Saoud Rizwan

Saoud Rizwan founded Cline in 2024, building an open-source coding agent that can read and edit files, run terminal commands, use a browser and connect to external tools through the Model Context Protocol (MCP). His consequential decisions include packaging the agent as a VS Code extension, turning users’ planning workflows into a Plan/Act interface, and favoring visible, correctable agent behavior. In 2026, his team moved the extension onto a shared SDK runtime after an initial migration failed.[3][1][7]

In his September 2, 2026 migration account, Rizwan explained the team’s decision to replace the extension’s approximately 76,000-line private core with a shared, modular SDK runtime. The extension, CLI and other interfaces needed the same agent loop, tool execution, provider integrations and context management; sharing that foundation allowed improvements to carry across them. The first attempt took months and broke badly enough on release to require an immediate rollback.[7]

The team’s second approach bundled both the legacy and SDK-based engines into one extension, with a loader selecting between them through a percentage feature flag. If the new engine failed during activation, the loader started the legacy engine in the same window and pinned that machine to it. Both engines shared settings, credentials and task storage. Flag changes applied on subsequent windows, so an agent would not switch engines midway through a task; build checks kept the two engines’ views and settings compatible.[7]

Rizwan emphasized instrumenting the old system before comparing it with the new one. After more than a week near a 50/50 cohort split, his account reported that 6.34% of tasks on the old harness reached three consecutive mistakes and asked for human guidance, compared with 0.62% on the new harness, using a full day of production traffic per variant. This was a company-reported measure of that guidance threshold, with no sample counts or uncertainty intervals supplied. The rollout flag reached 100% on August 23, 2026, while legacy sessions continued to drain as installations updated and windows reloaded.[7]

That migration replaced a harness originally built around Claude 3.5 Sonnet in 2024. Rizwan’s earlier accounts explain how Cline’s first interface and developer controls took shape.[7][1]

In his January 2026 retrospective, Rizwan recalled spending a week in his mother’s garage in Terre Haute building an Anthropic hackathon demo. He gave the model tools he used as a developer: file access, terminal commands and codebase navigation. The demo did not win. In his July 2025 interview, he described reading Anthropic’s model-card discussion of agentic coding and building around the model’s improved ability to work through tasks step by step and use long context.[6][1]

Plan/Act grew from users’ own workflows. Rizwan observed developers first asking Cline to write Markdown architecture plans, then letting it implement them. Discord conversations helped turn that pattern into two explicit modes: Plan explores files, gathers context and clarifies requirements; Act runs commands and edits files. He described the interface as a prompt shortcut that spared users from repeatedly asking for a plan before implementation.[1]

Rizwan chose a VS Code extension for its access to local files and the terminal, existing diff views, Marketplace distribution and lower installation and organizational approval friction. It also let Cline work within editors such as Cursor and Windsurf. He regarded maintaining a VS Code fork as costly; collaborator Nik Pash separately argued that avoiding a fork let the team concentrate on the agent loop and support for new models.[1]

In the July 2025 interview, Rizwan argued for exposing prompts, tool calls, commands, edits and errors so developers could correct an agent while it worked. He expected better models to reduce the supervision needed, but saw back-and-forth interaction as especially useful for complex work requiring the developer’s architectural intent. He also preferred separate Cline and Cursor rules because he wanted the tools to behave differently for different tasks.[1]

Rizwan studied Anthropic’s MCP documentation and added extensive protocol instructions to Cline’s system prompt so models could build MCP servers before they already understood the protocol. For codebase exploration, he described a tool that exposed directory-level class and function names, helping the agent decide which full files to read. These mechanisms supplied context and connected the agent to tools without requiring every integration to be built directly into Cline.[1]

Rizwan’s July 2025 position was that stronger long-context and search-and-replace capabilities reduced the need for RAG and fast-apply intermediaries, which could introduce additional failure points. He favored giving the principal model fuller context despite higher costs. Cline then let users bring their own API keys and connect directly to providers. His formulation was “inference is not the business”: users should see which models they were paying for and what inference cost. Pash identified enterprise offerings as the revenue model.[1]

Asked in July 2025 whether he regretted making Cline open source, Rizwan answered, “Not at all.” He valued the experiments and product insights from forks and believed open source reduced adoption barriers. Attribution extended beyond the founder: Pash credited Nick Bauman with the earlier Memory Bank concept, while Rizwan’s January 2026 account credited the community with developing and validating the feature for retaining project context across sessions. Rizwan also described hiring some contributors full time.[1][6]

In a post dated January 30, 2026, Rizwan announced a commitment of $1 million in Cline credits for open-source projects, with proposed awards of $1,000–$10,000 in credits. Projects did not need to integrate with Cline. He particularly sought to support solo developers and small teams without access to traditional funding; the announcement established a credit commitment rather than completed disbursements.[6]

Before Cline, Rizwan also built software for iOS. His Swift library Disk provides a concrete example of that earlier work.[4][5]

Rizwan’s Disk library simplifies persistence of Swift Codable values, images and raw data. It encodes a Codable value as JSON, writes it atomically to a file and decodes it into the requested type on retrieval. Operations are synchronous, with background execution recommended for large file operations. Rizwan described the aim as extending Codable’s simplicity from network JSON to the file system.[5]

Disk’s documentation credits @benpackard for suggesting support for appending structs and thanks @nixzhu and @mecid for custom JSONEncoder and JSONDecoder support.[5]