LatentGraph
Give your agent the bigger picture.
A structured view of your repository, dependencies and engineering knowledge helps agents work with codebase context, instead of repeatedly reconstructing it.
LatentStack is a modular AI software stack. This component is available on its own.
How it works
01
Code and dependency context
02
Relevant engineering knowledge
03
Context shared across coding tasks
Illustrative repository context
Works with
Agents
LatentCode, and supported MCP-compatible third-party coding agents. Context is supplied to the agent you already run.
Code hosts
Git-based code hosts, indexed per repository with your access rules preserved.
Models
Model-independent. Context is assembled before inference and routed through LatentGateway.
Supported-capability list v0.1 · owned by engineering · requires engineering sign-off before publication.
What it does not do
−It is not an agent. It supplies context; something else does the work.
−The published results measure repository-context task resolution and API cost on a named task set, not whole-factory productivity.
−It does not train model weights on your content.
Deployment note
Runs in your environment under on-premises and air-gapped configurations, so the index and the repository content stay inside your boundary.
What has been measured, and on what.
Claude Code tasks resolved
8 / 14
→
10 / 14
Without LatentGraph → with LatentGraph
14 selected hard tasks across two repositories. Company-reported repository-context evaluation, June 3, 2026; not a measure of whole-factory productivity.
API cost per resolved task
$3.93
→
$1.33
Without LatentGraph → with LatentGraph
Same evaluation. API cost includes failed attempts. Company-reported, June 3, 2026.
Works alone. Better with
Questions
Can we evaluate it on our own repositories?
We can scope a POC or paid engagement around one team, a repository and an engineering workflow, with agreed acceptance criteria and deployment and model requirements.
What evidence do you have?
Scoped, company-run evaluations with the method and limits published on the evidence page, plus one enterprise customer evaluation. None of it measures whole-factory productivity, and the site says so.