About Framewren
Built to answer the question every LLM team has.
Why did our API bill double this month — and which feature caused it? Framewren exists because that question was too hard to answer, and it shouldn't be.
Our mission
LLM cost and latency shouldn't be black boxes.
When LLM-backed features were internal experiments, teams accepted opacity around API costs. A few thousand dollars a month was a reasonable R&D line item. Now those features are production: the load is real, the token counts are large, and $40K API invoices arrive with no attached explanation.
Framewren instruments the API layer — not the prompts, not the completion content — and surfaces per-model cost, P50/P95/P99 latency, and token throughput in a dashboard built for the questions platform engineers actually ask. We are not a prompt tracing tool. We are not a model evaluation framework. We do one thing: make LLM infrastructure as observable as any other infrastructure your team runs.
Founder
Devin Okafor, CEO & Co-Founder
In 2022, Devin was the infrastructure lead at a growing SaaS company in Portland that had just shipped its first LLM-backed feature — a document summarization pipeline running GPT-3.5. Within six weeks, that pipeline's monthly cost had climbed past $18K, with no alerting, no attribution, and no way to know which product team owned the spend. The only answer available was to grep application logs and manually calculate token counts per feature.
The tooling gap was obvious. Existing APM tools like Datadog tracked HTTP latency but had no concept of token cost or provider pricing units. Billing exports from OpenAI arrived as monthly CSVs with no project-level breakdown. The only remediation was manual per-provider instrumentation — work that had to be repeated for every new provider, every new framework, every new team that started using LLMs.
Devin left to build Framewren in 2023, spending the first year in Portland with co-founder Sura Chen building the SDK and cost normalization engine. The company received angel backing in late 2025 and has been shipping to production teams since early 2026.
Framewren is the tool Devin needed in 2022: a single SDK that wraps any LLM provider, normalizes cost and latency data across pricing models, and surfaces it in a dashboard built specifically for the questions engineering teams ask at 2am when the spend alert fires.
Team
Small team, sharp focus
Framewren is a small, early-stage company. Three people, Portland, Oregon — building something that saves engineering teams from the invoice they didn't see coming.
Devin Okafor
CEO & Co-Founder
Infra lead background. Built Framewren after managing a $40K+ LLM overspend with no visibility tooling.
Sura Chen
Co-Founder & Head of Engineering
Previously built distributed tracing systems at a cloud infrastructure company. Designed the Framewren SDK's transport-layer interception architecture.
Rafael Morales
Head of Product
Previously led developer experience at a growing API platform. Owns the dashboard, alert UX, and provider normalization model.
Engineering principles
What we believe about observability tooling
Precision over dashboards
P99 matters more than "latency looks normal." Cost attribution down to the request matters more than a monthly total. We build for the question that comes at 2am, not the quarterly review.
Cost transparency, not cost estimates
Different providers price by tokens, characters, or API units. We normalize everything. You see one number per request — comparable across OpenAI, Anthropic, Gemini, and every other provider you connect.
Zero prompt storage — by architecture
We intercept at the transport layer. Prompt text and completion content never leave your infrastructure and are never stored by Framewren. This is a design constraint, not a configuration option.
Developer-first, operations-ready
The integration is a single framewren.wrap() call. No infrastructure changes, no proxy deployment, no prompt routing. Engineers ship it in minutes; ops teams get the visibility they need.
We're building in public.
Follow the product on the blog. Read the changelog. And if something looks useful, start for free — no credit card, no sales call.