The 867-file checkout stores Patterns as Markdown
Across 867 files, Fabric stores recurring task instructions in Markdown rather than burying them in chat history. The supplied collection covers work such as summarizing material, extracting ideas, checking claims, and rewriting documentation. You can run a Pattern against standard input, a URL, or supported media, then change the model without rewriting the task itself. Custom Patterns live separately and survive updates to the built-in collection.
That separation is Fabric's best reason to exist. The 867-file checkout contained about 50,726 lines of source and occupied 28.3 MB, yet the core artifact remains a readable instruction file. A team can review one, put it in version control, and call it through a shell alias or the REST API. You can also copy the Pattern into another AI product without taking Fabric along.
Fabric lists 11 native provider integrations
Our install pulled 281 packages before we configured any of Fabric's 11 native provider integrations. OpenAI, Anthropic, Gemini, Ollama, LM Studio, Bedrock, Vertex AI, and several other services appear in the current README's native list. OpenAI-compatible providers add more choices. fabric --setup collects settings, while environment variables can map an individual Pattern to a vendor and model.
The large provider menu means a team must decide which credentials may live on each machine, which services may receive sensitive text, and which model each Pattern assumes. Fabric puts those choices in one tool. It does not settle them. Reproducible work needs the Fabric revision, Pattern revision, provider, model, and relevant generation settings recorded together.
What happened when we ran it
Using our test method, we installed commit 22577eb in 115 seconds on 3 CPUs with 8 GB of RAM. It was an unprivileged Go 1.24 Debian container with no secrets. Installation succeeded with 281 packages, and the build completed successfully in 266 seconds. The checkout had 4 CI workflow files, no Dockerfile, and no dedicated tests directory. These facts cover repository mechanics, not the quality of a model's answer.
The test command failed after 74 seconds. Go reported 68 passing packages and 3 failed packages out of 71. The supplied tail places the failures in internal/plugins/template and repeats Description: Test extension before the package and overall run are marked failed. Server, tools, converter, custom-pattern, notification, Spotify, and YouTube packages shown nearby all passed.
The tail does not include the failed assertion or exception. It cannot tell us whether the cause was Debian, missing software, or the template code. The narrower finding is enough: commit 22577eb installed and built in our fresh container, but its full Go test run was not green. Anyone extending the plugin system should reproduce the 3 failures before treating the suite as a release gate.
Three input paths can leave the local model
Our 3-CPU sandbox did not exercise Jina AI scraping, provider-backed search, or YouTube retrieval. Those paths can send work outside the machine even when Ollama or LM Studio handles the model, and hosted models need credentials too. A local model only protects the input when the retrieval tools also stay local.
The 28.3 MB checkout also includes helper paths beyond prompt execution. Media splitting and visual extraction can involve FFmpeg, while the to_pdf helper requires a LaTeX installation with pdflatex. Open issue 2144 describes empty output from --transcript and --transcript-with-timestamps; it was updated on October 1, 2026. If transcripts are central to the job, retain the source and provide another extraction route.
fabric --serve exposes configuration as well as chat
fabric --serve exposes chat completions, Pattern management, sessions, contexts, model lists, YouTube extraction, and configuration over HTTP. An Ollama-compatible mode presents Patterns as model names to existing clients. The server has an API-key option and a configurable address. Our 3-CPU run did not start this service, so its authentication and behavior still need a separate deployment check.
Our 266-second source build passed even though the 867-file checkout had no Dockerfile. The README instead points to published images on Docker Hub and GHCR. An organization that cares about image provenance should match the chosen image to an approved release and source revision. Running Fabric as a shared service also puts provider keys, spending, stored configuration, logs, and updates behind one network endpoint.
An October 1 release and October 2 issue activity show maintenance
GitHub showed 44,147 stars, 28 open issues, and 21 open pull requests on October 2, 2026. The repository was pushed on October 1, and v1.4.505 shipped that day with another OpenAI-compatible provider. Issue 2246 was opened and discussed on October 2 with a concrete model-discovery reproduction. Recent code, a current release, and live issue work all point to an actively maintained project.
Fabric is easiest to justify when you want a reviewed Pattern to outlive today's model choice. Start with a few tasks where repeatability saves time, pin what you run, and check the outputs against real examples. The 3 failed packages out of 71 keep us from recommending the extension path without a rerun, but the plain Markdown format makes the useful part easy to test and easy to leave with.