For developers who work with agents
Stop making the agent rediscover the repository.
Kin is the system of record for AI-written software. It keeps a persistent graph of entities, relationships, changes, and provenance, so an agent can spend its context on the problem instead of rebuilding what the repository already knows.
Usable early alpha · Apache-2.0 · Runs beside Git today
The category matters
The graph is the standing source of truth.
Kin is not a graph index over files, a RAG layer, a visualization, or an agent memory cache.
The continuously updated graph is the source of truth for code, relationships, and history. Agents work from standing software truth instead of rediscovering the repository on every task.
- Today
- Kin runs beside Git.
- Long-term bet
- Replace Git and the disconnected tools used to adapt file-first systems for AI.
First-use challenge
Try it on a repository you already know.
Ask Kin a question you already know the answer to. Inspect what it returns, check the source, and judge it honestly. The point is not a polished demo. The point is whether standing repository context changes how your agent works.
Terminal input
curl -fsSL https://get.kinlab.dev/install | sh
exec "$SHELL" -l
cd /path/to/your/repository
kin init .
kin locate "where is authentication retry logic handled?"What to inspect
do not trust blindly- 01
Did it find the right entities?
Open the source spans and compare them with what you know.
- 02
Are the relationships useful?
Follow callers, dependencies, and provenance beyond the first match.
- 03
Did it report its limits?
Coverage and graph gaps should remain visible in the answer.
Initialization note
kin init reads complete reachable Git history. It can take minutes on a long-lived repository. Let it finish before you grade the answer. On a small repository, a fast kin init can outrun the 523 MB embedding model download in the background, so the first kin locate may run without semantic ranking; run it again after the download finishes.
The semantic loop
Ask once. Follow the graph.
After initialization, Kin answers from standing graph authority. It does not silently fall back to grepping files to make an empty answer look complete.
- 01
locate
kin locate "where is retry logic handled?"Rank relevant entities from graph-owned retrieval.
- 02
context
get_context_pack(entity_id)Assemble the focal code, dependencies, and dependents to a fixed budget.
- 03
trace
trace_data_flow(focal)Walk directional relationships without reopening the repository hunt.
- 04
review
kin review shadow BASE..HEADAttach relationship and impact evidence to the change. Review stays advisory.
Use the agent you already have
Put Kin behind your MCP client.
Kin exposes graph-backed tools through MCP for Codex, Claude, Cursor, Gemini, and other MCP-capable clients. Client integration and language depth vary, so test the path you actually use.
$ kin setup --intent agentKin detects supported clients and registers the local MCP server. Inspect the result, then ask your agent to locate behavior in the initialized repository.
- 01 Codex
- 02 Claude
- 03 Cursor
- 04 Gemini
- 05 Other MCP clients
Build it in the open
If the idea matters to you, help make it real.
Kin is Apache-2.0 open source. The wider open-core repositories welcome careful issues, hard questions, documentation, tests, language support, and code. Every contribution that makes the graph more trustworthy moves the whole model forward.
Current boundaries
Know the edges before you start.
Early alpha means useful paths and unfinished edges can exist at the same time.
- Strongest path
- macOS and Linux
- Start here for the most complete current experience.
- Early
- Native Windows x86_64
- Use WSL2 for the full experience today.
- Coexistence
- Kin runs beside Git
- Git remains part of the working path today.
Your repository is the benchmark
Start with one repository.
Ask one question you can verify. Then decide whether the graph earns another.