Several AIs in the same repository,
without losing rules, context or work
The knowledge base lives in your git. GuardianContext indexes, audits and reconciles, and each agent reads only what it needs.
Local MCP and safe updates
Published release · 0.1.9 · 2026-10-02
Sessions from different clients can discover peers and exchange messages within one repository. The desktop brings together MCP configuration, sessions, human messaging and update state.
Born from real incidents
One session's work disappears in another
Claude Code, Codex, Cursor and Kimi committing in the same tree. A git add -A picks up someone else's file, a checkout -- wipes the edit of whoever has not committed yet. That is how the product was born: lost commits and hours of wrong diagnosis in a real repository.
Context that costs more than the code
The work board of the source repository reached 245 KB, about 66K tokens read in every session, 43% of a 200K window, before a single code file was opened.
Each tool reads only its own file
Claude Code reads CLAUDE.md; Codex, Cursor and others read AGENTS.md. In the measured repository, security and contract rules existed for one family of tools and were invisible to the other.
Tokens your agents no longer spend
Measured on a real monorepo operated by four AI tools at the same time, with about 690 markdown documents. Tasks written the way an agent would ask; answer key taken from the repository itself.
| How the agent searches | Tokens per task | Found the right file |
|---|---|---|
| No knowledge base (grep and reading) | 57,976 | 2/12 |
| Code graph only | 33,248 | 5/12 |
| Well-maintained wiki with a router | 4,883 | 12/12 |
| GuardianContext | 4,699 | 7/12 |
Tokens estimated as characters ÷ 3.7, the same ruler in every scenario; reading capped at 2,000 lines per file. The no-knowledge-base and code-graph-only scenarios simulate a single search pass, without the iterative behavior of a real model. The tasks are about process and rules: a code graph is built for code questions, and the two tools answer different questions and add up. Against a wiki that is already well maintained, the gain is small: the value shows up where it does not exist or goes stale. Reproduce it on your repository withguardian benchmark.
Six pieces, one repository
Protocol in your git
Router, per-topic wiki, decisions and work board in markdown, versioned with the code. Works offline, shows up in the PR diff and does not create a second source of truth outside the repository.
Local snippet search
SQLite index with full-text and multilingual semantic search on your machine. The agent receives only the snippets that answer the question.
Audit and pre-commit hook
Measures the repository's discipline (bloated board, decisions without a status, expired claims) and refuses git add -A in a shared tree, without getting in the way of merges.
Reconciliation between tools
Shows, folder by folder, which rule each tool cannot see and reconciles CLAUDE.md and AGENTS.md with a pointer between markers, without copying content and without changing any tool.
Incident pattern library
45 rules curated from real incidents in repositories operated by several AIs (shared git, deploy, tests, database access rules). Each has the case and the command to check it.
Benchmark on your repository
guardian benchmark measures how many tokens your agents spend to find each answer, with and without GuardianContext. The numbers on this page came from it.
Cross-review with Claudex Loop
Claudex Loop is an open skill in which Claude Code and Codex take turns: one model plans and builds, the one from another company reviews in a fresh session, and whoever wrote the code never approves their own work. It is the scenario GuardianContext was made for.
The same rules for whoever builds and whoever reviews
Claude Code reads CLAUDE.md and Codex reads AGENTS.md. Without reconciliation, the reviewer judges by a different standard than the one the builder followed. guardian instrucoes sync links the two files.
Context ready for the session without memory
Every review starts from scratch. guardian brief delivers, in a few hundred tokens, the tool's rules, the snippets that answer the task and the related decisions, to paste into the plan.
The review log under control
PLAN.md and PLAN-REVIEW-LOG.md go into the local index, and the auditor warns when the round log grows beyond its budget, before it turns into a fixed cost of every session.
guardian instrucoes sync
guardian brief "task description" --tool agents --saida .guardian/brief.mdUse --tool agents when Codex reviews and --tool claude when Claude Code reviews; paste the brief into the context section of PLAN.md. Claudex Loop is an independent open source project with no commercial relationship with IronStack.
Start free, grow by repository
Plans differ by the number of repositories and users the account covers. The free plan has no time limit.
- 1 repository
- 1 user
- 3 repositories
- 1 user
- 10 repositories
- 2 users
- 25 repositories
- 5 users
- 100 repositories
- 20 users
Charged in Brazilian reais (BRL); international cards accepted. Dollar amounts are a reference at the Central Bank of Brazil PTAX selling rate of September 30, 2026 (R$ 5.1809 per US$); your card statement may vary with your issuer's exchange rate and fees.
One line, no need to talk to anyone
Installs the CLI and opens the Desktop Manager, where you create your account and connect the repository. Free plan for one repository, with no time limit. Your repository content stays in your git.
irm https://api.ironstack.dev.br/install.ps1 | iexcurl -fsSL https://api.ironstack.dev.br/install.sh | shTerminal only: uv tool install "guardiancontext[semantic]" and guardian signup. Terms of use at api.ironstack.dev.br/termos.