MiMo Code is Xiaomi's terminal-native coding agent, released on GitHub in June 2026 and now sitting at roughly 12,400 stars with 1,265 forks. It is a fork of OpenCode, and the project says so directly rather than burying the lineage — which makes the interesting question not what it inherited but what Xiaomi added on top. The answer is a persistent memory system, a deterministic workflow runtime, and a set of self-modification commands, all under MIT with an additional use-restrictions document. ## Memory as Infrastructure The headline addition is cross-session memory backed by SQLite FTS5 full-text search, and it is structured rather than a single append-only log. Project knowledge, rules, and architecture decisions live in `MEMORY.md`. Session state snapshots go to `checkpoint.md`, maintained automatically by a dedicated checkpoint-writer subagent. Agents get `notes.md` as scratch space, and per-task logs land in `tasks/<id>/progress.md`. What makes this more than a file convention is the context management wired around it. MiMo Code decides when to checkpoint based on the model's context window, and when context approaches the limit it rebuilds from the latest checkpoint plus project memory, task progress, and retained recent messages — so a long session degrades into a reconstruction rather than a truncation. Injection runs against a token budget with importance ranking, which is a more deliberate approach than dumping everything relevant into the prompt and hoping. A tree-shaped task system (`T1`, `T1.1`, `T1.2`) hooks into the same checkpoint machinery. ## Agents, Workflows, and Stop Conditions Three primary agents cover different postures: `build` with full tool permissions, `plan` as read-only analysis for exploration and design, and `compose` for specs-driven orchestration. Tab switches between them, and after the first message the mode locks — Build and Plan can still trade places, but Compose is isolated once entered. The stated reason is that a fixed skill and tool set from session start measurably improves tool-call reliability, which is a real tradeoff made explicit rather than a limitation hidden. Alongside conversational agents, MiMo Code ships a workflow runtime: deterministic JavaScript scripts that orchestrate multiple agents in a sandbox with fixed phase sequences, bounded retries, and automatic parallelization, running fire-and-forget with no user interaction. Four are built in. `compose` runs Brainstorm through Merge, parallelizing independent tasks into isolated git worktrees with TDD per task. `deep-research` plans research angles, runs parallel subagents for cited findings, then cold-reviews the citations. `fact-check` cross-checks extracted claims with a three-juror adversarial vote. `research-experiment` runs an autonomous hypothesize-implement-evaluate loop against a verifiable metric and audits for metric gaming. Custom workflows drop into `.mimocode/workflows/` or `.claude/workflows/`. The `/goal` command is a smaller feature with an outsized point behind it: it sets a stop condition, and when the agent tries to finish, an independent judge model evaluates the conversation to decide whether the condition is genuinely met — an explicit guard against premature optimistic stops during autonomous runs. ## Skills, Voice, and Self-Improvement Twenty-four builtin skills ship with the CLI, spanning `arxiv`, `deep-research`, `pdf-official`, `pptx-official`, `xlsx-official`, `modern-python-toolchain`, `frontend-design`, and `skill-creator`. Two are conditional: `claude-code` and `codex` appear only when those executables are installed, letting MiMo Code delegate work to competing CLIs. Skill matching runs on exact name, localized alias, and BM25 relevance, auto-loading high-confidence matches and ranking uncertain ones for the agent to judge. Two commands handle self-improvement. `/dream` scans recent session traces, promotes durable knowledge into project memory, and prunes stale entries. `/distill` finds repeated manual workflows and packages the high-confidence ones into reusable skills, subagents, or commands. There is also streaming voice input via `/voice`, built on TenVAD and MiMo ASR, and an experimental Max Mode running parallel best-of-N reasoning with judge selection. Setup is a one-line install script or `npm install -g @mimo-ai/cli`. Authentication options include MiMo Auto — an anonymous zero-config channel, free for a limited time — plus Xiaomi platform OAuth, Codex OAuth for ChatGPT Pro and Plus subscribers, one-step import of existing Claude Code credentials, and any OpenAI-compatible endpoint. ## Pros and Cons The strengths are architectural. Structured persistent memory with budgeted injection and context reconstruction addresses the failure mode that most affects long agent sessions. The workflow runtime is a genuine capability rather than a wrapper, and the judge-model stop condition shows attention to how autonomous loops actually fail. Provider flexibility is unusually broad, and the free zero-config channel removes the usual first-run friction. The caveats deserve weight. Being an OpenCode fork means inheriting upstream architecture decisions and the ongoing question of how divergence gets maintained. MiMo Auto is free only for a limited time, with no announced pricing after. MIT licensing sits alongside a separate use-restrictions document, so it is not unencumbered MIT. The ASR model is available only on MiMo's platform, making voice input effectively vendor-locked. Windows and WSL users hit documented rough edges around clipboard handling and CJK code pages. And the compose mode lock, while justified, means an early wrong choice costs a session restart. ## Outlook The terminal coding agent category is crowded, and MiMo Code's differentiation is not model quality — it is willing to run against Claude, Codex, or anything OpenAI-compatible. The bet is on the harness: memory, workflows, and self-modification as the layer that matters once frontier models are commoditized. Fourteen contributors, releases roughly every two to three days, and a same-day final push suggest Xiaomi is resourcing it seriously. ## Conclusion MiMo Code is worth evaluating for developers who run long multi-session projects and feel the pain of agents relearning context each time, and for anyone who wants deterministic multi-agent workflows without building the orchestration themselves. Teams that need licensing certainty or predictable long-term pricing should wait for the free tier's terms to settle before committing a workflow to it.