Kivgraph

Kivgraph

[!NOTE] Kivgraph began as a research project. It explored whether a local semantic code graph could reduce the amount of context coding agents need to inspect.

The resulting tool remains useful for structured code navigation, cross-repository relationships, and impact analysis. However, the research did not demonstrate a reliable net token saving over long, end-to-end agent conversations. While some isolated queries can use substantially fewer tokens than grep plus file reads, those savings did not consistently translate into lower token usage across realistic long-running agent workflows.

The benchmark results and their limitations are preserved in benchmarks/, including the end-to-end agent benchmark.

Kivgraph is a local cross-repository code intelligence MCP server for AI coding agents. It builds a canonical semantic code graph across multiple registered repositories and answers questions about symbols, repository relationships, callers, dependencies and change impact.

Watch the Kivgraph UI demo video It indexes a corpus once and serves an immutable graph: the edges are resolved by go/types, the TypeScript checker and rust-analyzer, not by matching names. That is the difference from a search tool, and it is what makes an empty answer worth something — an empty reference list means nobody calls it, not that nothing was found, and grep cannot tell those apart.

Kivgraph is focused on semantic code relationships, not automatic discovery of every HTTP, gRPC, Kafka or database runtime flow between services.

Documentation

Read the Kivgraph user documentation for installation, MCP clients, code intelligence, repository relationships and workspace code graphs. The same pages are the source of landing/src/content/docs in this checkout, which is what a reader on a fork or without a network still has.

What each tool answers

Thirteen read-only tools, plus two consent-gated mutations (index_project and start_index_project) that a client has to authorize before either can register a repository or publish a generation.

Every row that names a symbol carries its repository, path, qualified name and line range, so it can be opened without a second call, and every tool accepts that triple in place of an opaque key.

Where it loses. A rare name in one small repository is cheaper with grep, and indexing a small file costs more than reading it. It wins on common names, on transitive impact, on consumers in another repository, and on proving an absence. Measured over 29 questions against a 37-repository corpus (benchmarks/graph-tools-comparison/results-all.json, commit 954b9eb, tokenizer o200k_base): 35,961 tokens for Kivgraph against 267,980 for grep plus reading, both exact on 28 of the 29, median 5.95x per question in Kivgraph's favour. grep is cheaper on 5 of those 29, all of them at full recall on both sides: T1_go_trivial asks for a name the corpus declares twice, and there grep costs 0.53x what Kivgraph does.

A second harness, benchmarks/mcp-token-cost, compares against the host's own tool output captured verbatim, but it runs on Kivgraph's own single repository of 13,222 symbols: 7.64x on the answers themselves and 1.60x over a whole session, against a 2.41x floor set by the source bodies both arms pay for.

Status

Released and in use. kivgraph version reports the published release; the backlog and the acceptance gate of every phase are in TASKS.md.

• Languages: Go, TypeScript, Rust, Python and Dart. Python uses the bundled AST worker in fallback mode; those inferred references are CANDIDATE, never EXACT. Exact Python mode uses the bundled Pyright LSP adapter with an installed Pyright/BasedPyright server. Dart uses the Dart Analysis Server supplied by the Dart or Flutter SDK.
• Semantic dependencies: Python and Dart imports can publish a package dependency when exactly one registered provider owns the requested package; symbol-level cross-repository edges require an explicit provider identity.
• Surface: thirteen read-only tools over STDIO, plus two consent-gated mutations (index_project and start_index_project). The contract is docs/protocol/mcp-surface-v3.md.
• Storage: LadybugDB is canonical; queries are served from an immutable HotSnapshot published atomically, never from the database.
• Platforms: linux/amd64, darwin/arm64 and windows/amd64.
• Viewer: kivgraph ui serves a read-only 3D view of the published graph.

Installation

Install the MCP with one script

The installer detects the platform, downloads the latest published MCP release for it, verifies both the release archive and the bundle checksums, and installs it without requiring Go or pnpm. The release contains the Go server, the pinned LadybugDB library, the TypeScript worker, the bundled Python AST worker, the pinned rust-analyzer, the grammar manifest and the web viewer, whose assets are 2.3 MB of the bundle. scripts/build-bundle.sh --mcp-only produces a bundle without the viewer for anyone who wants one. --slim goes further for anyone packaging an .mcpb: it leaves out the pinned rust-analyzer and every symbol a debugger would read, which is 46.3 MB packaged against 24.9 MB. It downloads nothing later, so that bundle reads Rust only where the machine already has an analyzer on its PATH.

Published bundles: Linux amd64, macOS arm64 and Windows amd64.

Runtime requirements: Bash on Linux and macOS or PowerShell 5.1 or later on Windows, Node.js 22 or later, Python 3.10 or later when indexing Python, and on the POSIX platforms curl, tar, and sha256sum or shasum. The bundle carries its own rust-analyzer; indexing Rust repositories additionally needs cargo on the PATH, and indexing Dart needs the Dart or Flutter SDK. On Windows the installer also installs the Visual C++ redistributable, without which kivgraph.exe does not start.

On macOS the binaries are not notarized. A release downloaded with curl is not quarantined and runs; a copy downloaded with a browser needs xattr -dr com.apple.quarantine. See docs/development/macos.md.

Install the latest release in one command. On Linux and macOS the same line covers both, because the installer reads uname and picks its own archive:

curl -fsSL https://kivgraph.dev/install.sh | bash

On Windows, where install.sh cannot run because there is no POSIX shell:

irm https://kivgraph.dev/install.ps1 | iex

install.ps1 is a second implementation of the same pre-extraction checks, and internal/release/install_parity_test.go fails when either script grows a check the other lacks. Piping it into Invoke-Expression turns its #Requires -Version 5.1 into a comment; download it to a file and run it as one to keep that guard.

From a checkout, either installer can be run directly:

./scripts/install.sh

To install a specific release instead of the latest one:

KIVGRAPH_VERSION=v0.9.10 ./scripts/install.sh

The script installs the bundle in ~/.local/opt/kivgraph and puts launchers in ~/.local/bin; on Windows it is %LOCALAPPDATA%\Programs\kivgraph and %LOCALAPPDATA%\Programs\kivgraph-bin. It never modifies a registered repository, creates an index, or replaces configuration files. To use a different location, set KIVGRAPH_INSTALL_ROOT and KIVGRAPH_BIN_DIR.

Add the launcher directory to the current shell and verify both runtimes:

export PATH="$HOME/.local/bin:$PATH"
kivgraph version
kivgraph-ts-worker <<'EOF'
hello
EOF

Check for a newer release or update the installed bundle:

kivgraph update --check
kivgraph update

Bundle replacement is atomic, preserves the configuration and graph state, verifies the release and bundle checksums, and replaces the installed bundle. The post-install runtime refresh may partially complete, fail, and make the command exit non-zero. It also restarts an installed supervised daemon and refreshes Kivgraph-managed user hooks, skills and MCP registrations. A stale supervisor returns an error and is not restarted. Missing, foreign and project-scoped integrations are left alone. Client-owned serve and ui processes still need a restart, or --stop, to use the new binary.

Development builds use a separate prerelease channel. Install one explicitly, then select that channel for later checks:

release=vX.Y.Z-dev.N
curl -fsSL \
  "https://github.com/Luqueee/kivgraph/releases/download/$release/install.sh" |
  KIVGRAPH_VERSION="$release" bash
kivgraph update --channel dev

For a prerelease binary, omitting --channel already follows dev; stable installations continue to follow the stable channel. KIVGRAPH_UPDATE_CHANNEL can be used instead of the flag, including for the interactive update notice.

To remove the installed bundle and launchers without deleting configuration, repository registrations or graph state:

uninstall_url=https://github.com/Luqueee/kivgraph/releases/latest/download/uninstall.sh
curl -fsSL "$uninstall_url" -o /tmp/kivgraph-uninstall.sh
bash /tmp/kivgraph-uninstall.sh

Use bash /tmp/kivgraph-uninstall.sh --yes for a non-interactive removal. Windows users can run the corresponding uninstall.ps1 with PowerShell.

When kivgraph is invoked without a command from an interactive terminal, it checks for a newer release with an 800 ms timeout and a 24-hour cache in the platform cache directory ($XDG_CACHE_HOME on Linux and $HOME/Library/Caches on macOS), under kivgraph/update-check.json. The optional check never blocks the command when the network is unavailable.

Interactive command output uses semantic ANSI colors when the destination is a terminal. Set NO_COLOR or redirect output to keep it plain.

Configure Kivgraph and its coding agents

After an interactive release installation, the installer asks whether it should run this same guided setup. Run it later, or run it directly from a checkout:

kivgraph configure

configure opens one selector for the coding agents detected on this machine and installs each user-scoped surface that the selected agent supports. It also adds user-level Kivgraph instructions for every selected agent. It initializes the empty Kivgraph configuration when needed, but it does not register a repository or run an index. Repeat --target TARGET for scripted setup; omit it to open the selector.

Use ↑/↓ (or j/k) to move, space to toggle an agent, a to select all, n to select none, Enter to confirm, and q or Esc to cancel. The daemon is offered once after the selection; --daemon requires it without asking and --stdio keeps one serve process per client. --dry-run previews every surface without writing.

Supported MCP targets are claude-code, claude-desktop, codex, opencode, and oh-my-pi. Supported skill targets are claude-code, codex, opencode, and oh-my-pi; Claude Desktop has no local skill target. The standalone commands use user by default; use --scope project for project-local configuration. Use --dry-run to inspect a plan without writing. Existing incompatible entries stop with an error; --force is required to replace or remove one. Existing files are written atomically with mode 0600 and receive *.kivgraph.bak backup before replacement or removal.

The pre-tool-use gate supports claude-code, claude-desktop, codex, opencode, and oh-my-pi. Oh My Pi receives a native extension under ~/.omp/agent/extensions/ for user scope or .omp/extensions/ for project scope. The gate is fail-open when its graph query cannot be answered. Searches wrapped as rtk rg ... or rtk proxy rg ... are classified by their inner command, while RTK's own commands are left alone.

Add Kivgraph to user agent instructions

Add the Kivgraph navigation rules to the user context loaded by a coding agent:

kivgraph instructions install
# The selector can install one or more coding-agent destinations.
kivgraph instructions install --agent codex
kivgraph instructions install --agent claude
kivgraph instructions install --agent omp

With no --agent or --file, the interactive selector lets you choose one or more coding agents. Each installation owns a KIVGRAPH.md prompt beside the client configuration. Codex (~/.codex/AGENTS.md), Claude Code/Desktop (~/.claude/CLAUDE.md), and Oh My Pi (~/.omp/agent/AGENTS.md) receive only a small managed absolute-path reference to KIVGRAPH.md. OpenCode instead adds its canonical path to ~/.config/opencode/opencode.json's native instructions list; it does not modify OpenCode's AGENTS.md. The selector deduplicates shared destinations. Existing instructions are preserved, the managed reference is idempotent, and --dry-run previews the change. An edited Kivgraph prompt or reference requires --force to replace. --file is retained for compatibility and selects every matching global client; prefer --agent for new automation. The command never changes repository instructions. Use configure when you want these instructions and the compatible client integrations in one flow; the individual commands remain available for explicit changes.

When set, CODEX_HOME replaces ~/.codex and PI_CODING_AGENT_DIR replaces ~/.omp/agent, matching the configuration roots those clients use.

Inspect or remove a registration explicitly:

kivgraph mcp status --target claude-code --scope user
kivgraph mcp remove --target claude-code --scope user
kivgraph skill status --target claude-code --scope user
kivgraph skill remove --target claude-code --scope user
kivgraph hook status --target claude-code --scope user
kivgraph hook remove --target claude-code --scope user

Initialize and publish a graph before starting the MCP server:

kivgraph init \
  --repository project=/absolute/path/to/project \
  --languages go,typescript,rust
kivgraph doctor
kivgraph index --full

init writes a self-contained configuration: with --config pointing elsewhere, its state, cache and registry hang off that directory, so a throwaway index never touches the real one. index --full republishes atomically — a failure at any stage leaves the previous generation serving. A server already running follows the new generation on its own.

When you are inside one project, kivgraph index detects its supported languages, creates or reuses .kivgraph/, registers the current project as project in its local registry, and runs the same full rebuild. With neither --config nor --repositories, it does not alter the shared user registry; those overrides intentionally select the configuration and registry to update. Use kivgraph index --full when you want the explicit registered-repositories workflow; both forms preserve the full-indexing contract. The command does not install language toolchains implicitly; kivgraph doctor reports any prerequisite that is missing on the host. Optional analyzers can be managed explicitly by Kivgraph:

kivgraph toolchain status
kivgraph toolchain install pyright
kivgraph index --full

toolchain install pyright pins and installs Pyright under Kivgraph's state, then activates exact Python analysis in the selected configuration. For a project-local configuration, pass --config .kivgraph/config.yaml. Removing it requires an explicit confirmation and restores the bundled Python fallback when the selected configuration uses the managed analyzer:

kivgraph toolchain remove pyright --yes

The first install requires npm and network access. Later status checks and reusing an installed version work offline.

The command family is intentionally language-agnostic; more managed analyzers can use it without making index mutate the host or a repository.

Day to day:

kivgraph graph status      # what is published, and whether a tree has moved
kivgraph doctor            # toolchains, storage, and the type-checking ceiling
kivgraph ui                # read-only 3D viewer, default 0.0.0.0:7777
kivgraph logs --follow     # aligned history of what it indexed, served and queried
kivgraph tool-stats        # per-tool cost, calls, and failures
kivgraph stop              # terminate this user's serve and ui, never an index
kivgraph clean --keep-active

kivgraph ui binds a non-loopback address by default, because the graph is indexed where the repositories are and looked at from elsewhere; there is no authentication, so it logs exactly what it exposes and --addr restricts it.

logs and tool-stats read an append-only record in the state directory rather than asking a server, which is why they can answer at all: the per-tool counters a serve keeps are minted when it starts and gone when it stops. Reading the file also makes the answer span every server that ever ran.

logs renders a fixed-column table. Tool rows include a bounded, allow-listed query summary: find_by_intent records its exact intent and any keywords, while opaque cursors, consent flags, stable keys, and absolute project paths stay out of the record. --json preserves the individual append-only records. SYMBOL_NOT_FOUND renders as the neutral NOT_FOUND status with zero results, not as an operational failure.

Configure any MCP client to start the server over STDIO:

{
  "mcpServers": {
    "kivgraph": {
      "command": "/home/user/.local/bin/kivgraph",
      "args": [
        "serve",
        "--config",
        "/home/user/.config/kivgraph/config.yaml"
      ]
    }
  }
}

kivgraph serve starts before a graph exists: with no published generation it completes the handshake, publishes no query tool and puts the rebuild command in instructions. A client launches the process itself, so exiting would read as a crash. It writes MCP framing exclusively to stdout and logs to stderr.

Requirements

• Go 1.26 or later to build from source. The indexer type-checks with the go/types linked into the binary, so it can only read repositories and dependencies written for its own language version or older; kivgraph doctor reports that ceiling.
• Indexing Rust needs cargo and rust-analyzer. The release bundle carries the analyzer; it does not carry a Rust toolchain.
• Indexing TypeScript needs Node.js 22 or later for the worker.
• Indexing Python needs Python 3.10 or later for the bundled worker. It is a syntax-aware fallback and reports dynamic or unresolved names explicitly; exact mode additionally requires a Pyright-compatible language server.
• Indexing Dart needs the dart executable; a Flutter installation supplies it. The loader uses the Analysis Server protocol and does not modify the Flutter project.

What the graph carries, and what it refuses to

An edge is EXACT only with sufficient evidence and the right provenance. It is never created from a name, a path, an alias or a single candidate, and a reference that cannot be resolved is published as UNRESOLVED with its reason, repository and language rather than dropped. graph_status reports both, broken down.

That is why some answers are absences rather than edges. With the Rust standard library indexed, impl Add for u32 is generated by a macro and exists in no source range, so every use of it is declared PROVIDER_DEFINITION_NOT_INDEXED once per symbol instead of becoming an edge nobody could open.

The providers Kivgraph derives from the machine — today the Rust standard library, named rust:1.96.1 after the toolchain — are withheld from read results by default: one toolchain is around twenty thousand symbols, and a search for Clone would answer with core. include_derived asks for them, and graph_status breaks out what they contribute so the totals stay readable.

Development

make build
make test
make semantic-coverage
make test-ladybug

make test-ladybug is the only supported way to run the tag that links the pinned native library. Contributing conventions are in AGENTS.md, which CLAUDE.md links to.

make semantic-coverage is the release gate for Go, TypeScript, Python and Dart. It validates the machine-readable matrix in testdata/semantic-coverage/manifest.json, runs the exact TypeScript, Go and Dart suites, and requires a Pyright-compatible language server for the exact Python suite. A language is not considered complete when a capability has a fixture but no executable regression test.

Storage and graph benchmarks

The LadybugDB qualification, the synthetic corpus generator, the load and query benchmarks, and the doctor, rebuild, rollback and snapshot commands are documented in docs/development/storage-benchmarks.md. It concludes with ACCEPT_LADYBUGDB_WITH_LIMITS.

The public site

landing/ carries the landing page and the user documentation. It ships in no release bundle, is verified with make landing-check and make landing-build, and is served on port 6767. What it publishes, how the MCP reference was captured, and what is still open are recorded in docs/development/landing-site.md.

Structure

cmd/kivgraph/   Main executable.
internal/        Kivgraph internal packages.
ts-worker/        TypeScript worker.
web/              Graph viewer served by `kivgraph ui`.
landing/          Landing page and documentation site (not part of any release).
testdata/         Test fixtures and corpora.
benchmarks/       Benchmark results.
docs/             Documentation and ADRs.
scripts/          Auxiliary automation.

License

Kivgraph is distributed under the Apache License 2.0.

Third-party licenses

Notices and licenses for dependencies distributed with Kivgraph are recorded in THIRD_PARTY_NOTICES.md. The list is updated whenever a dependency is added to the distributable product.