Xiaomi’s MiMoCode: A 5‑Person, 14‑Day Open‑Source Claude‑Code Clone

Five developers at Xiaomi’s MiMo team answered the question “Can a 5‑person team build a Claude‑Code‑style terminal programming agent in 14 days?” by creating MiMoCode, an open‑source AI coding agent (v0.1.0) that supports reading and writing code, executing commands, managing Git, and invoking sub‑agents for complex tasks.

MiMoCode extends the OpenCode fork with three primary agent modes: build for direct development, plan for read‑only analysis and design, and compose for full workflow orchestration. It also enables on‑demand sub‑agent creation, cross‑session persistent memory, automatic context reconstruction when the model’s context limit is reached, and a goal‑based stop condition judged by an independent referee model.

To preserve long‑task continuity, MiMoCode stores different state types in separate files: MEMORY.md (project knowledge, rules, architecture decisions), checkpoint.md (structured snapshot of the current session), notes.md (temporary agent notes), and tasks/<id>/progress.md (progress of each task). These files are indexed with SQLite FTS5 for full‑text search and injected into the context during task recovery.

The authors contrast this approach with merely enlarging the context window, likening a long context to a bigger table while persistent memory and checkpoints act as a project archive that determines which information should be retained long‑term and where to resume after interruptions.

Compose mode strings together planning, execution, code review, TDD, debugging, verification, and merging into an orchestrated pipeline, so the agent moves beyond generating snippets to advancing a complete software delivery process.

MiMoCode also introduces two memorable commands: /dream, which scans recent session traces to distill long‑term valuable information into project memory and prune stale data, and /distill, which extracts recurring manual steps into reusable Skills, sub‑agents, or commands—what the team calls “self‑evolution.”

Despite the promising features, the project is still at an early stage (v0.1.0). Stability of long‑running tasks, memory quality, sub‑agent scheduling efficiency, and real‑world experience across different models require further validation. The free “MiMo Auto” channel is time‑limited, and future pricing, quotas, and service policies may affect adoption. The MIT‑licensed source also carries usage restrictions prohibiting malicious network activity and unsupervised high‑risk operations, which developers must review before integration.