Loom
Large-scale Orchestration Of Models — a Go framework for data pipelines
powered by AI models.
Define how records are classified, filtered, and summarized. Loom handles
concurrency, provider rate limits, retries, model escalation, and caching,
with explicit budgets and permissions for each task.
Use it for workflows such as triaging support tickets, summarizing document
collections, or running several agents that share model capacity and results.
Website · Watch the demo · Documentation
Quickstart
You'll need Go 1.24 or later and Git.
git clone https://github.com/zionrubin/loom.git
cd loom
go run ./examples/triage
This example uses a deterministic mock model: no API key or paid model calls.
Go may download dependencies on the first run; the example itself runs offline.
It classifies five support tickets, keeps the three urgent ones, and produces a
briefing. The terminal shows model-call events, the briefing, and a per-stage
report of tasks, tokens, cost, and cache hits.
To see completed work reused, run it twice with the same state directory:
LOOM_STATE=./state go run ./examples/triage
LOOM_STATE=./state go run ./examples/triage
The second run replays cached results with zero model calls. Read the
complete example to see the pipeline and model setup.
How it works
A pipeline starts with records and connects stages such as Infer,
Filter, and ReduceAI. Loom turns those stages into tasks, schedules their
execution, and returns the output with a run report.
- Control spending: estimate calls and cost before a run, then apply token
and dollar budgets during execution.
- Handle model failures: retry transient errors and escalate invalid output
to stronger models using your validation rules.
- Reuse completed work: persist task results so unchanged work can be
replayed across runs and process restarts.
- Declare access: task envelopes specify permitted models, tools, secrets,
network hosts, and sandbox profiles; deployment policies constrain them.
- Scale when needed: run concurrent pipelines in a fleet, process streams,
or distribute tasks across worker processes.
Loom is a Go library. The default executor runs locally; an envelope is an
access declaration, not an OS sandbox. See the
architecture and isolation model for execution boundaries.
Where to go next
