directions, and the project's own figure does not agree with itself: the headline line says 26 distinct local CLI executables and the comparison table further down the same README says 25 CLIs, both inclusive of the names listed rather than additional to them. This entry now carries the range and the disagreement rather than either number. The desktop app is macOS and Windows only, and the README states that Linux users run it from source because the latest official release ships no prebuilt Linux artefact.
vllm-project/vllmThe repository tagline calls it a high-throughput and memory-efficient inference and serving engine for LLMs, and the README opens by describing a fast and easy-to-use library for LLM inference and serving. It was originally developed in the Sky Computing Lab at UC Berkeley per its README, and is built around efficient management of attention key and value memory with PagedAttention, plus continuous batching of incoming requests, chunked prefill and prefix caching. The kernel line names five optimised attention kernels, FlashAttention, FlashInfer, TRTLLM-GEN, FlashMLA and Triton, plus optimised GEMM and MoE kernels for various precisions. The quantization matrix is wide, covering FP8, INT8, INT4, GPTQ, AWQ, GGUF and more. It is a serving layer applications sit behind rather than an end-user tool.
astral-sh/uvAn extremely fast Python package and project manager written in Rust, from Astral, the team behind Ruff and ty. The README bills it as a single tool to replace pip, pip-tools, pipx, poetry, pyenv, twine, virtualenv and more, and claims 10 to 100 times the speed of pip. Correction on licensing: this entry previously said it ships under Apache-2.0, which is only half of it. The repository is dual-licensed, Apache-2.0 OR MIT, and the licence field in the metrics block shows only the one the GitHub API reports. The ownership fact this entry lacked entirely, and it is the load-bearing one. Astral announced on 19 March 2026 that it had entered into an agreement to join OpenAI as part of the Codex team, stating that OpenAI will continue supporting the open source tools after the deal closes, with the deal reported as pending regulatory approval. That makes uv's maintainer a team inside the company that also ships openai/codex, which this corpus tracks as a separate entity. Added 11 September 2026, the commercial layer this entry had no fact about at all. Astral sells pyx, a Python-native package registry described on its own page as from the creators of uv and named as trusted by Ramp, Intercom and fal, and that page now states pyx is no longer accepting new signups. A paid registry closing its funnel while its parent sits inside a pending acquisition is worth recording, because it is the only commercial signal available on a company whose open-source tools carry no price.
ollama/ollamaA tool for running open large language models, and this entry's own description of it now needs correcting against the current README. Four of the descriptors it carried are no longer supported by that file. Locally, single command and quantization each appear zero times, and the only CLI heading is Terminal and CLI, a subsection of the community-integrations catalogue rather than a reference for Ollama's own command line. Only REST survives, at a Ollama REST API section. Note the trap in the counting. Local returns ten case-insensitive hits and every one is a false positive, being either the string localhost inside a curl example or the name of somebody else's project such as LLocal.in, Obsidian Local GPT or Local shell copilot. What the README says about Ollama, in full, is Start building with open models. That is five words in a file of 19,225 characters, of which the Community Integrations section runs from line 153 to the end, 204 of 356 lines. and the model list this entry used to quote from the README is no longer there. Checked against the canonical README this run, **Kimi, GLM, MiniMax, gpt-oss and Qwen return zero hits each**, and only Gemma survives at seven, while DeepSeek now returns two, both on one line and both from the same link, the anchor text and URL slug of a DeepSeek Harness entry added to the ollama launch integration list on 13 August 2026. The README now routes readers to the hosted library at ollama.com/library and to docs.ollama.com rather than enumerating models inline, so a named-model list is a claim about the hosted catalogue, not about this repository. Correction as of 3 August 2026: the entry previously said Ollama had moved off llama.cpp to its own GGML-based backend since mid-2025. That is wrong. The repo lists llama.cpp as its supported backend, and the 0.30 release is specifically about deeper llama.cpp integration, opening the full GGUF ecosystem with Vulkan by default and NVIDIA acceleration, alongside an MLX runner on Apple Silicon. Re-verified on 13 August and both halves hold, with two refinements. The six absent names each returned zero on that date and Gemma still returns seven, and one string has since moved without being a model, because later on 13 August a commit added a DeepSeek Harness coding integration, so a raw count now returns two for DeepSeek, and the repository About field names exactly those absent models, Kimi-K2.6, GLM-5.2, MiniMax, DeepSeek, gpt-oss, Qwen and Gemma, so the split between what the README says and what the tagline says is confirmed from both sides. The Gemma count needs its reason rather than its number. Reading the seven lines shows Gemma is the worked example in the quickstart, a Run and chat with Gemma 4 section plus code samples using gemma4, not a survivor of a deleted catalogue. The llama.cpp correction is also confirmed and this entry nearly mis-read it. A token count returns two, which is one line counted twice, and that line reads as a credit to the project and its founder. Reading four lines up settles it, because the heading directly above is Supported backends and llama.cpp is its only entry. The commercial layer, corrected against the live pricing page on 11 September 2026 because two of the three figures this entry carried had changed. Ollama sells a hosted subscription, Ollama Cloud, priced Free at zero, Pro at 20 dollars a month or 200 a year, Max at 100 dollars a month, Team at 500 dollars a month, and Enterprise on request. Two corrections matter for anyone who costed this. Max is NOT paused, the page carries an active signup and no waitlist or sold-out notice anywhere. And Team is a 500-dollar flat monthly price for UNLIMITED users carrying 1,000 dollars of shared usage credits, not 25 dollars per seat with a five-seat minimum, so the word seat does not appear on the page at all. The per-seat reading inverted the economics of the tier, since a flat 500 gets cheaper per head as a team grows while a per-seat price does the opposite.
browser-use/browser-useA Python library that lets an AI agent drive a real web browser. Correction to the opening line this entry carried. It quoted a README one-liner about making websites accessible for AI agents and automating tasks online with ease, and that sentence is not in the README. Case-insensitive searches for accessible, make websites and with ease each return zero. The README describes itself twice, once as image alt text reading The AI browser agent, and once as a heading at line 85, Python library, the easiest way to automate the web. The quoted sentence is the repository description field, not the file. This is the second time this entry has attributed a repository metadata field to the README, after the Playwright note below. It lets an AI agent drive a real web browser, opening pages, clicking buttons, typing and filling in forms the way a person does, so a model can complete tasks on sites with no API, and it is MIT-licensed. A correction this run makes to the benchmark passage this entry added on 31 August, and the correction is that the passage no longer describes the file. That passage reported a named claim of first place on the Odysseys leaderboard at an 87.4 per cent average across 200 long-horizon web tasks, separated a rubric average from a success rate, attributed the phrase success rate to a separate BU Bench V1 chart, named OpenAI, Anthropic, Google and Microsoft as the compared computer-use agents, and reported a ChatBrowserUse model completing tasks 3 to 5 times faster. None of it is in the README today. Case-insensitive searches for Odysseys, 87.4, leaderboard, long-horizon and 3-5 each return zero on the current default branch. What the file carries now is a section headed Browser Use Benchmark v2, an image captioned mean rubric score by model and cost per task, and one sentence describing a very hard benchmark aimed at the hardest browser tasks with results drawn from a 60-task subset of BU Bench V2, linking to the project's own browser-use/benchmark repository. So the external scoreboard is gone and the measurement that replaced it is the project's own. The Browser Use Cloud reading survives with its specifics withdrawn: the hosted product is still prominent and the FAQ still sorts a reader between a fully hosted cloud, a CLI and the Python library, but proxy rotation and more than a thousand integrations are no longer stated, and proxy returns zero occurrences. The open-source library still sits in front of a commercial hosted product; the feature list this entry attributed to that product does not survive. The earlier sourcing note still holds, the entry once said it acts through Playwright while the README names no automation backend at all and Playwright appears only as a repository topic tag.
penpot/penpotPenpot describes itself as the open-source design platform for teams that build digital products at scale, giving full ownership of the design infrastructure. It is written mainly in Clojure by the GitHub language metric, runs in the browser or deployed on your own servers, and the README says it works with open standards like SVG, CSS, HTML and JSON, the last of which this entry previously omitted. Named capabilities are real-time collaboration, best-in-class native Design Tokens positioned as a single source of truth between design and development, an MCP server, CSS Grid and Flex Layout, which the README frames as enabling multi-directional workflows between design and code, and which it pairs with an open API and plugin system it says make the workspace programmable and enable automation, AI-driven workflows and integrations, so MCP is positioned as the on-ramp to AI rather than as an alternative to it. This round adds two things previously left out. The first is the commercial layer, Penpot offers a SaaS version alongside self-hosting and the README carries an enterprise approach reading that if your organization is scaling and needs extra support, we are here to help. The second is its two superlative claims, that it is the only design and prototype platform that is deployment agnostic, and that its native Design Tokens are best-in-class. Both are vendor statements, recorded here without endorsement. The licence is MPL-2.0, a weak copyleft, which differs from the MIT or Apache-2.0 carried by most of this corpus.
duckdb/duckdbThe README opens by calling DuckDB a high-performance analytical database system, designed to be fast, reliable, portable and easy to use, with a rich SQL dialect covering arbitrary and nested correlated subqueries, window functions, collations and complex types. MIT-licensed, and that comes from the repository licence metadata rather than the README, which contains no licence mention at all. **Two corrections to this entry's own earlier sourcing correction, which was applied too widely.** First, the phrase in-process, which this entry carried as README wording, does not appear in the README in any form, hyphenated, spaced or otherwise, at zero occurrences. It is the project's documentation and website framing, not this file. Second, this entry stated that Parquet, CSV and Arrow all come from the documentation rather than the README. **Parquet and CSV are in the README**, twice each, under a Data Import heading showing that both can be queried by naming the file directly in a FROM clause. Only Arrow is genuinely absent. The earlier note that SQLite for analytics is community shorthand and not the project's own words still stands, at zero occurrences. The ownership of the commercial side changed on 26 August 2026, and it is now the load-bearing fact about this project even though nothing in this repository records it. DuckLabs, the Amsterdam company whose name this entry corrected last round, announced that it will join Amazon Web Services as a new subsidiary, and the transaction has since closed: the project's own post now carries an update stating that the process concluded successfully on 31 August 2026, with the team joining on 1 September. The project’s own post states that DuckDB, DuckLake, Quack and all the other extensions remain free and open source under the MIT licence under the stewardship of the non-profit DuckDB Foundation, that there are no changes for the projects’ roadmap, licensing and governance model, and that the Foundation will include a technical advisory board so that leading community members can provide input on the project's technical direction, which is narrower than the stakeholder advisory board this entry previously recorded. Co-founders Hannes Mühleisen and Mark Raasveldt continue to lead the team and set technical direction, and the team stays in Amsterdam. Terms were not disclosed, and press reporting puts DuckLabs at over 30 employees.
deepseek-ai/deepseek-harnessAn open-source agent harness from DeepSeek, run from npm with `npx @deepseek-ai/dsh web`, which starts a local web UI on 127.0.0.1:3080 and opens it in the default browser. The README labels it a developer preview and warns in capitals that there will be compatibility-breaking changes. Licensed MIT. Correction to the architecture this entry described. It had named six components as interchangeable plugins around a small kernel, the model adapter, tool registry, sandbox, session store, event loop and web interface. The README carries no such list. At 2,201 bytes it says only that the project is built on an everything-is-a-plugin architecture and is powered by Cordis, whose design it cites to an arXiv paper. The enumeration lives in docs/architecture.md, and it names four things rather than six, the model adapter, the tool registry, the session log and the agent loop itself, each replaceable from configuration. Two of the six names were wrong rather than merely unsourced. It is a session log, not a session store. It is the agent loop, not an event loop. The second is the one worth correcting, because a replaceable agent loop is a stronger claim than a replaceable event loop, and it is the claim the project actually makes. Cordis is described as supplying a shared context, not a kernel.
ClickHouse/ClickHouseThe README describes ClickHouse as an open-source column-oriented database management system that allows generating analytical data reports in real-time. C++ is the implementation language by the GitHub language metric, and the share is deliberately not restated here, because a volatile metric belongs in the metrics block rather than in prose, the same correction already applied to shadcn-ui, renovate and superset. Two further claims this entry has carried, vectorized query execution and distributed processing, are NOT in the README and come from ClickHouse’s own documentation, which defines vectorization as query-plan operators passing intermediate result rows in batches rather than single rows to exploit CPU caches and SIMD, and describes sharding a table across nodes of a cluster with queries executed on every node holding relevant data. One layer this entry had omitted. ClickHouse, Inc. sells ClickHouse Cloud, a managed service in Basic, Scale and Enterprise tiers metered on compute per minute, on stored compressed data and on transfer, and CONTRIBUTING.md requires contributors to accept the ClickHouse Individual CLA, so the open licence sits on single-vendor governance.
ggml-org/llama.cppThe README states the main goal as enabling LLM and VLM inference with minimal setup and state-of-the-art performance on a wide range of hardware, locally AND in the cloud, which is broader than the local-only framing this entry previously implied. It is built on top of the ggml library, confirmed in the README. The hardware list is also wider than this entry carried: Apple silicon is called a first-class citizen optimised via ARM NEON, Accelerate and Metal, with AVX, AVX2 and AVX512 on x86, custom CUDA kernels for NVIDIA GPUs, and Vulkan and SYCL backend support. Two sourcing notes stand from the earlier audit and still hold. The README makes no claim of authorship over the GGUF format, and it does not state any Hugging Face or ggml.ai ownership, so the reporting that ggml.ai and llama.cpp joined Hugging Face on 20 February 2026 is external to this repository. It remains MIT-licensed.
openai/codexOpenAI's terminal coding agent, distributed as a standalone binary and an npm package, which reads, edits and runs code against a local checkout and executes shell commands inside a sandbox unless the operator grants wider access. Rust-majority by the languages endpoint, and the hardcoded percentage this entry carried has been removed, because a volatile metric belongs in the metrics block rather than in prose, the same correction already applied to shadcn-ui and renovate. The repository also holds the desktop app, reachable as the `codex app` subcommand, while the IDE extension is installed from OpenAI's developer site rather than built here. Licensed Apache-2.0. Sourcing note on the sandbox claim, which was checked this run and is true, though not where a reader would look for it. The README does not carry it. At 3,334 bytes it contains no occurrence of sandbox, approval or permission. The repository's own docs/sandbox.md is 150 bytes and consists of a single pointer to an external page, which itself redirects, so the prose authority for the claim sits several hops from the repository. The code is here even though the prose is not, because codex-rs ships sandboxing, linux-sandbox, windows-sandbox-rs and shell-escalation crates. Verified against the external documentation, the sandbox modes are read-only, workspace-write and danger-full-access with workspace-write the default, and the approval policies are untrusted, on-request and never. Two descriptions in this entry are wrong. The approval policies are not untrusted, on-request and never. There are three selectable today and a fourth in the code: on-request is the default, never is the second, and a granular mode configured as a table of per-category flags is the third and is missing from this entry entirely. untrusted survives in the Rust enum as an internal policy for directories marked untrusted, but the documentation states that Codex no longer supports it and that leaving it in a configuration file can stop the client from starting, so listing it as a user choice is advice that breaks a setup. The repository also does not hold the desktop app. It holds the installer and the back end. The codex app subcommand is a launcher that downloads a prebuilt Codex disk image from an OpenAI asset host and copies the bundle into Applications, with a flag whose help text is an override for the installer download URL, and a search of the tree finds no desktop UI source in any framework. What is built here is the protocol layer the desktop client talks to, app-server and app-server-protocol, so the shipped desktop binary is as external as the IDE extension while its server half is not. Label, metric and artefact disagree here, and the honest sentence names all three.