The knowledge graph that makes every team AI-native.
Guard maps your repositories into a deterministic knowledge graph and runs an agentic SDLC on top of it, so agents read pointers instead of entire codebases.
One graph, three outcomes
Deterministic and statistical: parsers, graph metrics and git history. No LLM builds it, and no source code is stored, only metadata and pointers.
Issue to pull request with guard rails: intake, implementation, independent review, a deterministic security gate and a human merge.
Token-budgeted context packs, a compact data map built for prompt caching, and per-run and monthly budgets.
From repository to reviewed pull request
GitHub App or PAT, github.com or Enterprise Server. Pick orgs and repos.
Analyzers build the graph: APIs, models, modules, dependencies, findings, history.
Claim an issue, label it guard:intake, and the agent writes a spec or asks questions.
Implementation, review and a security gate end in a PR. A human always merges.
Issue state machine
Pointers instead of whole files
Agents receive a context pack of node summaries and repo/path/line pointers, then read only the line ranges they need. The stable prefix (system prompt, tools, schema legend, data map) is prompt-cached.
89% fewer input tokens for the same task.
- A change that touches 3 modules in a mid-sized service.
- Naive approach: the agent loads 12 related files whole (≈7,500 tokens each).
- Guard: a 6,000-token context pack of pointers, then ≈4,000 tokens of targeted line-range reads.
- Excludes the cached prompt prefix (system prompt, tool definitions, data-map legend).