Semantic infrastructure for code change.
Kai understands what code means — functions, dependencies, behavior impact — not just which lines changed. This semantic graph powers precise CI, context-aware IDEs, and verifiable AI coding agents.
atlas.kaicontext.com · Docs · Slack
# curl
curl -sSL https://get.kaicontext.com | sh
# Homebrew
brew install kaicontext/kai/kaikai init # One-shot setup: graph, MCP, account, remote
kai code # Launch the interactive coding experience
kai capture -m "Initial"# Snapshot your code with a message
kai push # Push to atlas.kaicontext.com
kai diff # Semantic change impactkai code is the single front door to Kai's interactive coding agent. On
first use it downloads its backing component to ~/.kai/bin automatically,
then launches it — no extra setup, on any install path (curl, Homebrew, or
manual). Every flag after code is forwarded to the experience unchanged.
kai init runs a single, low-friction setup flow:
- Creates the semantic graph in
.kai/and builds the first capture. - Imports git history automatically (for repos with ≤1000 commits) — no prompt.
- Installs git hooks for auto-capture on commit and auto-push on
git push. - Installs the Kai MCP server for any detected AI coding tool (Claude Code,
Codex). If the
kaiMCP is already registered, it's left alone; otherwise it's added automatically — no prompt. - Signs you up for atlas.kaicontext.com in-place: asks for your email, emails you a login link, you paste the token back, and Kai creates a personal org named after you plus a repo for the current directory. Your first push goes out at the end of init.
Every step runs by default. Press Ctrl+C to skip the account step if you'd
rather stay local; everything else is automatic.
For full command reference, see docs/cli-reference.md.
Kai ships an MCP server that gives AI coding assistants access to call graphs, dependency maps, impact analysis, and test coverage.
# Claude Code
claude mcp add kai -- kai mcp serve
# Or without installing kai (npx downloads it automatically)
claude mcp add kai -- npx -y kai-mcpNo setup required — the server lazily initializes the semantic graph on first use.
12 tools: kai_status, kai_symbols, kai_files, kai_diff, kai_impact, kai_callers, kai_callees, kai_context, kai_dependencies, kai_dependents, kai_tests, kai_refresh.
See docs/mcp.md for Cursor setup, tool reference, and troubleshooting.
Kai's agentic coding experience (kai code) uses kailab as the default LLM provider — you sign in with kai auth login and the server holds the upstream key. If you'd rather use your own API key (or a local model), set KAI_PROVIDER and the matching key.
Kailab credentials are still used for sync, shared safety gates, and team features when present, regardless of which LLM provider you pick.
export KAI_PROVIDER=anthropic
export ANTHROPIC_API_KEY=sk-ant-...
kai codeSame wire shape and prompt-caching behavior as kailab. Costs and rate limits are billed to your Anthropic account.
export KAI_PROVIDER=openai
export OPENAI_API_KEY=sk-...
export KAI_OPENAI_MODEL=gpt-4o
kai codeWorks with OpenAI, Together, Groq, OpenRouter, and any endpoint that speaks the chat.completions protocol. No prompt caching — every turn re-sends the conversation at full cost. Expect 5–10× the per-session cost of an equivalent Anthropic run.
# `openai-compatible` (or `local`) is an alias for `openai` —# more honest about the fact that the endpoint isn't actually# OpenAI. Same wire protocol, different vendor.export KAI_PROVIDER=openai-compatible
export KAI_OPENAI_BASE_URL=http://localhost:11434/v1
export KAI_OPENAI_MODEL=llama3.1:70b
kai codeOPENAI_API_KEY is optional for endpoints that don't require auth. Tool-use quality varies sharply by model — pick one that natively supports the OpenAI tool-call protocol (Llama 3.1 70B+, Qwen 2.5 32B+, Hermes-tuned models). Smaller models often emit tool calls as text, which Kai can extract with limited fidelity (set KAI_OPENAI_TOOL_FORMAT=hermes|llama3 to opt into a specific text format).
| kailab | Anthropic direct | OpenAI / compat | Local | |
|---|---|---|---|---|
| Setup | kai auth login | API key | API key | Local server |
| Prompt caching | Yes (server) | Yes | No | No |
| Per-session cost | Lowest | Low | ~5–10× higher | Free (compute) |
| Tool-use fidelity | High | High | High | Varies by model |
| Sync + team gates | Yes | Only if logged into kailab | Only if logged into kailab | Only if logged into kailab |
| Variable | Purpose |
|---|---|
KAI_PROVIDER | kailab (default), anthropic, or openai. Aliases: openai-compatible / oai-compat / local for openai; anthropic-direct / claude for anthropic. |
ANTHROPIC_API_KEY | Required for anthropic |
KAI_ANTHROPIC_BASE_URL | Override Anthropic endpoint (rare) |
KAI_ANTHROPIC_MODEL | Override default model on Anthropic |
OPENAI_API_KEY | Used by openai (optional for local) |
KAI_OPENAI_BASE_URL | OpenAI-compatible endpoint URL |
KAI_OPENAI_MODEL | Override default model on OpenAI |
KAI_OPENAI_TOOL_FORMAT | raw (default), hermes, or llama3 |
KAI_MAX_SESSION_COST_USD | Pause and prompt before exceeding this cap |
Run kai auth status to see which provider is active and whether kailab sync is connected.
→ Full reference, troubleshooting, and per-provider model recommendations: docs.kaicontext.com/bring-your-own-model
Kai reviews are anchored to semantic changesets, not line diffs.
kai review open --title "Add auth middleware"# Create a review
kai push # Push to atlas.kaicontext.com
kai fetch --review abc123 # Sync comments from web
kai review comments abc123 # View inline comments locallyOn the web, reviews show semantic diffs (what functions changed, not just which lines), inline commenting, and one-click merge that updates snap.main.
Kai CI runs workflows defined in .kailab/workflows/ with semantic checkout, parallel jobs, and 30-minute default timeouts.
kai ci runs # List CI runs
kai ci logs 42 # View logs for run #42
kai ci cancel 42 # Cancel a run
kai capture -m "Fix bug"# Message shows as CI run headlineEmail notifications on pipeline completion are sent to the snapshot author via Postmark.
For every capture, Kai constructs a queryable semantic model:
| Layer | What It Captures |
|---|---|
| Functions & methods | Signatures, bodies, call graphs |
| Dependencies | Module relationships, imports, data flow |
| Behavior changes | What actually changed in meaning, not just text |
| Test coverage | Which tests cover which source files (static + transitive) |
This graph is immutable, content-addressed, and designed for machine reasoning.
Kai determines which tests actually need to run based on behavioral impact, not file diffs. Result: 80% CI time reduction for early users.
12 MCP tools give AI assistants structured access to your codebase's dependency graph, call graph, impact analysis, and test coverage.
Semantic diff shows what changed (function added, condition modified, API changed) instead of raw line diffs. Inline comments anchored to symbols, not lines.
Agent proposes edit → Kai validates impact → agent executes with proof, not generation with hope.
Kai uses tree-sitter for parsing and builds semantic graphs for 9 languages. Each language is validated with end-to-end tests against real open-source projects:
| Language | Symbols | Callers/Callees | Dependencies | Tests | Impact | E2E Fixtures |
|---|---|---|---|---|---|---|
| Go | functions, methods, interfaces, structs | cross-package via import aliases | Go package resolution | same-dir _test.go + transitive | via dependency graph | chi, lo |
| Rust | functions, structs, enums, traits, impl methods, macros | cross-file via exports | crate::, super::, self::, mod, wildcards | filename pattern matching for integration tests | via dependency graph | just, miniserve |
| TypeScript/JavaScript | functions, classes, methods, variables | via imports | relative + workspace resolution | *.test.ts, __tests__/ patterns | via dependency graph | — |
| Python | functions, classes, methods | via imports | dotted module resolution | test_*.py, *_test.py patterns | via dependency graph | — |
| Ruby | methods, classes, modules | via exports | require, require_relative, Zeitwerk autoload | *_spec.rb, *_test.rb patterns | via dependency graph | — |
| SQL | — | — | — | — | — | — |
| PHP | functions, classes, methods | — | — | — | — | — |
| C# | functions, classes, methods | — | — | — | — | — |
92 E2E tests across Go, Rust, and push completeness, all passing.
The Kai CLI is open source under Apache 2.0. The core engine (kai-core) ships
as a prebuilt module statically linked into the kai binary you download; the
server components are distributed separately. Kai Cloud is the hosted version —
managed infrastructure on top of the same engine.
kai capture → local semantic graph (SQLite)
kai push → atlas.kaicontext.com (Postgres + GCS)
├── File viewer with search, language breakdown
├── CI with SSE live updates
├── Code reviews with semantic diffs
└── Email notifications (Postmark)
See LICENSE.