Skip to main content
Glama

Kin is a graph-native code repository for AI agents and the people who build with them. It stores source, recorded code relationships, and versioned history as repository state. The graph is the repository model, not a search index maintained beside another repository.

Functions, types, and the relationships between them are data you commit, branch, and merge. Exact source is preserved byte for byte, and filesystem projections let supported tools keep working with ordinary files.

Public beta. Try Kin on a real project you know well. Expect rough edges.

Quickstart · Documentation · Browser demo

A star helps other people find Kin, and Discussions is where to bring a question about it.

See what connects

Before changing a shared function, what else should you inspect? Kin lets you look up its recorded callers and explore related code. The CLI and MCP server query the same graph, so you and your agent can work from the same record.

Recorded on a prepared ripgrep graph at e89fff89ac9af12e8d4ce9d5fd07beb408ca730f. Raw run artifacts are not public. This illustrates a workflow, not a performance benchmark.

Related MCP server: GraphHub

Use it from Claude Code, Codex or Cursor

For Claude Code:

/plugin marketplace add kinlab-ai/kin
/plugin install kin@kin

For Codex, add the marketplace and install the plugin, or let kin setup --intent agent write the MCP server into ~/.codex/config.toml for you.

For Cursor, add the MCP server by hand; see plugins/kin-cursor for the exact snippet.

After installing, run kin init . in a fresh clone of a project you know well; the kin-setup skill walks you through the rest.

Kin is not published on crates.io. The kin crate there is an unrelated project.

Quickstart

Start in a fresh, full clone of a project you know well. kin init reads your Git history without changing it or your tracked files. It keeps its store outside your checkout, in ~/.kin/stores, and writes nothing into the repository; kin store path prints where a checkout's store is. By default it admits the checked-out commit first, so questions answer as soon as that tree is read, and then imports every commit your branches and tags reach in the background, at low priority; kin init --history full imports the whole history up front, and --history none never imports it. Importing history takes time, memory and disk that grow with history rather than with the size of the checkout. Before it starts, kin init checks memory and free disk and stops with the numbers when it can already tell the machine is short. A shallow clone admits only its checked-out commit, and submodules and Git LFS are not supported, and Git hooks, a sparse checkout or an unfinished merge or rebase also stop the import. kin init names the fix for each, and a fresh clone avoids most of them.

1. Install

On macOS, Linux and WSL2, one command installs Kin and connects the AI coding tools it finds. It needs Node.js 20 or newer:

npx -y @kinlab/kin setup

On a machine without Node, run the installer instead. get.kinlab.ai and get.kinlab.dev serve the same script:

curl -fsSL https://get.kinlab.dev/install | sh

Either way, reload your shell as a separate command, because the install puts kin in ~/.kin/bin and adds that directory to your shell profile for new sessions:

exec "$SHELL" -l

On native Windows x64, install from PowerShell with irm https://get.kinlab.dev/install.ps1 | iex. It installs the kin CLI and does not connect AI coding tools, because WSL2 remains the recommended path on Windows. The Windows entry under Beta limits says what works there. No native Windows ARM64 build is published. On an ARM64 machine, run that line from x64 PowerShell to install the x86_64 build under emulation, or use WSL2. The installer adds kin to your user PATH, so open a new PowerShell window in place of exec "$SHELL" -l before the next step.

For other installers or troubleshooting, see the full quickstart.

2. Initialize the repository

At the new prompt, clone a repository with Kin:

kin clone https://github.com/pallets/itsdangerous

Or initialize one you already have, replacing the path below:

cd /path/to/your/repository &&
kin init . &&
kin overview &&
kin status

Both end with the next command to run, a kin refs question about a function in that repository. They also connect Codex CLI and Grok CLI to it when you allowed that in setup, because those clients keep one entry that names a repository.

Windows PowerShell 5.1 has no &&, so on native Windows run the same commands one at a time and stop if one fails:

cd C:\path\to\your\repository
kin init .
kin overview
kin status

kin overview shows the entities Kin imported. kin status shows what was admitted and the working tree's state against it. kin graph status reports the daemon's live query graph and coverage. Uncommitted and untracked changes are not part of the imported Git history; kin init reports what it left out.

3. Ask a question you can check

Look for something you already know is in the code:

kin locate "<something you already know is in this repository>"

Replace ExactEntityName below with a symbol from the result:

kin refs ExactEntityName
kin trace ExactEntityName
kin impact ExactEntityName

refs returns recorded references, trace brings in nearby context, and impact explores potential effects through the graph. Check the results against the source.

kin setup already asked before connecting the AI coding tools it found, and connected them if you said yes. There is no separate linking step. After you install another tool, run kin setup again, and kin setup status shows what is connected. On Windows, connect AI tools inside WSL2.

kin runtime status --json checks the shared CLI and daemon installation, including their exact source identity. Normal commands refuse mismatched pairs; development builds use isolated state. See runtime installation.

Graph and lexical retrieval work without a model. Embeddings and reranking are optional, and their model providers default to none. Standard CLI and daemon builds include external HTTP model support. Built-in models require a build with optional local inference enabled.

To build an embedding index using an OpenAI-compatible endpoint you operate:

KIN_MODEL_EMBEDDINGS_PROVIDER=external \
KIN_MODEL_EMBEDDINGS_BASE_URL=http://localhost:1234/v1 \
KIN_MODEL_EMBEDDINGS_MODEL=your-served-embedding-model kin embed

Model endpoints receive the selected entity text needed for their stage. Model settings describe providers, model identities, and credential environment variable names.

Kin supports Claude Code, Codex, Cursor, Gemini, and other MCP clients. Use kin setup --intent editor for VS Code. Kin also includes kin agent run for local or hosted OpenAI-compatible model endpoints.

Client configuration · MCP tools · Built-in agent

Local imports, storage, and graph queries run on your machine. Installing Kin needs network access; the default graph retrieval path needs no model download. When an embedding provider is selected, queries report its vector coverage and any gaps.

Review a change

After running kin init on the Git branch you want to review, compare explicit commit SHAs against main:

kin review shadow "$(git rev-parse main)..$(git rev-parse HEAD)"

The report returns PASS, NEEDS ATTENTION, or WOULD BLOCK, with graph-derived impact and supporting evidence. It is advisory: it does not block a merge or change graph state. Authorship is declared, not independently verified.

Use Kin with or without Git

Kin has its own commits, branches, merges, diffs, and history, including in repositories with no Git underneath. Existing Git repositories can be imported, and supported workflows can export a new Git repository.

Native version-control walkthrough · Git interoperability and export limits

Beta limits

Coverage is incomplete. Supported languages are parsed into entities and relationships; other files remain available as content and history. An empty result does not prove there are no callers or dependencies. Keep using your compiler, tests, and review. See language support.

Compatibility varies. Release archives do not include the deprecated kin-vfs filesystem projection, so there is no transparent projection to rely on. Check platform notes for what each platform supports.

Back up Kin-only state. Commits, branches, reviews and specs you record with Kin live in the repository's Kin store, not in Git. Deleting the store and re-importing from Git does not recover them, so run kin backup create first. kin store prune deletes stores only when you pass --yes, for the same reason. Read the import, recovery, and upgrade notes.

Windows. Native Windows x86_64 support is early. Repository admission works: kin init imports a Git repository and publishes graph authority, and graph, lexical, and daemon-backed queries answer natively. The end-to-end install proof also runs agent setup on native Windows and gets graph-backed answers from the installed MCP server. Transparent filesystem projection is not shipped on Windows, and review workflows are not yet tested there, so WSL2 remains the recommended path for the full Kin experience.

Why I built Kin

I kept watching coding agents piece together parts of a codebase we'd already worked through. Then I'd do a version of that work myself to review their changes. I started wondering why more of that structural understanding wasn't part of the repository itself.

That's what I'm building with Kin.

Troy

Learn more and contribute

CLI reference · Architecture and detailed limits · Contributing · Issues · Security

Kin and its local stack, including kin-db, are Apache-2.0. KinLab is the separate proprietary hosted product. Public repository onboarding is still in development.

Available Tools

22 tools
find_referencesB
Read-onlyIdempotent

Callers, importers and references of one entity. Chains? trace_data_flow.

Inputs: query, entity_id, max_chars, answer_only, +1 more.

get_context_packC
Read-onlyIdempotent

Bodies, signatures and routes around entities or a question, in one bundle.

Inputs: depth, entities, question, entity_id, +4 more.

get_entity_sourceC
Read-onlyIdempotent

One entity's exact source, by id.

Inputs: entity_id (required).

graph_neighborhoodB
Read-onlyIdempotent

What one entity depends on and what depends on it. Paths: trace_data_flow.

Inputs: entity_id (required), depth, limit, direction, +1 more.

impact_analysisC
Read-onlyIdempotent

Every entity a change could affect, from entity_ids or change_ids.

Inputs: base, head, files, max_chars, +2 more.

kin_graph_statusC
Read-onlyIdempotent

Graph and embedding counts. last_settled_selected_graph marks a last settled reading.

Inputs: no named parameters.

kin_initA
Idempotent

Set a folder up as a Kin repository when it is not one.

Inputs: path.

kin_mutateB
Destructive

Commit targeted entity changes atomically, in one call.

Inputs: operations (required), summary.

kin_provenance_queryB
Read-onlyIdempotent

Who changed an entity, when, and whether it was approved.

Inputs: entity_id (required), limit.

kin_session_endB
Idempotent

Session plumbing for writes. Close an open session.

Inputs: session_id (required).

kin_session_execB
Destructive

Build, test or run this project with its toolchain. Session needs can_execute.

Inputs: argv (required), session_id (required), env, summary, +2 more.

kin_session_heartbeatB

Session plumbing for writes. Keep a session alive.

Inputs: session_id (required).

kin_session_startB

Session plumbing for writes. Open one; get a session_id.

Inputs: cwd (required), vendor (required), client_name (required), pid, +3 more.

kin_transaction_abortA
DestructiveIdempotent

Transaction plumbing for writes. Discard what is staged and close it.

Inputs: transaction_id (required), session_id.

kin_transaction_beginA

Transaction plumbing for writes. Open one; get a transaction_id.

Inputs: scope (required), session_id (required).

kin_transaction_commitB
DestructiveIdempotent

Transaction plumbing for writes. Publish everything staged, atomically.

Inputs: transaction_id (required), message, operations, session_id.

kin_transaction_stageC

Transaction plumbing for writes. Stage an update, create, delete or rename.

Inputs: operations (required), transaction_id (required), session_id.

lexical_lookupB
Read-onlyIdempotent

Exact literals, punctuation included, in stored graph fields. Lexical evidence only.

Inputs: kind, limit, cursor, literal, +1 more.

semantic_locateA
Read-onlyIdempotent

Find code by describing what it does. Know the exact name? semantic_search.

Inputs: limit, query, cursor, max_chars, +2 more.

trace_data_flowA
Read-onlyIdempotent

The ordered call chain out from one entity. Two endpoints? trace_path.

Inputs: focal (required), depth, target, compact, +5 more.

trace_pathA
Read-onlyIdempotent

How one entity reaches another, as hops. One endpoint? trace_data_flow.

Inputs: to (required), from (required), limit, to_file, +4 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.1/5.0

    Scored across 22 tools

    Disambiguation4/5

    Tools cover distinct retrieval modes (lexical, semantic by name vs. meaning, trace flow vs. path, references vs. neighborhood) and descriptions explicitly cross-reference to guide selection. Some overlap remains among graph query tools and between kin_mutate and transaction staging, but boundaries are mostly clear.

    Naming Consistency3/5

    All names are snake_case, but prefixes and patterns are mixed: kin_* appears on session/transaction/init/mutate/status/provenance tools, while query tools use bare verb_noun or noun_noun forms. The set is readable but not fully consistent.

    Tool Count3/5

    22 tools is borderline heavy for a code-graph server. The scope spans read queries, write/session/transaction plumbing, provenance, init, and execution, so most tools earn their place, but the count is on the high side.

    Completeness4/5

    The surface covers graph querying, entity mutation with transaction support, sessions, provenance, initialization, execution, and direct source retrieval. Minor gaps like bulk entity listing or file-level operations exist, but core lifecycle appears complete.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Knot is a semantic and structural codebase indexer designed for AI coding agents and developers navigating large projects. It combines vector search and graph traversal to find code by meaning, analyze impact via reverse dependencies, and explore file architectures.
      7
      11
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Provides a dependency graph of any local repository with tools for change impact, transitive dependents, health audits, and more, enabling AI coding agents to see structure and refactor safely.
      4
      MIT