Indago Retrieval · on device

One local server. Every MCP-aware client.

Indago hosts an MCP server on your machine. Connect Claude Code, Claude Desktop, Codex, Cursor, or any MCP-aware client with one command. Tools resolve to the local daemon — no SaaS hop.

Six first-class wirings.

Claude Code

terminal$ indago connect claude-code → MCP server registered (user scope) → conversation sync hooks installed

Adds SessionStart:compact hook so the full transcript is indexed before truncation.

Claude Desktop

~/Library/.../claude_desktop_config.json{ "mcpServers": { "indago": { "command": "indago", "args": ["mcp", "serve"] } } }

Restart Claude Desktop; Indago tools appear in the tool tray.

Codex

~/.codex/config.toml[mcp_servers.indago] command = "indago" args = ["mcp", "serve"]

Available as the indago namespace inside Codex.

Cursor

.cursor/mcp.json{ "mcpServers": { "indago": { "command": "indago", "args": ["mcp", "serve"] } } }

Project-scoped wiring; one entry per repo.

Ollama · llama.cpp · MLX

For local LLM workflows, Indago's HTTP API is the integration layer: have your local model call POST /v1/find / POST /v1/read / POST /v1/report. JSON in, JSON out. No MCP required.

Generic MCP client

Anything that speaks MCP can attach. The transport defaults to stdio; indago mcp serve --transport sse opens a local SSE endpoint for browser-class clients.

Tiered. Handshake stays small.

The core MCP handshake ships under 3K tokens. Full per-tool prose is loaded on demand via indago_help(tool="indago_find"). That keeps your LLM's context budget for the work, not the schema.

Past sessions, searchable.

Every MCP-aware client writes session transcripts to a known path. Indago indexes those transcripts. indago_find(mode="conversation") returns past Q&A across every client. The Claude Code compaction hook captures the full transcript before truncation — nothing is lost in /compact.

Who called what, and what it cost.

indago usage --hours 24 returns a per-agent ledger of tool calls, work-units, p50/p95 latency, and cache hit rate. Every tool call records its caller (client name, session id) for cost attribution and policy enforcement.

No outbound network.

In sovereign mode, the MCP server still works end-to-end; only the public-API indago_learn calls are blocked. Pre-stage your source bundles on the network; pre-stage your LLM. Offline licence activation. No telemetry.