
Codeflash
Codeflash is an AI agent that rewrites Python, JavaScript, TypeScript and Java functions for speed, proves behavior is unchanged with generated regression tests, and ships each gain as a mergeable pull request. Free plan, then 20 USD per user monthly.
What is Codeflash?
Codeflash is an AI performance-engineering service built around codeflash-agent, an autonomous optimizer that reads your repository, rewrites slow code and proves the rewrite is both correct and genuinely faster. The homepage promise is blunt: Cut your infra bill by 90%. Then keep it there. The pitch rests on a diagnosis, namely that most companies sit far behind the performance-engineering frontier, that cloud bills are typically 40 to 90 percent reducible, and that coding agents have made matters worse: Codeflash says it found 118 functions up to 446 times slower than necessary in just two AI-written pull requests.
The agent works on Python, JavaScript, TypeScript and Java. It installs through each language's package manager, pip, uv or poetry, npm, yarn, pnpm or bun, uv with Maven or Gradle, and reads its settings from pyproject.toml, package.json, pom.xml or gradle.properties. The optimizations are algorithmic rather than architectural: better algorithms, redundant computation removed, caching, cleaner logic, more efficient library methods. The documentation states plainly that it does not modify the system architecture of your code, although Codeflash claims the agent understands abstractions well enough to rewrite a six-step flow as a three-step one.
Correctness is the pivot. Every candidate rewrite runs against your existing tests plus regression tests the agent generates itself, inside an isolated sandbox with no production access and, per the vendor, no exfiltration paths. Real speed is then measured by runtime benchmarking, GPU benchmarking included. What arrives is a mergeable pull request carrying the reasoning, the percentage gained and the correctness evidence, reviewed first by senior performance engineers, because the model is deliberately hybrid: humans don't drive it; they steer it.
Published results include a 90 percent bill cut at Unstructured, from 10,000 to 1,100 dollars a month, RF-DETR inference five times faster, a 13.7x token-decoding gain merged into vLLM, and merged pull requests at pydantic, langflow and Hugging Face Diffusers. You reach it through a CLI, a GitHub Action, a web app, a VS Code extension or a Claude Code plugin, on SaaS, your own cloud or on-premises.
What it does
- Find the fastest correct version of a slow function and rewrite it
- Optimize a single file or function, an entire repository, or a complete end-to-end workflow
- Optimize every new pull request automatically through a GitHub Action
- Prove correctness by generating and running regression tests alongside your existing test suite
- Measure the real gain through runtime benchmarking, GPU workloads included
- Deliver a mergeable pull request carrying the reasoning, the percentage gained and the correctness evidence
- Watch for performance regressions, close stale optimization pull requests, and catch repository-wide wins that file-by-file work misses
When to use Codeflash / When not to
A quick filter to help you decide if Codeflash is the right fit.
When to use Codeflash
- Engineering teams carrying a cloud bill they suspect is 40 to 90 percent larger than it needs to be; the homepage pitches the saving as the cleanest dollar a CFO will see this year
- ML and inference teams: the published work covers RF-DETR, SAM3, YOLOv8, vLLM and Hugging Face Diffusers, with GPU optimization and custom CUDA kernels
- Maintainers of open source libraries, who already have optimization pull requests merged at vllm, diffusers, pydantic and langflow, and whose public GitHub projects fit the free plan
- Teams shipping code written by coding agents, where Codeflash says it found 118 functions up to 446 times slower than necessary in just two AI-written pull requests
- Large organizations with data-residency or compliance constraints, served by the on-premises and VPC deployment options
When not to use Codeflash
- Anyone working in a language Codeflash does not cover: C, C++, C#, Go, Rust, Kotlin, PHP and Ruby are all absent from the supported list
- Developers looking for a copilot to write features; Codeflash positions itself against other AI copilots and states that it does not modify the system architecture of your code
- Teams whose code does not live in a GitHub repository, since continuous optimization relies on a GitHub Action workflow and on the Codeflash GitHub app
- Anyone hoping to work without a terminal or a repository: there is no mobile app and no code-free path into the product
- Organizations that need private repositories, a Zero Data Retention guarantee or published pricing above 20 USD per user per month, since the free plan is public-projects-only and permits AI training while Enterprise is quote-only
How to use Codeflash
A typical end-to-end flow, from setup to results.
- Create an account from the Start Free button, which leads to app.codeflash.ai/login
- Install the client with your language's package manager: pip, uv or poetry for Python, npm, yarn, pnpm or bun for JavaScript and TypeScript, uv for Java
- Declare your settings in the project file Codeflash reads: pyproject.toml, package.json, pom.xml or gradle.properties
- Optimize one function first to see how it behaves: codeflash --file path/to/file --function my_function
- Widen the sweep to the whole repository with codeflash --all
- Trace and then optimize a complete workflow with codeflash optimize myscript.py
- Turn on continuous optimization with codeflash init-actions, which installs the GitHub Action workflow that runs on your own runners for every pull request
- Install the Codeflash GitHub app so that the codeflash-ai bot account can execute those actions
- Review each proposal in the pull request, which carries a detailed explanation, the percentage gained and the correctness evidence; the Staging-Review option lets you vet optimizations before any GitHub pull request is opened
- Work from the VS Code extension, the Claude Code plugin or Cursor instead of the terminal, or take the Book a Call route for a twenty-minute diagnostic call
Pros & Cons
Pros
- Gains you can audit yourself: optimizations merged upstream at vLLM (#20413), Hugging Face Diffusers (#11665), pydantic and langflow
- Correctness is verified rather than promised, with generated regression tests run alongside the existing suite
- Delivery inside the workflow you already have, since everything arrives as a pull request with no process to change
- Every optimization is reviewed by a senior performance engineer before it reaches you
- Strong data posture on paid plans: Zero Data Retention, no training on your code, and zero-retention agreements with the LLM providers
- SOC 2 Type II certified, with the report and the penetration tests available under NDA
- Three deployment modes, SaaS, your own cloud and on-premises, described as the same product, and on-premises can point LLM calls at your own Azure OpenAI, OpenAI Enterprise, Anthropic or compatible endpoints
Cons
- Scope is limited to Python, JavaScript, TypeScript and Java, with no systems language covered
- The site contradicts itself: the pricing FAQ still claims Codeflash currently only optimizes performance for Python, while the documentation ships installation and configuration pages for all four languages
- The free plan permits AI training on your code, covers public GitHub projects only, restricts tracing to local runs and caps you at 25 credits a month
- Privacy mode and the Zero Data Retention pledge are reserved for paying customers
- Heavy dependence on GitHub: a GitHub Action workflow plus a GitHub app with write access to Checks, Contents and Pull Requests
- No DPA published or offered, no Article 27 EU representative, and a GDPR claim that stops at "GDPR Readiness"
- Enterprise pricing is not public and on-premises is Enterprise-only; the Azure hosting region is never named, SOC 2 and pentest reports come only under NDA, Aggregated De-Identified Data and Usage Data stay freely exploitable by the vendor, and the postal address differs between the homepage and the other pages
Pricing & Plans
Codeflash publishes a permanent free plan at 0 USD, presented as free with no hidden fees. The cheapest paid entry point is the Pro plan at 20.00 USD per user per month, preceded by a 14-day free trial, and the Enterprise tier is quoted on request. The terms of service specify monthly invoicing in advance, payment within 30 days, amounts in United States dollars, non-refundable fees, interest of 1.5% per month on late payment, and quota overages billed pro rata at the applicable overage rates. A monthly/yearly toggle appears on the pricing page, but the annual figure is injected by JavaScript and was therefore not collected.
- 25 function optimization credits per month
- community support
- public GitHub projects only
- local tracing only
- AI training permitted
- 500 optimization credits per user per month
- advanced optimizations and developer code quality metrics
- priority support
- private GitHub projects
- no AI training ever
- dashboard and user analytics
- unlimited credits
- no AI training ever
- admin usage analytics
- on-premises deployment
- 24/7 premium support
- custom SLAs on proposed optimizations
- onboarding and optimization discovery
- plus a shared Slack channel opened for enterprise customers
Data, GDPR & hosting
A consolidated view of how Codeflash handles your data.
GDPR overview
The privacy policy, updated on 10 July 2025, lists the GDPR under a heading stating that CodeFlash maintains compliance with industry security standards and regulations, next to SOC 2 Type 2, CCPA alignment and the NIST Cybersecurity Framework. Read closely, the wording is softer than the heading: the item is titled "GDPR Readiness" and says only that privacy practices are designed with GDPR principles in mind. Concrete measures are named, including privacy by design, data minimization and mechanisms for exercising rights, and the policy enumerates access, rectification, erasure, portability, objection and opt-out, answered within 30 days at privacy@codeflash.ai. Transfers from the EEA, the United Kingdom, Asia and Switzerland rely on standard contractual clauses or another valid mechanism. What is missing is just as concrete: no DPA published or offered, no Article 27 EU representative, no named DPO, no Do Not Track support, minimum age 16.
Who owns the data?
The terms of service (23 July 2025) draw a clean line: the customer owns and retains all right, title and interest in the Customer Data, and Outputs, the agent's code suggestions, are explicitly part of it. CodeFlash keeps the intellectual property in the Product itself. Internally, and without disclosing it, the vendor may use and modify Customer Data and Usage Data to operate the Product and to generate Aggregated De-Identified Data, which it can then exploit freely for improvement, testing, operation, promotion and marketing. Feedback is licensed perpetually and royalty-free. Some features pass Customer Data to third-party generative AI providers, which you must not feed personal data. Paid privacy mode adds a no-training pledge.
Reuse rights
Because Outputs belong to the customer, the optimized code can be merged, shipped and sold as part of your own software without asking CodeFlash for anything. The licence to the Product itself is far narrower: non-exclusive, limited, personal, non-sublicensable, non-transferable and for internal use only. Reverse engineering, derivative works of the Product, resale, rental or transfer, use for a third party's benefit and building a competing product are all prohibited. Outputs carry no warranty at all: they are supplied as is, neither promised non-infringing nor necessarily protectable, and the AI clause requires you to verify them before relying on them. Customer Data and Outputs may be deleted after 90 days, irreversibly once an account has been unpaid for 90 days or more. CodeFlash may also cite your name and logo as a commercial reference unless you say otherwise. (terms of service, 23 July 2025)
Data retention & training
Hosting summary
The backend runs on Microsoft Azure, presented as Microsoft's enterprise-grade security and compliance framework, and Codeflash states that no sensitive data is persisted there. No Azure region is ever named, which leaves the exact hosting country unconfirmed; the privacy policy says only that the company is US-based, that US law governs, and that information may be transferred, processed and stored in the United States and other jurisdictions. In cloud mode with privacy mode on, the Codeflash client scans and runs your code inside your own environment and sends the backend only the snippet to optimize and its context. LLM work goes through Azure OpenAI Service and Anthropic Claude under zero data retention agreements, Auth0 handles authentication for app.codeflash.ai, Stripe processes payments under PCI-DSS with no card data stored, and the optional Staging-Review feature holds optimization information for 30 days. The terms add that the Products are installed in the customer environment, which CodeFlash does not host. On-premises puts every Codeflash service inside your infrastructure or VPC, with LLM calls pointed at your own endpoints and no telemetry sent back. SOC 2 Type II is audited annually, report under NDA.
Things to keep in mind
Risks and trade-offs to weigh before adopting Codeflash.
- Automated rewriting of production code: human review of every pull request stays essential, the vendor itself warns that Outputs can be wrong and must be verified before you rely on them, and every published gain figure comes from the vendor, so reproducing your own baseline remains prudent
- The GitHub app receives write access to Checks, Contents and Pull Requests on your repository
- On the free plan, AI training is permitted, so code from public projects can feed model training
- No warranty on Outputs: they are supplied as is, with no promise of non-infringement or protectability
- Skill erosion: delegating optimization can quietly hollow out the in-house performance culture that the homepage itself says most companies struggle to prioritize
- A third party becomes a link in your delivery chain, sitting in your CI and in your pull requests, and in cloud mode your code is passed to third-party generative AI providers
- Contractual exposure: liability is capped at the fees of the last 12 months or 150 USD if none were paid, arbitration before the AAA is mandatory with a jury waiver, and Customer Data and Outputs can be deleted after 90 days
Setup & Integrations
Technical difficulty
This is a developer tool: expect a terminal and a code repository. Installation goes through your language's package manager, pip, uv or poetry, npm, yarn, pnpm or bun, uv for Java, and configuration lives in a project file, pyproject.toml, package.json, pom.xml or gradle.properties. Continuous optimization adds one command, codeflash init-actions, plus a GitHub app install with write permissions. No performance-engineering expertise is required: Codeflash advertises no manual effort and no expertise needed. The VS Code extension, the Claude Code plugin and the web app are gentler routes. On-premises is another matter, a real infrastructure project, with Enterprise onboarding included.
Deployment
Integrations
Behind Codeflash
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
How is Codeflash different from an AI coding copilot?
Which languages does Codeflash optimize?
How much faster will my code actually get?
How does Codeflash make sure the rewrite is correct?
Is my code used to train models?
Where is my data processed, and what security certifications back that up?
How does Codeflash plug into my repository?
Is there a free plan, and what does the paid tier cost?
Who owns the code Codeflash produces?
How do I get support?
Should you pick Codeflash?
Codeflash makes a claim that is unusually easy to check: it says it can cut infrastructure cost and latency, and it points at public pull requests merged upstream at vLLM, Hugging Face Diffusers, pydantic and langflow, alongside named customers including Unstructured, Roboflow, Pydantic and Langflow. That is a stronger form of evidence than a testimonial, and it is the main reason to take the product seriously.
Two things are worth understanding first. This is a hybrid offering: a self-service product, with a CLI, a GitHub Action, a VS Code extension and a Claude Code plugin, wrapped in an engagement where senior performance engineers review every optimization before it reaches you. And the technical scope is narrow and openly so, covering Python, JavaScript, TypeScript and Java, with no systems language and no architectural rewrites.
The data posture splits sharply by tier. Paid plans bring Zero Data Retention, no training on your code, SOC 2 Type II certification and an on-premises option; the free plan permits AI training and covers public repositories only. Choose accordingly.
Several points still need human checking. There is no published DPA and no Article 27 EU representative, the GDPR wording stops at "GDPR Readiness", the Azure hosting region is never named, the postal address differs between the homepage footer and the other pages, and the pricing FAQ still says Python only while the documentation covers four languages. The company is young, with the domain registered in August 2023 and a first Wayback capture in April 2024, and it is built around a single founder who came from NVIDIA and Cresta. Finally, every published speed-up figure comes from the vendor: reproducing your own baseline before and after remains the sensible move.
- Choosing a selection results in a full page refresh.
- Opens in a new window.