Skip to main content

Documentation Index

Fetch the complete documentation index at: https://kasava.dev/docs/llms.txt

Use this file to discover all available pages before exploring further.

The Kasava web app is where you create plans, generate documents, manage settings, and view analytics dashboards. The Claude Code plugin gives you those plans plus product intelligence while you code — without context-switching back to the web. The plugin bundles MCP tools, 6 auto-trigger Agent Skills, 8 slash commands, and 4 subagents into a single installable package.

​
Installation

​
From the official marketplace

Anthropic auto-registers claude-plugins-official. Once Kasava is listed there, this is the simplest path: 
/plugin install kasava@claude-plugins-official

​
From our marketplace (today)

/plugin marketplace add kasava-dev/claude-code-plugin
/plugin install kasava@kasava-dev

​
Test locally

git clone https://github.com/kasava-dev/claude-code-plugin
claude --plugin-dir ./claude-code-plugin

​
Setup

1

Create an API key

Go to Settings → API Keys in Kasava and click Create API Key. Copy the key immediately — it’s only shown once. Format: kasava_live_… or kasava_test_….
2

Enter the key in Claude Code

When Claude Code prompts for your API key, paste it. The key is stored securely in your system keychain.Already running the Kasava CLI? The same API key works in both. Just paste it again, or eval "$(kasava env)" in your shell to share it via KASAVA_TOKEN.
3

Optional: install the CLI alongside

The plugin auto-discovers a local .kasava/ artifact if you’ve run kasava install + kasava sync in the repo. That gives the plugin zero-token-cost access to plans, graph, and symbols.
npm i -g @kasava/cli
cd /path/to/your/repo
kasava install   # binds the repo to a product
kasava sync      # pulls plans, graph, symbols, work items
4

Start coding

Claude automatically uses Kasava tools and skills when relevant. You can also invoke slash commands explicitly with /kasava:<command>.

​
What’s Included

​
MCP connection (8 tools)

The plugin connects to Kasava’s MCP server at api.kasava.dev/v1/mcp/sse, exposing:
ToolPurpose
code-toolSymbol search, definitions, usages, dependencies, semantic code search
commit-toolCommit history with semantic analysis, hotspots, changelog
search-toolUnified search across code/issues/plans/docs/competitors
document-search-toolCustomer signals + plans + product docs
competitor-toolLive competitor intel and content scrapings
entity-analysis-toolGraph traversal of features, layers, symbols
spec-export-toolPlan spec in agent-prompt format
record-decision-toolWrite decision against the bound product
Plus 3 Mastra agents (ask_codeIntelligenceAgent, ask_productInsightsAgent, ask_architectureInvestigationAgent) and 2 workflows (run_bugAnalysisWorkflow, run_deepResearchWorkflow). The toolset is intentionally small (cut from 23 in May 2026 — see ADR-0015). Cacheable data — product graph, work breakdowns, ownership, rationale — lives in your local .kasava/ artifact instead, refreshed by kasava sync. The plugin’s skills prefer the artifact over MCP for that data, then fall through to MCP for live queries.

​
6 auto-trigger skills

These activate when Claude detects matching user intent.
SkillTriggers when you ask…
kasava-contextAnything in a .kasava/-bound repo (paths-scoped)
kasava-product”what’s the architecture / features / tech stack”
kasava-code”where is X defined”, “what depends on Y”
kasava-bugPasting a stack trace, “debug this”, “investigate this bug”
kasava-plan”implement the plan for X”, “what does the spec say”
kasava-signals”what are users saying”, “any bug reports”

​
8 slash commands

User-invoked only — these don’t auto-trigger so you control when they run.
CommandEffectCost
/kasava:installBind this repo to a Kasava productFree
/kasava:syncRefresh .kasava/ artifactFree
/kasava:plan <name>Print the agent-spec for a planFree
/kasava:ask "<q>"Offline keyword search across .kasava/Free
/kasava:ask-ai "<q>"Stream AI-synthesized answer from the Kasava agentBillable
/kasava:stalenessList documents whose code references have changedFree
/kasava:decision "<text>"Record a decision on the bound productFree
/kasava:statusShow binding info, last sync, next-step suggestionsFree
The two billable variants (/kasava:ask-ai and kasava sync --ai) are deliberately not in the allowlist — Claude Code prompts every invocation so credits never get spent without your click.

​
4 subagents

Delegate read-heavy investigations instead of running them in the main thread. They run in isolated context and return a focused summary.
SubagentWhen to delegateModel
kasava-developerStarting a coding task; want auto-loaded context + project memoryinherit
kasava-bug-triagerBug or stack trace investigationhaiku
kasava-spec-loaderStarting work from a plan name/sluginherit
kasava-impact-assessorBefore non-trivial edits — walks dependents, owners, related planshaiku (read-only, no Write/Edit)
Invoke with @<agent-name>: 
@kasava-bug-triager investigate this stack trace: <paste>
@kasava-spec-loader load the plan "improve-onboarding"
@kasava-impact-assessor what would change if I modify processPayment

​
UserPromptSubmit hook

The plugin installs a hook at ${CLAUDE_PLUGIN_ROOT}/scripts/inject-context.sh that fires on every prompt. When you’re inside a .kasava/-bound repo, it injects:
  • A reminder to read .kasava/PRODUCT_MAP.md, plans/*.md, and the local artifact first
  • The CLI commands available (kasava plan, kasava ask, kasava ownership, etc.)
  • A “View in Kasava → app.kasava.dev/products/…” deep link
  • A staleness hint if .kasava/config.json is more than 7 days old
Outside a bound repo it exits silently. Always exits 0 (fail-open).

​
Usage Examples

Load product context before coding: 
What's the architecture of our product?
The kasava-context and kasava-product skills activate, read .kasava/PRODUCT_MAP.md and graph.json, and surface feature areas, architecture layers, and tech stack. Search code with the symbol graph: 
Find where AuthService.validateToken is defined and what depends on it
The kasava-code skill activates, calls code-tool action search_symbols then find_dependencies, and returns the definition with its dependents. Investigate a bug (delegated): 
@kasava-bug-triager investigate why /api/billing returns 403 for some users
The subagent runs the 3-way investigation (code + commits + customer signals) in isolation and returns a one-page triage summary with hypothesis, recent risky commits, and customer impact. Implement from a plan: 
@kasava-spec-loader load the plan for payment-retry-logic
The subagent pulls the spec from .kasava/plans/, resolves named symbols to current implementations, and returns a starter brief with acceptance criteria and a suggested first step. Record a decision: 
/kasava:decision "Use Drizzle for ORM" --type technology --rationale "Better PG support than Prisma"
Direct shell-out to the CLI’s kasava decisions add. Returns the new decision ID and a deep link to the product timeline.

​
Troubleshooting

Run /plugin validate . from the plugin directory. Confirm Claude Code is on a recent version.
Verify your API key at Settings → API Keys. Test connectivity with:
curl -H "Authorization: Bearer $KASAVA_TOKEN" https://api.kasava.dev/v1/mcp/sse
Try invoking directly: /kasava:kasava-product. If that works, the skill description may not match your phrasing — try rephrasing or just use the slash command. Skill descriptions are tuned for natural phrasing like “what’s the architecture” rather than abstract terms like “structural overview”.
Ensure you have at least one indexed repository in Kasava. Check the repository’s index status in the Kasava dashboard at app.kasava.dev.
The plugin works without the local artifact (it falls back to MCP for everything), but you’ll get more value with it. Run kasava install then kasava sync in the repo root. See the CLI guide for setup.
The plugin’s permission allowlist deliberately excludes kasava sync --ai and kasava ask --ai. Claude Code should always prompt before running them. If it doesn’t, your ~/.claude/settings.json may have an overly broad allowlist — check for a Bash(kasava *) rule that bypasses the per-command exact-match.