What is senkani?

Senkani is a native macOS binary that sits between your AI coding agent (Claude Code, Cursor, Copilot, any MCP-compatible tool) and your filesystem. Every tool call — read, grep, shell, build, search — is intercepted by a hook relay, answered by a local symbol index or filter pipeline where possible, and forwarded with only the output that matters.

The one-sentence version

Senkani is a trust boundary. Your agent on one side, your filesystem on the other; the boundary compresses, redacts, caches, and learns.

The category

Senkani is a new category of developer tool. It is not a new coding agent, not an IDE plugin, not a cloud service, not a framework your code imports. It is a local compression-and-context layer that makes the agent you already use cheaper, faster, and safer — by intercepting the tool calls the agent makes on your behalf.

Who it's for

• Developers running coding agents on real codebases. If your Claude Code session is eating 40k+ tokens on a medium refactor, the compression layer pays for itself on the first prompt.
• Operators who care about trust boundaries. Secret redaction, SSRF hardening, prompt-injection scanning, and socket-auth handshakes all ship on by default.
• People tired of configuring LLM stacks. senkani init, one command, idempotent, zero config files required.

How the pieces connect

Senkani has three layers, and all three ship in the same binary:

1. MCP tools — 19 tools the agent can call directly. Compressed reads, filtered shells, symbol lookups, sandboxed parse, local embed/vision.
2. Hook relay — intercepts the agent's built-in Read/Bash/Grep and routes to senkani equivalents or denies with cached results.
3. Smart denials — re-read suppression, command replay, trivial routing, search upgrade. The filter decides a tool call doesn't need to run at all.

What it replaces (or augments)

• Read → senkani_read (compressed, cached, redacted) or senkani_fetch (symbol-only).
• Grep → senkani_search (local tree-sitter index; 99% fewer tokens than text grep).
• Bash → senkani_exec (24+ command-specific filter rules; mutating commands pass through unchanged).
• Per-repo briefs → senkani_bundle (budget-bounded Markdown or JSON).
• Cloud embed / vision API calls → senkani_embed + senkani_vision on Apple Silicon.

Next

• Install senkani — clone, build, init, launch.
• Read about the three-layer stack.
• Browse the MCP tool reference.