SuperLocalMemory
Local, governed long-term memory shared by coding agents, with recall that says when it has no answer
Local project memory, rule-pack routing and worktree cache tools for coding agents
By Permanu TypeScript Local · stdio Knowledge and memoryDeveloper tools
40 popularity score ★ 19 ↓ 311/wk
Opens the app, which asks before it adds the server.
claude mcp add superskill -- npx -y --prefer-online superskill@latest{
"mcpServers": {
"superskill": {
"command": "npx",
"args": [
"-y",
"--prefer-online",
"superskill@latest"
]
}
}
}{
"mcpServers": {
"superskill": {
"command": "npx",
"args": [
"-y",
"--prefer-online",
"superskill@latest"
]
}
}
}{
"servers": {
"superskill": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"--prefer-online",
"superskill@latest"
]
}
}
}[mcp_servers.superskill]
command = "npx"
args = ["-y", "--prefer-online", "superskill@latest"]gemini mcp add --scope user superskill npx -- -y --prefer-online superskill@latestVAULT_PATH |
Optional: Folder of the markdown vault; must be under your home directory. Default ~/Vaults/ai. |
|---|---|
MAX_INJECT_TOKENS |
Optional: Token cap for injected context, default 1500, clamped 100-50000. |
SESSION_TTL_HOURS |
Optional: Hours before an agent session without heartbeat expires, default 2. |
SUPERSKILL_MAX_RESPONSE_BYTES |
Optional: Byte budget per MCP response, default 65536, max 10 MiB. |
SUPERSKILL_TELEMETRY |
Optional: 1 opts in to local-only telemetry, 0 turns recording off. |
SUPERSKILL_CACHE_ROOT |
Optional: Root folder for shared worktree caches; defaults to the platform cache folder. |
SUPERSKILL_FORCE_MCP |
Optional: Set to 1 to force server mode instead of the CLI. |
SuperSkill is a local stdio MCP server, plus a CLI from the same npm package, that sits between a coding agent and the project it works on. Its main tool, `superskill`, takes a task description, picks curated rule packs by language, phase and specialty (for example Go, QA, security review), and returns a short brief and a rule plan cut to a token budget, with a record of what was picked and what was left out.
Around that it keeps a per-project memory: a markdown vault under `projects/<slug>/` (by default in `~/Vaults/ai`) holding decisions, tasks, learnings, sessions and notes, with an SQLite FTS5 index and link graph built from it. Notes can be opened in Obsidian, and `knowledge_viz` writes Mermaid architecture and code notes into the vault. Further tools cover spec, ticket and evidence gates, a tree-sitter code graph (`impact`, `claims`), sharing package and compiler caches across git worktrees, a read-only disk-space report, and a watchdog that reviews past agent sessions.
In version 0.15.1 the server registers 52 tools; the docs' tool table lists 51 and leaves out `skill_validate` (src/core/registry.ts). Calls are rate-limited per server in a 60-second window: 30 write calls, 60 reads and 10 expensive diagnostics, after which it returns `RATE_LIMITED` (src/mcp-server.ts).
| Runtime | Node.js 22+ |
|---|
Node.js 22.16 or later on the 22.x line, or Node 24+, built with SQLite FTS5; official Node 22.13 to 22.15 builds lack FTS5 and cannot run knowledge search. npx for the launch command. The vault folder (`VAULT_PATH`) must sit under your home directory. Running `superskill skill init` inside a repository is the docs' first step and makes vault tools find the project without naming it.
Vault access is confined to the current project's folder; cross-project search is refused, and a workspace that has not been registered cannot use implicit project memory. Responses over 64 KiB by default (`SUPERSKILL_MAX_RESPONSE_BYTES`, up to 10 MiB) come back as a truncated JSON preview. Context injection is capped at 1500 tokens by default (`MAX_INJECT_TOKENS`, 100 to 50000). The docs' launcher uses `superskill@latest` with `--prefer-online`, so each start fetches the newest release and may fail to start offline. The makers say the secret and prompt-injection scans are heuristics and that rule-pack effect on model output is unproven.
Suits a developer who runs coding agents (Claude Code, Cursor, OpenCode, Codex CLI) on their own machine and wants project decisions, tasks and session notes kept as local markdown, with rule packs chosen per task instead of a large skill catalogue. It also fits teams juggling many git worktrees who want shared caches and cleanup reports. It is a large tool surface (52 tools) that writes to your home directory and client configs, so someone who only wants a simple notes or memory store, or cannot run Node 22.16+, should pick something smaller.
superskillreadwritesearchproject_contextgenerate_contextdecidetasklearnbrainstormsessionresumestatsprunedeprecateinitstatustelemetryregisterdoctorskill_installskill_validateskill_list_installedskill_removelinkextractcapturetemplategraph_relatedgraph_traversegraph_cross_projectknowledge_rebuildknowledge_vizsnapshot_repo_stateenv_factscred_refsrollbackspecticketsevidencegateimpactclaimsworktree_statusworktree_auditworktree_envworktree_activateworktree_applyworktree_gcworktree_uninstallhygiene_reportwatchdog