Integrations overview

Sigil’s wedge is one memory layer underneath everything. Setup detects every coding agent on your machine and wires each one — hooks for Claude Code, MCP registration plus client-native steering rules for the rest.

Integration depth

Agent	How	Auto-wired
Claude Code	4 hooks · MCP server · @~/.sigil/CLAUDE.md hot-context import	✓
Codex CLI	~/.codex/config.toml MCP entry · ~/.codex/AGENTS.md steering	✓
Cursor	~/.cursor/mcp.json MCP entry	✓
Kiro	~/.kiro/settings/mcp.json · ~/.kiro/steering/sigil.md	✓
Continue	MCP server	one line
Cline	MCP server	one line
Windsurf	MCP server	one line
Claude Desktop / any MCP client	MCP server	one line
Custom agents, CI, scripts	sigil CLI, or MCP, or HTTP	—

sigil connect --all      # (re)wire every detected client
sigil doctor --deep      # prove each integration actually round-trips

Two interfaces

The CLI is universal

If your agent has a Bash tool, a terminal, or any way to shell out, it can use Sigil with no client-specific integration at all:

sigil search "JWT auth setup"
sigil remember --bg "we use jose, not jsonwebtoken"
sigil ingest ./DESIGN.md

That’s how Claude Code (via its Bash tool), Codex CLI, terminal agents, and CI scripts use Sigil today. sigil is discovered on PATH the same way git and node are.

MCP for clients that prefer structure

For Cursor, Continue, Cline, Windsurf, Claude Desktop, Kiro, and any spec-compliant client, the same memory is exposed as a 13-tool MCP server.

sigil register --print

{
  "mcpServers": {
    "sigil": {
      "command": "/Users/you/.sigil/bin/sigil-mcp",
      "args": []
    }
  }
}

Drop the inner sigil object into your client’s MCP config.

Why it points at a shim

The registration targets a stable launcher in ~/.sigil/bin/, never a versioned package path. That’s what keeps every integration working across Node version switches, reinstalls, and updates — the path agents were told about doesn’t move.

Or register the HTTP transport, whose URL never changes at all:

sigil register --http

Capture differs by client

This is the one asymmetry worth understanding.

	Claude Code	Everything else
Recall	Automatic, before every prompt	The agent calls search, or you run sigil search
Session priming	Automatic via CLAUDE.md import	The agent calls prime
Capture	Automatic — Stop and PostToolUse hooks	The agent calls remember, or you run sigil remember

The memory engine is identical. What Claude Code has that others don’t is a hook layer — extension points that let Sigil act without the model deciding to.

For every other client, the system prompt does that work. Setup writes a client-native steering file for Cursor, Codex CLI, and Kiro. For the rest, add this to the client’s rules:

You have access to a persistent memory system called Sigil. Call prime at the start of every session. Before answering questions about projects, preferences, or past decisions, call search. When the user shares a preference, decision, or constraint, call remember.

When something stops working

The single most common failure is a stale path after upgrading Node, switching between nvm and fnm, or reinstalling:

sigil connect

Re-pins the shims and re-syncs every client’s generated config. Touches no database, no provider, no API key.

sigil doctor --deep    # if that didn't do it

See Troubleshooting.

Removing an integration

sigil uninstall [--dry-run]

Walks every detected client and removes Sigil’s MCP entry, steering file, and hook entries — preserving anything else in those files. Your data is untouched.