The Context Layer for AI Agents
A pair programmer for your agents that knows your platform and shows them how things are done.
Used for waitlist only. No spam.
Your best engineers are already cleaning up after their agents. Now imagine giving agents to everyone else.
Rework. Diverging patterns. Decisions revisited. Bugs reintroduced. At small scale, it’s annoying. At org-wide scale, it compounds.
The problem isn’t the agents. It’s that they’re blind. They don’t understand your architecture. They don’t know your standards. They don’t remember what your team has already learned.
Self-learning knowledge graph infrastructure for 1000s of autonomous coding agents
Control and govern at scale — during agent planning and execution. Git-first, instruction hierarchy, existing workflows.
Govern agent planning and execution states. Instructions hierarchy (AGENTS.md, skills, MCPs), review in PRs, promotion and demotion based on usage. Know what agents can do and what they're allowed to do.
Agent-agnostic MCP interface. Cursor, Claude Code, Copilot, custom workspaces — one connection.
Git-native, transparent memory + observation-driven graph enrichment. Context persists. Agents build on what others learned and how the system changed.
Git is the source of truth. Instruction hierarchy lives in your repo. Works with your branch strategy, CI, and tooling. No separate instruction store—it's all versioned and reviewed.
The graph learns from input and execution
Instructions, skills, and ADRs maintain a log in your repo—creating a feedback loop that keeps the knowledge graph current. Memory is an MCP tool that stores to files in the repo; it stays local until git ingestion. Personal memories sync to the server for cross-session continuity.
ADRs, instructions, and skills maintain a log in the repo. The graph ingests it, learns from it, and closes the loop.
Memory is a local MCP tool that persists to .md files in your repo. All memories are transparent, readable and versioned, shared via commit and push.
Personal memories sync to the server for continuity across sessions. Your preferences and context follow you.
Control and governance at scale—planning and execution
AGENTS.md, skills, and MCPs form an instruction hierarchy. Git is the source of truth; review happens in PRs.
AGENTS.md, skills, and MCPs define languages, libraries, and conventions. All versioned in the repo.
Repos roll up to domains, then org. Usage proposes promotion and demotion—putting you in the driver's seat of how agents build.
Review in PRs. No separate instruction store—agents read from files in the repo.
Connect git and other sources
Point ctx at your repos, monorepos, and tools. Git is the source of truth—works with your existing workflows, branch strategy, and instruction hierarchy (AGENTS.md, skills, MCPs).
System indexes and learns
The knowledge graph ingests your codebase, dependencies, and metadata. It learns from ADRs, instructions, and agent interactions. Indexing runs locally or in CI—your graph stays where you want it.
Connect your agent via MCP
Use the MCP server to connect Cursor, Claude Code, Copilot, or any agent. One connection, full graph access. Agents traverse the knowledge graph during planning and execution—with control and governance at scale.
Built for Agents + Engineers
Cline
Kilo CodeOpen source
Built in the open. ELv2 licensed. The core is open-source: audit it, contribute to it. We believe infrastructure this critical should be open.
Repo available soon. Currently in alpha.