Gimso2x Labs builds a source-linked knowledge layer for AI coding agents, with draft promotion gates and evidence-review labels. Session decisions, fixes and runbooks can become a searchable internal wiki that MCP clients, including Claude Code, can consult and cite.
The problem
Coding agents are now doing real engineering work, but what they learn disappears when the session ends. Teams re-debug the same incident, re-explain the same conventions, and lose hard-won context in chat transcripts. Naive "agent memory" makes it worse: unreviewed notes turn into confidently wrong instructions.
Root causes, rejected approaches and operational gotchas live in transcripts nobody re-reads.
Snippets cached from last month quietly disagree with the code and configs running today.
When an agent writes its own memory, nobody can tell which facts were verified and where they came from.
Product
Our platform turns agent work into durable, reviewable knowledge and serves it back to agents through the Model Context Protocol.
Search, read and propose tools exposed over MCP, so Claude Code and any MCP client can consult team knowledge mid-task.
Budget-bounded task and resume packs that reread canonical wiki files on each request and return source references with content revisions.
MCP proposals create concept drafts. Separate create-only promotion gates use redaction and pre-publish checks; evidence-review labels record review status.
Pages can carry sources linking to the sessions, commits or messages they came from, alongside indexed inbound and outbound page links.
Full-text ranking plus bounded one-hop linked excerpts, tuned for mixed Korean and English engineering notes.
Remote MCP uses authenticated access and per-client project ownership controls for writes. The current deployment shares read access.
How it works
Session summaries, commits and incident notes are collected automatically as agents work.
An LLM CLI pipeline drafts concise concept pages: what happened, why, and what to do next time.
Drafts use create-only promotion gates, redaction and pre-publish checks. Evidence-review labels record review status.
Agents request context packs over MCP, with canonical wiki excerpts and content revisions for citation.
Claude & MCP
Status
We use the platform across our own engineering, with an internal knowledge base of more than 700 pages from real sessions and commits. We are preparing a private beta for small engineering teams.
| Capability | State |
|---|---|
| Wiki MCP server (search, read, propose) | In use |
| Context Packs: canonical file rereads and revisions | In use |
| Draft → promotion gate with redaction | In use |
| Authenticated MCP with project-scoped writes | In use |
| Hosted multi-team workspace | Preparing private beta |
| Direct Claude API distillation | Planned integration |
한국어 소개
Gimso2x Labs는 AI 코딩 에이전트의 작업 기록을 출처가 연결된 팀 위키로 정리하고, 초안 승격 게이트와 근거 검토 상태 라벨을 제공하는 개발자 도구를 만듭니다.
에이전트는 MCP로 위키를 검색하고, 요청마다 정본 위키 파일을 다시 읽어 내용 리비전과 출처 참조를 담은 컨텍스트 팩을 받습니다. MCP 제안은 초안으로 생성되며 정식 문서 승격은 별도 게이트를 거칩니다.
원격 MCP는 인증과 프로젝트별 쓰기 제어를 제공하며 현재 읽기 접근은 공유됩니다. 세션 요약 합성은 일반 LLM CLI 파이프라인을 사용하고 Claude Code와는 MCP로 연결할 수 있습니다. 직접 Claude API 연동은 계획 단계입니다.
현재 내부 엔지니어링에 쓰고 있으며, 소규모 개발팀 대상 비공개 베타를 준비 중입니다.
Contact
We're looking for a handful of teams who run Claude Code daily and want their agents to stop forgetting.
hello@gimso2x.com