Codium AI

Codium AI builds AI tools for testing, generating and reviewing software, with an emphasis on whether code matches developer intent. Co-founded by Itamar Friedman and Dedy Kredo, the company became Qodo on September 30, 2024. Its products expanded from IDE test generation into pull-request review and regression testing; by March 2026, it described its focus as enterprise code review and governance. That expansion began with code-level testing in 2022.[3][2][7][8]

Friedman recalled that the idea grew out of difficulty verifying code logic during his career managing research and development, alongside experience building LLM-based developer tools at Alibaba. After leaving Alibaba, he partnered with Kredo, whom he described as his first employee. In the May 2023 interview, Friedman framed “zero bugs” as a goal and explicitly said the product then helped reduce bugs rather than eliminate them.[1]

The early product was an extension for VS Code and JetBrains IDEs. It analyzed code and related components, generated tests, ran them with the developer’s approval, displayed failures and offered fixes or code changes. Developers could supply input data, reference tests and instructions to guide generation. The company’s launch article describes a Python closed alpha followed by JavaScript and TypeScript support.[1][6]

Codium combined static code analysis with TestGPT-1 language models for analysis and generation, drawing on code structure, comments and user guidance. The company described its solution and model as based on OpenAI foundation models. Friedman explained that the team benchmarked models and selected different ones for tasks such as semantic analysis and structured output. Its initial decision was to prioritize the surrounding interface, engineering and analysis before building independent models.[6][1]

A bank-account demonstration showed why intent mattered: Friedman changed a deposit operation from addition to subtraction, and Codium suggested restoring addition. He also raised the possibility that subtraction could be intentional in an unusual banking system. The developer therefore remained responsible for deciding what the code should do, rather than treating either the implementation or generated tests as automatically correct.[1]

Friedman called the broader vision “extreme DRY”, extending “don’t repeat yourself” beyond duplicated code to specifications, tests and implementations. A developer could begin with any of those artifacts and use agents to generate and align the others. Where they disagreed, the system would ask the developer which reflected the intended behavior. In May 2023, this was a proposed development workflow; the demonstrated product already generated analysis and tests from code and supported interactive test editing.[1]

Friedman also used Codium AI to contribute tests to AutoGPT and reported finding problems that way. He described advising and working with AutoGPT’s maker, Toran Bruce Richards. Codium’s approach at the time was to build narrowly designated agents that could create, run and fix tests, while AutoGPT explored agents operating toward more general goals.[1]

By May 25, 2023, Codium had emerged from stealth with a reported $11 million seed. The company identified Vine Ventures and TLV Partners as co-lead investors, naming Daniel Povitsky and Brian Sack respectively as directors, alongside 18 angel investors. In the May interview, Friedman reported 50,000 downloads and “high thousands” of weekly active users. He said both downloads and weekly usage had doubled every two to three weeks during an early period of roughly seven weeks, while cautioning that the pace might not continue.[1][6]

Tal Ridnik, Dedy Kredo and Itamar Friedman submitted Code Generation with AlphaCodium: From Prompt Engineering to Flow Engineering on January 16, 2024. AlphaCodium used a test-based, multi-stage iterative flow to solve competitive-programming problems from CodeContests. The authors reported that GPT-4’s pass@5 on the validation set rose from 19% with a single direct prompt to 44% with AlphaCodium. The result concerned that benchmark, rather than production defect reduction.[4][5]

The official implementation explains a consequential design choice: iterate on generated code rather than alter input-output tests to match the code’s predictions, using “test anchors.” The authors reported 15–20 model calls per solution and acknowledged that this could be too expensive for some applications. They also observed that longer contexts—roughly above 4,000 tokens in their experiments—could cause models to ignore details, despite fitting within an 8,192-token window. The flow therefore had to balance useful intermediate results against retaining the problem’s requirements.[5]

The September 2024 name change accompanied a broader product suite. Qodo Gen, formerly Codiumate, provided context-aware code and test generation inside IDEs. Qodo Merge, formerly PR-Agent, supplied AI-powered pull-request reviews, insights and recommendations. Qodo Cover generated regression tests to increase coverage. The company positioned these tools as parts of a code-integrity platform spanning writing, testing and reviewing code, while continuing to support switching between leading model providers.[2][7]

On September 30, 2024, Qodo announced a $40 million Series A, led by Susa Ventures and Square Peg, with Firestreak Ventures, ICON Continuity Fund, TLV Partners and Vine Ventures participating. It stated total funding of $50 million. Qodo also reported that more than one million developers had used its tools during the 18 months since emergence from stealth and that thousands of teams were using them; the developer figure measured cumulative usage.[7]ⁿ¹

By March 2026, Qodo described an enterprise code-review and governance platform built around three components: specialized agents handling bugs, rules and requirements; a context engine learning from codebase structure, pull-request history and business logic; and rules management capturing organizational conventions. The company described delivering prioritized findings and suggested fixes in pull requests, alongside review workflows inside IDEs.[8]

On March 30, 2026, Qodo announced that it had raised $70 million in Series B funding, bringing its stated total capital raised to $120 million. Qumra Capital led the round, with Maor Ventures, Phoenix Capital Partners, S Ventures, Square Peg, Susa Ventures, TLV Partners and Vine Ventures participating.[8]

Qodo announced PR-Agent’s handover to the community on April 23, 2026, separating the self-hostable open-source project from its enterprise platform. The announcement, credited to Dana Fine and Naor Peled, named Peled as the first external maintainer, with Ofir Friedman continuing on the maintenance team. Peled, Friedman and Fine were forming a governance committee; a possible foundation transfer remained prospective. The community repository identifies PR-Agent as community-owned and distinct from Qodo’s primary commercial offering.[10][9]