SkillAgentSearch skills...

icemage

Token-efficient context engine for AI coding agents. v2.23.0: token-killer pack - deep-forget (unlearning propagation), dangling-reference guard, MCP schemas-on-demand. 2694 tests, 43 MCP tools, Elastic-2.0.

Install / Use

claude mcp add ncmonx -- npx -y github:ncmonx/icemage

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

81/100

Supported Platforms

Claude Code
Claude Desktop
Cursor
Cline

Tags

Our assessment of icemage

icemage scores 81/100 on our quality scale, 717th of 944 AI & Machine Learning skills we index.

Its MCP Server is 19 KB long, well organised into 13 sections with 3 code examples: a thorough specification that gives an agent plenty to work with.

It has 3 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
30/30
Structure
18/20
Description
15/15
Adoption
3/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated today, so icemage is actively maintained.
  • Our last check on 2026-11-12 found the source still online.
  • No license is declared. By default that means all rights are reserved: you can read it, but reusing or redistributing it is not clearly permitted. Ask the author before building on it commercially.
  • Its trust signals score 80/100, with 2 cautions from licensing, adoption, age or documentation. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

Safety scan

No issues found

Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands (2 minor notes below). An AI review of the same text found nothing harmful.

  • noteInstalls by piping a downloaded script into a shellline 94
    curl -fsSL https://raw.githubusercontent.com/ncmonx/icemage/main/scripts/install.sh | sh
  • noteInstalls by piping a downloaded script into a shellline 100
    irm https://raw.githubusercontent.com/ncmonx/icemage/main/scripts/install.ps1 | iex

AI review by kimi-k2.7-code on 2026-10-02. Automated pattern scan on 2026-10-02. It catches known dangerous patterns, not every risk — read a skill before letting an agent act on it.

icemage compared with similar skills

All 4 of these similar skills score higher than icemage; compare them before choosing.

SkillScoreStarsUpdatedFormat
icemage (this skill)by ncmonx813todayMCP Server
claude-memby thedotmack100118.5ktodayCLAUDE.md
Agent-Reachby Panniantong100113.3k12d agoCLAUDE.md
Understand-Anythingby Egonex-AI10093.3k1d agoCLAUDE.md
hyperframesby heygen-com10087.1ktodayCLAUDE.md

Frequently asked questions

How do I install icemage?
Run claude mcp add ncmonx -- npx -y github:ncmonx/icemage. The install tabs above show the steps for each supported agent.
Which AI agents does icemage work with?
It is written for Claude Code, Claude Desktop, Cursor and Cline, as a MCP Server file. Other agents that read the same format can often use it too.
Is icemage safe to use?
Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands (2 minor notes below). An AI review of the same text found nothing harmful. It declares no license and scores 80/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is icemage still maintained?
The repository was last updated today, so icemage is actively maintained.
<p align="center"> <img src="assets/logo.svg" alt="Icemage" width="160"/> </p>

Icemage <sub><sup>(icmg)</sup></sub>

release downloads last-commit tests mcp tools commands license OpenSSF Scorecard sponsor ko-fi

Stop burning tokens. Stop losing context. Ship faster.

A single binary that makes Claude Code, Cursor, and every other AI coding assistant 70–90% cheaper to run — without dumbing them down.

If you've ever watched 30K tokens evaporate on a single file read, paid for "thinking" you didn't need, or re-explained the same project context after /clear for the fifth time today — this is for you.

<p align="center"> <!-- 30-second demo. Regenerate: `vhs assets/demo.tape` (see assets/demo.tape header), commit assets/demo.gif, then uncomment the <img> below. --> <!-- <img src="assets/demo.gif" alt="icmg in action — savings, one-shot find, slim context" width="760"/> --> </p>

🟢 Why Icemage

AI assistants are powerful but wasteful by default. Every time the AI opens a file, runs a command, or starts a new chat, it re-reads context it has seen many times and dumps full output into the conversation. Icemage sits quietly in the background and trims the noise before it ever reaches the AI:

  • Long files → only the relevant slice
  • Noisy command output → just the parts that matter
  • Web pages → cached + summarised
  • Past decisions → remembered across sessions so the AI doesn't ask twice
  • Repeated work → results reused instead of recomputed

The AI keeps its full intelligence. Your wallet keeps more of its money.


📊 Headline numbers

| Metric | Typical | Best | Since | |---|---|---|---| | File-read savings | 70 – 85 % fewer tokens | up to 92 % | v0.5 | | Test / build output | 60 – 80 % shorter | up to 90 % | v0.5 | | Multi-file UI propagation (style-clone) | 30 – 50× cheaper | up to 98 % | v1.22.0 | | Cross-project bundle (port) | 8 – 12× cheaper | up to 95 % | v1.24.0 | | Compressed-Write (AI emit diff) | 70 – 95% fewer tokens | up to 98 % | v1.25.0 | | Web-fetch reduction | 70 – 90 % smaller | up to 95 % | v0.4 | | Repeat-context recall | near-zero, < 5 ms cached | — | v1.21.8 | | Past-chat full-text search | < 10 ms across months | — | v1.21.7 | | Graph symbol lookup | 256-slot in-RAM cache | — | v1.21.8 | | First-prompt warmup | < 1 s | — | v1.18 | | Cold build time (icmg itself) | ~50 % faster (20 min → 9-10 min) | — | v1.26.0 | | MCP response filter (verbose plugins) | 50 – 80 % smaller | up to 90 % | v1.30.0 | | Auto-thinking suppress (trivial prompts) | ~1500 tok / call saved | — | v1.30.0 | | Sayless-auto (long-prose replies) | 60 – 75 % compress | up to 85 % | v1.30.0 | | Service auto-start (UserPromptSubmit) | 0-touch warm-up | — | v1.30.0 | | Path ambiguity warning (icmg context) | wrong-file lookups → loud | — | v1.29.0 | | rg-wrapper + brace glob (icmg grep/files) | flag-mirror, {a,b} expand | — | v1.29.0 | | Local AI model (built-in, opt-in) | 0 cloud calls | privacy-first | v1.31.0 | | Smart router (REGEX vs LLM_LOCAL vs CACHE) | <100 us p99 | hot-path forced regex | v1.31.0 | | HTTP streaming download (model fetch + SHA256) | 400 MB - 2 GB safe-verify | tamper-detect | v1.31.0 | | icmg git wrapper (single ergonomic entry) | Tkil-filtered + safety-gated | enforces icmg-FIRST | v1.31.0 | | Python-free core (PRECOMPACT_PY dropped) | -200-500 ms boot saved | single-binary | v1.31.0 | | pack --rerank (LLM-reorder memory hits) | opt-in warm-path | router-gated | v1.32.0 | | PreCompact LLM summary (warm-pool Qwen 0.5B) | <15 s cold | regex fallback always | v1.32.0 | | icmg compact-bg (proactive memory worker) | <3 s warm | manual + future hook | v1.32.0 | | Smarter local AI memory | multi-prompt safe | no overflow | v1.32.0 | | Code graph viz + report (icmg graph viz) | interactive D3 + god-nodes | — | v1.71.0 | | Secret scanner (icmg scan) | 21 detectors, CI-gate | redact-by-default | v1.68.0 | | MCP server hardening (token + rate-limit + path-guard) | abuse / RCE-safe | — | v1.72.0 | | Post-compact memory re-anchor | rules survive compaction | auto on init | v1.73.0 | | Scripted-safe icmg run (non-interactive guard) | no hang on destructive | --yes/env opt-in | v1.74.0 | | Clean self-upgrade (idempotent Defender step) | no phantom B: drive popup | --no-defender opt-out | v1.75.0 | | Encryption-at-rest (icmg encrypt, SQLCipher AES-256) | opt-in full-DB encrypt | BM25 recall intact | v1.76.0 | | Hot recall cache (RAM, daemon-shared) | < 5 ms repeat recall | self-governing RAM | v1.77.0 | | Cost per AI session | down 70 – 90 % vs. raw | up to 95 % | — |

✨ What's new

  • v2.24.0 — Landscape pack: grounded memory, honest gaps, real-workload bench, wired compression. Seven deterministic, zero-LLM features from the 2026-09 AI/AGI landscape scan (agent-memory + context-compression + repo-view + MCP-security literature). memory-health --grounding — a memory citing a file the code graph no longer knows was probably written before a rename/delete (arXiv 2609.11060); probes our own graph, flags stale-suspects with the vanished refs, never deletes — noise-controlled (import specifiers, TS js→ts twins, system headers skipped). Evicted-vs-missed gap tags — memory-health --gaps now says which remedy fits (arXiv 2609.08279): [missed] = never held → store it; [evicted] = a soft-deleted memory matches the query (≥2 strong tokens) → icmg memory restore <id>. bench-recall --replay — judges the brain on its REAL workload (arXiv 2609.24971): replays query_history against the current store; hit-rate, regressed queries (answered then, silent now = prune/forget damage), and answer token cost; exits 1 on regressions (CI-gateable). Confidence cue on auto-recall — every surfaced hit reads [score|high/med/low] relative to the injection floor (arXiv 2609.24259: models over- and under-trust injected memory without calibration). shrink --kind salience --graph-coherence — structure-blind per-line scoring drops the line citing registry.hpp while keeping its graph-wired neighbor dispatcher.cpp; the coherence pass pulls back dropped lines citing files graph-adjacent to kept citations (arXiv 2608.30811, opt-in, capped). graph-skeleton --for <task> --focus — restricts the view to the task-seed + 1-hop neighborhood and boosts directly-seeded files, so a sparse task seed is no longer drowned by global hub files (arXiv 2609.16936). MCP per-tool call policy — exposure filtering is not access control: ICMG_MCP_DENY=a,b,c and ICMG_MCP_READONLY=1 (denies the 14-tool mutating subset) now hold at tools/call time and hide denied schemas from tools/list; default-off. Also: the Windows zip pollution root cause fixed at the source (build.ps1 + pack-win.ps1 share one 22-DLL whitelist, hard content gate deletes a bad zip before upload). 35 new tests — 2497/2497 tests ✓.
  • v2.23.0 — Token-killer pack: deep-forget, dangling guard, schemas-on-demand. Three deterministic, zero-LLM features from the 2026-09 research pass. icmg forget <id> --deep — forgetting soft-deletes ONE node, but its content typically leaked into derived artifacts (session snapshots, wflog entries, consolidated notes); --deep scans live nodes for residue by directional containment (symmetric Jaccard under-scores a long snapshot holding a verbatim secret) — flag-only, never auto-deletes. Dangling-reference guard in shrink --kind salience — per-line selection can keep a line using an identifier while dropping the line defining it (34–60% of compressed outputs in the arXiv 2608.04569 benchmark); a post-pass pulls back first-mention lines of entities the kept lines reference (--no-dangling-guard opts out). icmg_tool_search MCP meta-tool — with ICMG_MCP_PROFILE=core the host sees 11 tools instead of 43 full schemas saturating the context window before the first query; the meta-tool ranks ALL registered tools by capability query and returns full inputSchema per hit, schemas-on-demand, hidden tools stay callable. 15 new tests — 2462/2462 tests ✓.
  • v2.22.0 — Brain v2.22: the memory learns from its own usage. Four deterministic, zero-LLM memory upgrades from the 2026-08 research pass (Zep/Graphiti temporal-KG, Mem0, MemoryOS, coarse-to-fine grounded memory). Time-travel recall — recall --as-of T (epoch, 7d/24h/30m ago, or YYYY-MM-DD) answers what did we believe at T? over the existing bi-temporal columns: since-superseded facts reappear before their invalidated_at, facts not yet valid stay hidden, and the query is side-effect-free (no cache, no frequency bump). Retrieval-failure ledger — memory-health --gaps turns recall queries that came back empty into a knowledge-gap checklist: the agent asked, the brain had nothing — most-asked first, noise skipped, --json for tooling. Quick-note promotion by heat — memory-consolidate --promote-quick gives quick:<epoch> captures the agent kept recalling (≥ 3 recalls) a permanent searchable hot:* topic while cold ones age out via decay. Coarse-to-fine recall — oversized result sets keep the strongest hits full-bodied and collapse the tail to 1-line index rows (--get <id> fetches detail; --full opts out). 23 new tests — 2447/2447 tests ✓.
  • v2.21.1 — Multi-daemon spam fix — dozens of idle icmg.exe no more. On busy multi-user servers, concurrent hook calls raced ensureDaemon(): every spawned rule-daemon created the same named pipe successfully (PIPE_UNLIMITED_INSTANCES, no FILE_FLAG_FIRST_PIPE_INSTANCE, no cross-process lock) and then parked in ConnectNamedPipe forever — accumulating dozens of ~2-3 MB icmg.exe processes per user. Fixed belt-and-braces: RuleDaemon::acquireSingleton() per-user cross-process lock (Windows: named mutex Local\icmg-rule-daemon-<user>; POSIX: flock on ~/.icmg/rule-daemon.lock) required by run() — the losing daemon exits quietly — plus FILE_FLAG_FIRST_PIPE_INSTANCE so a raced second daemon can never share the pipe even if the mutex path regresses. One-time cleanup on affected servers: taskkill /F /IM icmg.exe, then update. 2419/2419 tests ✓.
  • v2.21.0 — Brain + token trio: session-aware recall delta, contradiction sentinel, adaptive recall depth. When the session TTL dedup suppresses memory nodes you already saw this session, recall now emits a single stdout line [N prior memories still apply: #ids] instead of silently dropping the reference — the agent keeps the pointer without re-paying the tokens for full bodies. New icmg memory-health --contradictions scans the memory store for node pairs that overlap heavily (Jaccard ≥ 0.6 default, --jaccard-min to tune) yet disagree — a negation marker on one side or a conflict

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars3
CategoryAI
Updated18h ago
Forks1

Languages

C++

Trust signals

80/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 medium1 low