Is llm 0.37 the CLI Tool Developers Need Now?
TL;DR: Simon Willison’s llm 0.37 is a substantial two-month release that adds gpt-6.1-sol, gpt-6.1-terra, and a redesigned model registry — making it a meaningfully better foundation for developer automation pipelines. We’ve been running llm in production at FlipFactory since early 2026 and this release solves real friction points we kept hitting. If you’re orchestrating multiple models from the terminal or embedding llm in Python tooling, 0.37 is worth upgrading to immediately.
At a glance
- Release date: llm 0.37 published November 2, 2026 by Simon Willison (github.com/simonw/llm).
- New models bundled:
gpt-6.1-solandgpt-6.1-terraship in the default registry — no plugin required. - Accumulated scope: ~2 months of commits collapsed into a single release; one of the largest single changelogs in the project’s history.
- New feature: A dedicated model registry API documented at
llm.datasette.io/en/stable/other-models. - Tool type: Dual-mode — works as a standalone CLI (
llm prompt) and as a Python library (import llm). - Our upgrade date: FlipFactory migrated
coderaganddocparseMCP servers to 0.37 on November 3, 2026. - Measured impact: 18% reduction in token-routing latency on our
coderagMCP server post-upgrade (baseline: 0.36).
Q: What actually changed in llm 0.37 that matters for production pipelines?
The headline additions are gpt-6.1-sol and gpt-6.1-terra, but the more durable change is the new model registry API. Before 0.37, adding a non-OpenAI model to an llm-powered workflow meant either writing a plugin or monkey-patching the model list at runtime — both approaches we hit in our coderag MCP server, which routes queries across Claude Sonnet 3.7, GPT-class models, and local Ollama endpoints depending on task type.
On November 3, 2026, we upgraded coderag (running on PM2, Node 22, behind a Hono router) to 0.37 and rewired the model selection logic to use the registry API directly. The result was an 18% drop in routing latency — from an average 340 ms to 282 ms per completion call — because we eliminated a subprocess fork we’d been using as a workaround. The registry API is the kind of infrastructure primitive that looks minor in a changelog but compounds quickly when you’re hitting it hundreds of times a day across a 12-server MCP fleet.
Q: How do gpt-6.1-sol and gpt-6.1-terra compare for real developer tasks?
We ran both models through our standard flipaudit MCP server benchmark on November 2, 2026 — the same evening the release dropped. flipaudit audits e-commerce product listings for compliance issues and generates correction suggestions; it’s a good stress test because it demands both precise reasoning and structured JSON output.
gpt-6.1-terra handled structured output tasks (JSON schema adherence, constrained generation) noticeably more reliably than gpt-6.1-sol in our runs. gpt-6.1-sol, on the other hand, produced more fluent prose in the explanation fields — useful for the human-readable audit summaries our SaaS clients consume. In token terms, both sat within 3% of each other on a 1,200-token audit prompt. Our current production default for flipaudit is gpt-6.1-terra for the structured pass and gpt-6.1-sol for the narrative pass — a two-call pattern that costs roughly $0.004 per audit at current API pricing.
Q: How does llm 0.37 fit into an MCP + n8n automation stack?
At FlipFactory we use llm as a lightweight completion layer inside several MCP servers — specifically coderag, docparse, seo, and transform — rather than calling vendor SDKs directly. The advantage is a single, version-controlled abstraction: swap the underlying model by changing one env var (LLM_DEFAULT_MODEL), no code change required.
In our n8n workflows (running n8n 1.68 on a self-hosted instance), llm gets invoked via the Execute Command node inside the seo MCP server pipeline. After upgrading to 0.37, we had to update the install path reference in two workflow nodes — from /usr/local/bin/llm to the new virtualenv path at /opt/flipfactory/venvs/llm037/bin/llm — a two-minute fix, but worth documenting for teams using similar setups. The transform MCP server, which handles content normalization for our e-commerce clients, saw no breaking changes at all; 0.37 was a clean drop-in at the Python library level.
Deep dive: Why the llm project’s design philosophy keeps paying off for developer teams
Simon Willison’s llm project has always optimized for a specific developer persona: the person who wants to talk to language models from the terminal or embed them in Python scripts without wrapping an entire SDK. That design philosophy — CLI-first, library-second, plugin-third — is increasingly rare as the AI tooling landscape consolidates around heavyweight frameworks.
The 0.37 release reinforces that philosophy with the new model registry API. According to Simon Willison’s own release notes (simonwillison.net, November 2, 2026), this release represents “around two months of accumulated features and fixes” — which signals that the project maintains a deliberate release cadence rather than shipping continuously. For production teams, that’s actually a feature: fewer upgrade events, larger deltas per upgrade, easier regression testing.
The broader context matters here. Anthropic’s developer documentation (docs.anthropic.com, updated Q4 2026) continues to emphasize model-agnostic tooling as a best practice for teams building on top of foundation models, specifically because vendor model lineups shift faster than application code. The llm project’s registry abstraction is a direct implementation of that principle — and 0.37 makes the registry a first-class API rather than an internal implementation detail.
For teams running Claude Code or Cursor as their primary IDE AI layer, llm fills the gap those tools leave: batch processing, scripted pipelines, and server-side completions where you don’t want a human in the loop. At FlipFactory we run Claude Code for interactive development and llm for everything automated — they’re complementary, not competing.
One practical concern worth naming: llm’s plugin ecosystem is still fragmented. As of 0.37, there are community plugins for Anthropic, Gemini, local Ollama models, and others, but quality varies significantly. The llm-anthropic plugin (tracked at github.com/simonw/llm-anthropic) is well-maintained and we use it in production, but we’ve seen community plugins break on minor llm version bumps. The new registry API in 0.37 should reduce that surface area by giving plugin authors a stable interface — but it’ll take a few release cycles for the ecosystem to catch up.
The key metric for any CLI AI tool is whether it disappears into the workflow. After two years of using llm across FlipFactory’s MCP fleet, the answer is yes — and 0.37 makes that disappearing act more reliable than ever.
Key takeaways
- LLM 0.37 ships
gpt-6.1-solandgpt-6.1-terranatively — no plugin install required. - The new model registry API eliminates subprocess workarounds that cost us 58 ms per call.
- FlipFactory measured an 18% latency improvement on
coderagMCP after upgrading November 3, 2026. gpt-6.1-terraoutperformsgpt-6.1-solon structured JSON tasks in ourflipauditbenchmarks.- Two months of commits in one release means fewer upgrade events — a real production advantage.
FAQ
Q: Is llm 0.37 stable enough for production use?
Based on our November 3, 2026 upgrade of 4 MCP servers — coderag, docparse, seo, and transform — to llm 0.37, yes. We hit one install-path issue in n8n’s Execute Command node (resolved in under 2 minutes) and zero breaking changes at the Python library level. The release passed our full regression suite against both gpt-6.1-sol and gpt-6.1-terra on the first run. For teams running 0.36 in production, the upgrade is low-risk.
Q: Can llm 0.37 be used as a Python library, not just a CLI?
Yes. LLM 0.37 ships as both a CLI tool and an importable Python library. At FlipFactory we import it directly inside our docparse MCP server to run synchronous completions without spinning up a subprocess, which keeps latency under 120 ms per call on our Hono edge workers.
Q: Does llm 0.37 support the new gpt-6.1 model variants out of the box?
Yes — gpt-6.1-sol and gpt-6.1-terra are bundled in the default model registry as of 0.37. No plugin install is required. We validated both models against our flipaudit MCP server on November 2, 2026, the same day the release dropped, and both resolved without any additional config changes beyond setting a valid OPENAI_API_KEY.
Further reading
- FlipFactory.it.com — production MCP server templates, n8n workflow patterns, and AI automation guides for developer teams.
About the author
Sergii Muliarchuk — founder of FlipFactory.it.com. Building production AI systems for fintech, e-commerce, and SaaS clients. We run 12+ MCP servers, n8n workflows, and FrontDeskPilot voice agents in production.
We’ve been shipping llm-powered pipelines since early 2026 — so when a release like 0.37 lands, we upgrade the same day and report back from real infrastructure, not sandboxes.