Skip to main content
Glama

IcyLoom

One MCP server. One SQLite file. One ledger that tells you what it actually saved.

Python License MCP Tests Bind PyPI GitHub stars README views

IcyLoom V2 — the MCP Tool Gateway

What's new — 2026-11-04 — V2.2.0 (fail-closed, honest status)

Change

What it means

Unknown tools are denied

A tool absent from the registry used to be allowed with side_effects="none". Now it is unknown — denied by default (`unknown_tool_policy: deny

tool_simulate matches invoke_tool

One single decision point for both. A reader asking for a destructive tool gets deny, not approval.

pwsh classified by the command

Get-Date, Get-ChildItem, Test-Path pass for a reader; anything else stays destructive.

Real per-server status

up / down / no-tools + tool count + last error, exposed by proxy_info and /api/tools/catalog. The console badges no longer lie.

Drift noise gone

Schemas are normalized before hashing ($schema stripped). 910 cosmetic entries purged — only real changes are journalised.

Cache metrics are real

cache_read/cache_write are actually recorded (they were hard-coded to 0).

A live economy metric

The toolbar shows a number that moves — 💰 N tok économisés · X % — re-queried from SQLite every second, with a ticker that eases to the true value and a live session clock. The frozen 143 left the headline; the ledger row count stays as a secondary reference.

The economy is counted once

Disjoint domains only — cache LLM, compression, tools, tool-cache. The previous formula counted the same compression twice, so the toolbar and the registry card disagreed. They now agree, and /api/live returns the breakdown per domain.

Catalogue: 102 server entries

18 active, 84 dormant in the catalogue, listed with a real per-server Tester and a pager (25 per page); the dormant ones stay disabled until enabled.

What's new — 2026-11-04 — V2.1.0 (trust signals)

Change

What it means

The gateway reads what the servers declare

The MCP annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) complete the side-effect classification. They can only escalate it — never lower it — unless the server is explicitly trusted.

A changed tool definition is journalised

Every definition is hashed (description, schemas, annotations). A change lands in tool_drift with the field, the old value and the new value. A server that changes behaviour between two versions must be visible, not silent.

Structured results are checked

The structuredContent of a result is verified against the outputSchema the server declares. A mismatch is reported (schema_valid: false) instead of being served silently.

Results are scanned for injections

Instruction override, role hijack, credential leak, forged tool call, hidden markup: a flagged result is served with an untrusted block and a warning. The call itself is never blocked.

Two new meta-tools

tool_drift(server?) — what changed; tool_simulate(server, tool, args) — the decision the policy would take, without executing anything. 21 tools in total.

The console shows the trust signals

A new Trust signals panel in the Gateway V2 view: what the downstream servers declare, the definition drift, and the results that were flagged.

A new mark, visible in the console

The ink-and-paper identity is kept, the palette turns bluer (blue warp threads, violet weft). The mark now shows at 132 px in the console hero instead of hiding in the favicon.

What's new — 2026-11-03

Change

What it means

The system tray ships with the package

The gateway starts with your session — one click from the tray, no terminal juggling.

Agent clients are read-only and keep their own roots

An agent cannot reach outside the project roots it declares; a write outside them is refused outright.

Downstream MCP servers get a usable environment

An empty environment used to kill uvx / npx at spawn — the servers now start.

Every active server is indexed at startup

SQLite and the others were missing from the registry; the FastMCP lifespan now receives the server, so a refresh actually starts.

The registry accepts the real tool_proxy.servers shape

active_servers reads the list the config actually uses.

Dashboard rebuilt

Window mask, cropped relief/topo, honest axes, legends, tooltips, ranked cache chart.

Figures republished from a real measurement

194 tools across 14 servers, real tokens (tiktoken), gain per domain. The numbers below are the measured ones.

V2.0.0 is released (release notes). V2 turns a pile of MCP servers into one governed gateway: every tool catalogued, the right tool selected from a plain-language request, permissions enforced, context compressed, every decision logged.

194 tools · 14 servers · 21 MCP tools · zero embeddings · 127.0.0.1 only

V1 solved the cost problem: it cached LLM answers, compressed a context drawer, relayed tools. It did not solve the control problem — a client that could see every downstream tool could also use every downstream tool. V2 is the missing half.

V1

V2

Tools

relayed, opaque

catalogued — schema, intent, side effects, scope, risk

Selection

the client picks

the gateway picks, from a natural-language request

Permissions

one global switch

roles, rules, approvals, project scope

Context

raw tool schemas

compressed schemas and results

Trail

token ledger

token ledger + decision audit

Install

pip + config

pip + config — still no embeddings, no GPU, no external service

Eight meta-tools

The agent talks to the gateway, not to 194 individual tools.

Meta-tool

What it does

search_tools(query)

Finds the tools matching an intent written in 12 languages. Lexical index (SQLite FTS5/BM25), no embeddings.

get_tool_schema(server, tool)

The compressed contract of one tool — what the agent actually sees.

invoke_tool(server, tool, args)

Runs one downstream tool through policy, cache and compression.

run_tool_script(script)

Chains several calls inside a sandboxed Python script.

tool_policy()

The policy in force: roles, scopes, approval rules.

gateway_stats(days)

Tokens per category, registry, audit — the observability view.

tool_drift(server?)

Definition drift: a downstream tool changed — schema, description, behaviour.

tool_simulate(server, tool, args)

The decision the policy would take, without executing anything.

21 tools in total (13 from V1 + these 8). A destructive action without confirm=true is refused with no side effect.

The six pillars

  1. The registry is the source of truth — every tool described once (schema, domain, intent, side effects, security level, scope, owner, risk), in the same SQLite file as everything else.

  2. Selection that understands the request — FTS5 narrows the candidates, Python scores them, and the request is translated into the canonical concepts first, so a French sentence reaches the English-named tool that answers it.

  3. Policy, roles and approvals — reader / writer / operator / admin, rules allow / deny / approval, most specific rule wins. A destructive effect always needs an approval; a write outside the project roots is refused outright.

  4. Compression that keeps the prompt small — schemas and results are compressed, property order is preserved (reordering would invalidate the provider's prompt cache), secrets are detected on the way through, and the raw result stays retrievable by result_ref.

  5. Observability and audit — every decision is written to tool_audit: actor, role, intent, tool, decision, reason, compression applied, tokens saved.

  6. Trust signals — the server's declarations are read, not assumed — the MCP annotations (readOnlyHint, destructiveHint…) complete the side-effect classification (they can only escalate it, never lower it), a changed tool definition is journalised in tool_drift, a structured result is checked against its declared outputSchema, and a result that tries to inject instructions into the model is flagged instead of being served silently.

Benchmarks — measured, not claimed

Measured on the development instance (Windows 11, i7 vPro, no GPU) against a live stack of 14 real MCP servers, after a 46-call load pass through the gateway:

Metric

Value

Tools indexed

194 across 14 servers

Read-only / write / destructive

134 (69.1 %) / 45 (23.2 %) / 15 (7.7 %)

Decisions traced

58 — 40 allowed, 17 approvals, 1 refused (outside project scope)

Tokens saved by tool compression

118 654 — real tokens (tiktoken)

Tokens measured (window)

837 431 — 812 407 in tool schemas, 25 024 in tool results

Test suite

20 mission tests (one per finding) + full suite green, exit code 0

The console was then put under load in the same run — 40 real calls through the gateway (10 prompts × 4), no mock:

Metric

Value

Calls through the gateway

40 — 0 error, 10 distinct prompts

Served from IcyLoom's own cache

30 — ~215 ms each, 0 token spent

Real generations (upstream miss)

10 — p50 6.2 s, p95 10.4 s

Hit rate on the console window

78.5 % — 102 hits / 28 misses (13 errors excluded)

Tokens avoided

373 190 — cache 20 250 + compression 234 286 + tools 118 654

Gain on the window

48.4 % — 373 190 avoided / (398 367 paid + 373 190 avoided)

Local cache after the run

8 entries, 4 reuses each

Ledger rows

623

Every figure above is read from the SQLite file the console itself reads — the same file you can open.

End-to-end run against the real MCP server over stdio, no mocks:

TOOLS=19  META_OK=True
lis un fichier du depot   -> github.get_file_contents | desktop-commander.read_file | filesystem.read_file
supprime un fichier       -> github.delete_file | local-rag.delete_file | basic-memory.delete_note
cree un ticket github     -> github.issue_write | memory.create_entities | github.add_issue_comment
read_file README.md       -> ok=True, real content returned
write_file                -> ok=True, file really written
github.delete_file (no confirm) -> error=approval_required, call never reached the server
run_tool_script           -> ok=True, sandboxed
policy: role=reader env=dev roots=[] (client-declared roots only)

Gateway V2 view of the console: five KPIs with success badges, three charts, the registry, the policy in force and the decision audit — all live.

Release status

V2.2.0 is on main (changelog). It adds: registry, lexical selection in 12 languages, the six meta-tools, roles and approvals, compression, the sandbox, the audit trail, the console view, the system tray icon, the trust signals (MCP annotations, definition drift, structured-output verification, untrusted-content detection) with two extra meta-tools, tool_drift and tool_simulate, a new Trust signals panel in the console, a fail-closed relay (an unknown tool is denied by default), real per-server status, a catalogue of 102 server entries (18 active, 84 dormant) with a pager and a real per-server Tester — and a live economy metric in the toolbar: GET /api/live re-queries SQLite every second and sums disjoint domains, so the headline and the registry card always agree.

Full announcement: ANNOUNCEMENT.md · Session log: progress.md


A hybrid gateway for local and cloud LLM workflows that puts an LLM cache, a compressed context drawer and an MCP tool relay behind one MCP server, one configuration file and one SQLite database.

IcyLoom is a standalone project: no imports, links or paths to any other project. The code is written from scratch; the only files taken from elsewhere are data (normalisation rules, a tool registry), because the cache hash depends on them.

Cache

Every prompt normalised, hashed and stored. A hit is a zero-token answer.

Context drawer

Sources filed, chunked, anchored, invariants checked — compressed on demand, opt-in.

Tool relay

Downstream MCP servers, cached by canonical signature, deny by default.

Registry

v_token_ledger: tokens paid upstream and tokens avoided, in one unit, one view.

Console

A hand-drawn web console on :8153 — side panel, live views, working buttons.

Privacy

Binds 127.0.0.1. Local mode stays on-device; configured cloud upstreams receive only the traffic you route to them.

The numbers are real. A cache hit is counted as avoided, a miss is counted as paid with the upstream's own usage figures — never estimated. The 35 % threshold is read straight from the ledger.

Overview: the unified registry, five live KPIs, calls per bucket, latency, outcomes, models, upstreams. Captured on the live development instance: every tile is a real call read from SQLite.



Hermes Agent Plugin (Catalog Ready)

IcyLoom is packaged as a Hermes Agent plugin — installable by name once listed in the plugin catalog:

# After catalog merge (pending PR)
hermes plugins install icyloom

The plugin declares 21 MCP tools and runs the full gateway stack (ports 8150-8153). See plugin.yaml for the manifest and __init__.py for the entrypoint.


Quick start (Standalone)

Five minutes, on a clean machine. No GPU is required, no provider lock-in is imposed, and no cloud account is needed for local mode. Route to local or cloud upstreams when configured.

Option 1: PyPI (Python)

pip install icyloom
# or with pipx (recommended for CLI tools):
pipx install icyloom

Option 2: NPM (Node.js wrapper)

npm install -g icyloom
# Requires Python 3.11+ — the wrapper delegates to Python

Linux / macOS

python3 -m venv venv
source venv/bin/activate
python -m pip install --upgrade pip
pip install icyloom

icyloom init
icyloom doctor
icyloom tray --install-autostart

# in a second terminal
icyloom serve
icyloom check --transport all

Windows (PowerShell)

py -3.12 -m venv venv
.\venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install icyloom

icyloom init
icyloom doctor

# the system tray icon — starts with your session
icyloom tray --install-autostart

# in a second terminal
icyloom serve
icyloom check --transport all

Pinned to a specific release (git required):

python -m pip install "icyloom @ git+https://github.com/MlkBlk-IT/icy-loom.git@v2.2.2"

Expected output of icyloom check: 21/21 vérifications réussies.

icyloom check plays the role of the client: it starts the server, connects, lists the 13 tools and calls them (ctx_store, ctx_search, ctx_verify, cache, ledger, proxy_info). It tests the three transports (stdio, streamable-HTTP, SSE) and prints 21/21 vérifications réussies when all is well. If HTTP or SSE fails, the stack is not running (icyloom serve); the stdio mode, on the other hand, is self-contained — it is the one Claude Code and Cursor use.

Then plug IcyLoom into your MCP client:

icyloom client --target cursor --install     # writes the config on your machine
icyloom client --target claude               # prints the block to paste

Targets: cursor, claude (Desktop), claude-code, windsurf, codex, json (raw block).

icyloom.yaml      single configuration (Pydantic)
src/icyloom/data/ externalised data, shipped inside the package (JSON)
var/icyloom.db    single SQLite database (WAL)

  HTTP gateway :8150                      MCP server (one build function)
    /v1/chat/completions                    stdio             dedicated process
    /v1/messages                            streamable-HTTP   :8151
    /v1/models, /v1/embeddings              SSE               :8152
    /{upstream}/v1/...                      13 tools, one surface
    cache -> routing -> upstream

  admin API + dashboard :8153
  gateway, MCP and admin all write to one database and one ledger

Outside a repository checkout, nothing is written inside site-packages: the configuration, the database and the logs go to the user data directory — %LOCALAPPDATA%\icyloom on Windows, $XDG_DATA_HOME/icyloom (or ~/.local/share/icyloom) on Linux and macOS. Override it with ICYLOOM_DATA_DIR.

Related MCP server: mcplex

The system tray

The icon is part of the package: on Windows pip install icyloom already brings it in, nothing extra to install. It starts the stack, opens the console, shows the state of the four services, writes the MCP configuration of your clients in one click, and can start with your session.

icyloom tray                        # launch the icon
icyloom tray --install-autostart    # start it with Windows
icyloom tray --status               # stack, services, autostart

Full guide: docs/INSTALL.md · docs/TRAY.md

Status & known limitations

This section is the honest summary. Raw measurements live in docs/archive/PROGRESS.md.

Area

Status

Installation from a wheel on a clean machine

Verified. init, doctor (exit 0), serve, check 21/21, from a venv outside the repository

Test suite

223 tests collected by pytest (pytest tests --collect-only -q)

Cache round trip (miss then hit)

Verified by an automated test: the first request calls the upstream, the identical second request does not, the served response is identical, and the ledger records both

Tool relay

Verified by an automated test against a real MCP stdio server: the tool is reached, it answers, and the second identical call is served from the cache

Tool-result cache correctness (stale hits, replayed writes, cached errors, loop detection)

Verified. Synthetic harness: 0 stale hits out of 140 (before the fix: 425 out of 436). Real fetch server: 8 of 8 checks pass

Gateway latency

p50 13.5 ms (MISS) and 10.7 ms (HIT); p95 about 16 ms against a 200 ms budget

Latency gain on a slow tool

One real fetch call: 1527 ms first, 24 ms when repeated (one static page)

Token savings on tool calls

Not delivered and not measured. A tool-cache hit is ledgered with 0 tokens saved

Context compression

Opt-in, off by default. ctx_verify proves the invariants survive a round trip, but answer quality is not proven (see Known limitations)

Python versions

3.11, 3.12, 3.13 supported. Python 3.14 is not in the supported range yet

The "50 percent tokens saved" printed by tests/proof_e2e.py is true by construction: that test sends every request twice. It is not a traffic figure and must not be quoted as one.

Getting started

icyloom doctor        # diagnostics
icyloom serve         # gateway + MCP + admin
icyloom mcp --transport stdio   # for an MCP client
icyloom gateway       # gateway only (8150)
icyloom admin         # admin API only (8153)
icyloom paths         # every resolved path

An MCP client connects with a single entry, for example:

{ "mcpServers": { "icyloom": { "command": "/path/to/venv/bin/python",
                              "args": ["-m", "icyloom", "mcp", "--transport", "stdio"] } } }

An LLM client points its base_url at http://127.0.0.1:8150/v1.

icyloom doctor also probes each enabled upstream (/models) and checks that the tool-server executables are on the PATH. A missing upstream or a missing tool executable is a warning, not a failure: a default installation has neither, and still works. When server.remote_access_token is set, every admin and gateway route except /health requires it, and a non-loopback server.host is refused without one.

The tool cache

IcyLoom relays calls to downstream MCP servers and can cache their results. The rules:

  1. Deny by default. Only tools set to true in a server's cache_tools are ever replayed from the cache. There is no guessing from the tool name. A server with cache_enabled: true and no cache_tools caches nothing.

  2. State epoch per server. Any call that is not cacheable (a possible write) advances that server's epoch. The epoch is part of the cache key, so a read cached before a write is never served after it. The epoch also carries a boot token: the cache does not survive a relay restart, because the server may have changed while the relay was down.

  3. Errors are never cached. A result flagged is_error is not stored.

  4. Schema-guided canonicalisation. A near-identical call is recognised only according to the tool's inputSchema: integer, number and boolean parameters, and path-like string parameters. Without a schema, only an exactly identical call matches. Never: case changes, whitespace folding inside strings, reordering of lists. Schemas are loaded when list_tools is called.

  5. The loop guard counts cache hits. A loop answered from the cache is still a loop: the 5th identical call within the window is blocked.

  6. A batch stops at the first error. The remaining calls are reported as skipped_after_error.

tool_proxy:
  servers:
    - name: fetch
      command: uvx
      args: ["mcp-server-fetch@2026.10.10"]
      cache_enabled: true
      ttl_seconds: 3600          # protects against changes made outside the relay
      cache_tools:
        fetch: true              # explicit allowlist: nothing else is cached

The TTL is the only protection against a downstream server whose state changes outside the relay. Choose it per server.

Loading tools fast

Two things make the relay quick to use, both verified by tools\proof_tools.py:

  1. One live session per downstream server. The server process stays alive between calls instead of being respawned for every call. The session lives in its own thread and its own event loop, because the MCP SDK opens a stdio session inside an anyio task group that must be entered and left in the same task. Measured on a local server: the first call costs ~840 ms (process start), the next ones ~21 ms — 40× faster.

  2. The tool catalogue is cached and searchable. proxy_info(what="tools", server=…) lists the tools of a server and keeps the result (with the inputSchema) in memory; proxy_info(what="search", query=…) then finds a tool by name or description — "git" finds git_log, "file" finds read_file. The catalogue is also loaded automatically on the first call to an unknown tool, because the schema is what makes the near-identical recognition possible.

.\.venv\Scripts\python.exe tools\proof_tools.py   # 14 checks: near-hit, no false hit, reuse, search

The schema-guided merge is deliberately conservative: limit: 10 and limit: "10" are one and the same call (the schema declares an integer), but " 8150 " and "8150" are not merged (a string is a value — the downstream tool answered "unknown port" for the padded one).

The console (dashboard)

A web console, served by the admin API, shows all metrics with matching charts:

http://127.0.0.1:8153/dashboard      (alias: http://127.0.0.1:8153/)
.\.venv\Scripts\python.exe -m icyloom dashboard          # URL + current state
.\.venv\Scripts\python.exe -m icyloom dashboard --serve  # serve the console

IcyLoom console

The console on a demo database seeded by tools/seed_demo.py (synthetic traffic) — not production data.

Reading the console

Every chart is live and interactive:

  • hover a bar, a point or a slice → the values appear in a tooltip;

  • click a point → the tooltip is pinned, and it survives the 2-second refresh;

  • click a legend label → that series is hidden, or shown again;

  • the window selector (1 h / 24 h / 7 d / 14 d) re-reads the database on the spot.

One console, two consoles fused

The console is the merge of the IcyCache console and the Loom dash, rebuilt on the unified registry v_token_ledger and drawn in the rough-sketch style (warm paper, brown ink, hand-drawn strokes, hatched fills, stickers, doodles).

It is an application shell, not a long scrollable page: a side panel on the left switches the view, and the content area only ever renders the view you are on.

View

Fused from

Content

Overview

IcyCache + Loom

5 KPIs with sparklines, the unified registry card (tokens paid / avoided / baseline, latency avoided, 35 % threshold), calls per bucket, latency p50/p95, hit/miss donut, per-model breakdown, upstreams & clients

Activity

IcyCache

3D traffic mountain (day × hour, drag to rotate), topographic map with isolines, hourly traffic table, latest requests

Economy

IcyCache + Loom

14-day curve (paid vs avoided, hatched area, dashed average, circled record), gain gauge against the 35 % line, split of the baseline per domain, economy per day

Cache

IcyCache

most reused entries (hatched pseudo-3D bars), searchable table of every stored prompt (hash, model, provider, hits)

Context

Loom

sessions, sources, chunks, invariants, anchors, redactions, stored sources table, sources per session

Tools

Loom

downstream MCP servers, calls per tool, tool cache (entries, bytes, latency avoided), per-tool breakdown

Audit

Loom

live LLM journal and tool journal, both searchable

Guide

IcyCache Help

request-path diagram, how to read the metrics, the 13 MCP tools, shortcuts & tips, live system state

The side panel carries the eight views (each with its own live counter), the system actions — Pause / Refresh / Export / Compact, which really act — and a live status card (calls, hit rate, tokens avoided, gain, ledger rows, database size). A toolbar above the content shows the current view, the LIVE state, the measured request rate and the event count; the control bar holds the analysis window (1 h / 24 h / 7 d / 14 d — it really changes the query), a Reconnect button and the live database pills.

Every button is wired to a real action: Pause freezes the figures (the polling stops, the pill turns PAUSED), Refresh and Reconnect re-query the admin API, Export downloads the exact JSON the page is drawing, Compact shrinks the layout for a laptop screen. Keyboard shortcuts: 1…8 switch view, R refreshes, P pauses, E exports, C compacts.

The eight views: Activity (3D mountain + topo map), Economy (curve, gauge, split), Cache (entries, ranking), Guide (path of a request). Context, Tools and Audit complete the set — see docs/screenshots/.

Everything is computed in real time; no value is frozen in the page:

  • /api/dashboard is polled every 2 s; the server keeps the snapshot fresh in a background thread, so the answer is instant (≈ 10 ms) even on a busy base;

  • requests/min is computed between two real samples and smoothed;

  • each value is interpolated toward its new target, so counters scroll instead of jumping;

  • the window and the pause button apply immediately;

  • every timestamp is UTC — one single clock, from the SQLite defaults to the charts.

The data comes from a single read-only route:

GET /api/dashboard   ->  { kpi, config, registry, surface, series, breakdown, tables, context, recent, db }

registry is the unified v_token_ledger (cache + compression + tool tax, all in tokens), surface is the day × hour traffic grid behind the 3D mountain and the map, series holds the time series (per hour over 24 h, per day over 14 d). GET /api/gateway relays the gateway's live state (circuits, upstreams) from port 8150 so the page stays on a single origin.

The console is self-contained: one HTML file, hand-written SVG charts and a hand-written sketch engine (no rough.js, no framework), no external dependency, no CDN, no network call except /api/dashboard. The tokens shown are the real tokens reported by the upstream.

Reference screenshots (1440 × 960) live in docs/screenshots/: overview, activity, economy, cache, context, tools, audit, guide. The logo is src/icyloom/static/logo.svg — the same file is served by the console at http://127.0.0.1:8153/assets/logo.svg. The mark is the flame of the old logo, now extruded in 3D: ink hatching on top, hatched side faces, a bevelled light edge and a cast shadow on the paper.

Demo data set (fake data, isolated stacks, ports 8150-8154):

.\.venv\Scripts\python.exe tools\seed_demo.py     # continuous traffic, about 4 req/s
.\.venv\Scripts\python.exe tools\check_js.py       # JavaScript syntax check of the console

The functional test campaign

Re-run on 2026-10-28 with the current code: 88/88 checks green (report in var\campaign-report.json).

A hermetic campaign that exercises every layer on an isolated stack (ports 8160-8164, temporary database), with entirely different inputs: 24 distinct prompts, 6 distinct documents, 3 distinct markers, 3 distinct tool calls:

.\.venv\Scripts\python.exe tools\campaign.py       # report + var\campaign-report.json

It covers the cache (variants, anti-collision, key scope), the ledger (real tokens, gain), the context drawer (byte-exact, invariants, discriminating search, isolated sessions), the tool relay, the gateway formats (OpenAI, Anthropic, SSE streaming), MCP HTTP/SSE parity, the safeguards (confirmation, concurrency) and the consistency of the console with the database. Output: one verdict per check and a non-zero exit code at the first failure.

Two blocks play the role of an agent doing real work:

  • N, agentic work: the agent looks up its spec in the drawer, files a real document, asks 8 really different requests (plan, risks, test, talk, threat, rollback, measure, strategy), iterates one request ("v1" then "v2"), then calls 6 distinct downstream tools and closes with an invariants check.

  • O, anti-mixing: interleaved requests and tools, purge, replay.

To prove that a hit returns the right answer, the campaign uses tools/echo_upstream.py: the upstream returns the fingerprint of the request it received (echo:<sha256>). An answer served from the cache therefore carries the fingerprint of its own request, never another one's.

The 21 tools

V2 — the meta-tools (gateway)

Tool

Role

Destructive action

search_tools(query)

selects the tools matching an intent (12 languages, lexical index)

none

get_tool_schema(server, tool)

compressed contract of one tool

none

invoke_tool(server, tool, args)

runs one downstream tool through policy, cache, compression

yes, needs confirm=true for a destructive effect

run_tool_script(script)

chains calls in a sandboxed script

yes, needs confirm=true for a destructive effect

tool_policy()

the policy in force: roles, scopes, approval rules

none

gateway_stats(days)

tokens per category, registry, audit

none

tool_drift(server?)

definition drift: a downstream tool changed (schema, description, behaviour)

none

tool_simulate(server, tool, args)

the decision the policy would take — no side effect

none

V1 — the base tools

Tool

Role

Destructive action

cache(action)

stats · search · normalize · clear

clear needs confirm=true

tool_proxy(calls[])

relay to downstream MCP servers, in a batch (stops at the first error)

none

proxy_info(what)

servers · tools · stats · loop

none

tool_cache_clear(server)

purge the tool cache

yes, needs confirm=true

ctx_store(source)

text · file · cmd: stores the bulk of the context

none

ctx_get(chunk_id)

re-reads the original bytes (byte-exact)

none

ctx_search(query)

search the drawer

none

ctx_verify(session)

proof that no invariant was lost

none

ctx_check(session)

loss and fidelity check

none

ctx_profile(mode)

strict (default) · balanced · aggressive

none

ctx_session(action)

list · purge · transfer

purge needs confirm=true

ctx_admin(action)

stats · health · vacuum · purge_expired · rebuild_index

vacuum / purge_expired / rebuild_index need confirm=true

ledger(view)

summary · daily: measured cost and gain

none

Every tool carries ToolAnnotations (read-only, destructive, idempotent, open-world). A destructive action without confirm=true is refused with no side effect.

Capability map: tools and routes

(Capabilities from spec section 5. ✔ = covered and tested · ◐ = partly covered · ✘ = not implemented, announced)

Cache and network

#

Capability

Where

A1

OpenAI /v1/chat/completions

gateway.py ✔ (tested)

A2

Anthropic /v1/messages

gateway.py ✔

A3

/v1/embeddings, /v1/models

gateway.py ✔

A4

routing /{upstream}/v1/...

gateway.py ✔

A5

priority: URL prefix > header > body > default, plus sticky

upstream.py ✔

A6

canonical normalisation + hash

normalize.py ◐ (1122 rules; 70 % agreement with the old cache, see below)

A7

LLM cache (TTL, LRU, cap)

cache.py ✔

A8

semantic cache (Qdrant)

✘ opt-in, not implemented (config and cache key ready)

A9

real tokens (never estimated on the upstream side)

gateway.py, ledger.py ✔

A10

prefix cache

◐ prefix_events table present, proxy not implemented

A11

tool relay (single + batch)

toolproxy.py ✔ (batch stops at the first error)

A12

tool cache: per-server TTL, deny-by-default allowlist, state epoch, no cached errors, schema-guided canonicalisation

toolproxy.py ✔ (synthetic harness + real fetch server, 2026-10-27)

A13

loop guard (counts cache hits)

toolproxy.py ✔

A14

resilient pool: drain, recycle, retry

pool.py ✔

A15

circuit breaker with real re-arming

breaker.py ✔

A16

WebSocket /ws/events

✘ not implemented

A17

REST API (health, status, metrics, cache-clear, tools, ledger...)

gateway.py, admin.py ✔

A18

HTML dashboard

dashboard.py + static\ ✔ (covered by tests/test_dashboard.py)

A19

stdio / SSE / streamable-HTTP transports

mcpserver.py ✔ (all 3 tested)

A20

tool registry, assertion total == 100

cli.py doctor ✔

A21

IcyCache's 11 tool-relay tools

merged into the 13 (cache, tool_proxy, proxy_info, tool_cache_clear) ✔

A22

CLI doctor/paths/version/serve/mcp/gateway/admin/migrate/init/config get

cli.py ✔

A23

cache warm-up (warmup)

✘ not implemented

A24

stdout = pure JSON-RPC, logs on stderr, UTF-8 forced

logsetup.py, mcpserver.py ✔ (tested)

Context

#

Capability

Where

B1

17 context tools

merged into the 13 (ctx_*) ✔

B2

3 profiles, strict lossless by default (for invariants; see Known limitations)

context/compress.py ✔

B3

invariants (numbers, paths, URLs, versions, fingerprints, ports, dates, constraints)

context/anchors.py ✔

B4

ctx_verify: proven coverage

context/verify.py ✔

B5

FTS5 drawer + search

context/drawer.py ✔ (vector search ◐ absent)

B6

UUIDv7 sessions, TTL, isolation, purge

context/sessions.py ✔

B7

cross-session transfer

context/store.py ✔

B8

database / column encryption

✘ not implemented

B9

secret redaction before storage

context/redact.py ✔

B10

fail-closed sandbox

context/sandbox.py ✔

B11

rate limiting

context/ratelimit.py ✔

B12

JSONL audit + rotation

context/audit.py ✔

B13

auto-store of a tool output

✘ not implemented (threshold exposed in the config)

B14

FAISS vector index

✘ opt-in, not implemented

B15

admin CLI (health, stats, sessions, vacuum, purge)

admin.py, cli.py ✔

B16

dashboard

dashboard.py + static\ ✔

The ledger

A single ledger_entries table records everything that costs and everything that saves: LLM calls (real upstream tokens), cache hits, context compression, tool calls. The 35 % threshold is readable there continuously:

tokens saved / (tokens paid + tokens saved)

The tokens of an LLM cache hit are the real tokens measured on the original call: they are never estimated. Compression tokens come from the local counter (tiktoken if available, otherwise an estimator), and the counter used is recorded in every entry.

Tool path: a tool-cache hit is ledgered with saved_latency_ms and the tokens actually spared (tool_tokens_saved, measured by the local counter on the cached payload). A tool-cache hit therefore saves latency and tokens: the model never re-encodes a result it already has.

Known limitations

  • Tool-call token savings are measured by a local counter, not by the upstream API: exact with tiktoken installed, estimated otherwise. The counter used is recorded on every ledger entry. See the Status section.

  • Compression quality is not proven. In a prototype run on a real model (carried over from earlier notes, details in PROGRESS.md), estimated prompt tokens dropped by about 84 to 92 percent, but the model answered incorrectly: the anchor keeps only a header, the invariants and the first 800 characters, so a key line in the middle of a log was lost. Compression therefore stays opt-in; "strict" guarantees that the invariants are kept, not that the text the model sees is unchanged.

  • The gateway cache uses one slot per key shared by JSON and streaming responses, so a streamed request can miss a hit created by the JSON one.

  • An upstream ReadTimeout during streaming is not caught in gateway.py (around line 578) and can end a client run.

  • The tool-cache epoch is verified on a synthetic server; a real server with both reads and writes has not been tested yet.

  • Downstream state that changes outside the relay is only covered by the TTL.

Design decisions

  1. One MCP server, two HTTP processes. mcp>=1.2,<2 exposes one transport per instance: HTTP and SSE are therefore two instances built by the same function (build_server), which guarantees an identical tool surface. stdio stays a dedicated process.

  2. Own environment variables. ICYLOOM_CONFIG, ICYLOOM_HOME, ICYLOOM_STATE_DIR. The historical names from other projects are not reused: the project is independent.

  3. Ports 8150-8153 strictly reserved, bound to 127.0.0.1; tests run on 8154-8159.

  4. One transaction per request. Cache, LLM journal and ledger are written in one transaction: three successive transactions cost about 20 ms per request, more than the gain from the cache.

  5. Circuit breaker really re-armed: after reset_seconds the state becomes half-open and one probe is allowed.

  6. Upstream failover: a failing upstream no longer permanently excludes the others; the priority is never mutated during a call.

  7. Compression off by default, strict profile, never lossy implicitly.

  8. Fail-closed sandbox: without an allowed directory, no command is run.

  9. Tool cache: deny by default, with a per-server state epoch (see The tool cache).

  10. No git repository for now: the licence is still undecided, so the project is not published. The safety net is a folder copy.

Cache hash compatibility

The normalisation engine follows the reference semantics (NFKD, lowercase, Aho-Corasick rules, number-words, diacritics, sentence punctuation). Operators are kept (12+2 is not 1+22), and volatile fields (stream, user, metadata) are excluded from the key.

Measured on the historical cache (read-only): 70 % of normalised texts coincide. The remaining 30 % were written by an earlier version of the engine that removed :, /, +, * and did not convert numbers to words. Those entries are already incompatible with the reference engine; honouring them would break compatibility with the current engine. tests/test_normalize.py measures this agreement and checks the equivalence between the fast engine and an independent naive implementation.

Tests

.\.venv\Scripts\python.exe -m pytest tests\ -q                                          # 223 tests collected
.\.venv\Scripts\python.exe tests\proof_e2e.py                              # end-to-end proof
.\.venv\Scripts\python.exe tests\test_cache_correctness.py                 # stale-hit measurement (prints a report)
.\.venv\Scripts\python.exe tests\real_mcp_fetch.py                         # real MCP server; needs uvx and Internet
  • proof_e2e.py starts a fake upstream on 8154, launches the stack on 8150-8153, checks the 4 services, verifies the tool list over HTTP and SSE, then measures latency over 200 requests (direct / MISS / HIT) and the ledger gain. Its "50 %" gain is by construction (every request is sent twice).

  • test_cache_correctness.py runs a stateful fake tool server under interleaved reads and writes and compares every cached answer with the server's true state. Run it as a script to get the numbers, or through pytest for strict thresholds (zero stale hits).

  • real_mcp_fetch.py drives the real mcp-server-fetch through the relay: schema capture, MISS then HIT, integer canonicalisation, no merge across different URLs, a real error never replayed.

  • Tests use a temporary database and never touch var\icyloom.db. Set ICYLOOM_TEST_PORT to run the suite while a stack occupies 8154.

Layout

D:\ICY-LOOM\
├── icyloom.yaml              single configuration
├── pyproject.toml            package (requires-python >=3.11,<3.14)
├── LICENSE                   MIT
├── README.md
├── PROGRESS.md               what is proven, what is not, what is next
├── MERGE_PLAN.md             the merge plan (IcyCache + Loom -> IcyLoom)
├── RESEARCH.md               the research notes behind the choices
├── ICYLOOM_PROMPT.md         the original brief
├── docs\
│   ├── INSTALL.md            installation, configuration, start, troubleshooting
│   ├── TRAY.md               the system tray icon, end to end
│   ├── logo.png              the mark (source: src\icyloom\static\logo.svg)
│   └── screenshots\          the eight console views
├── data\
│   ├── normalization_rules.json   1122 rules (data)
│   └── tools_registry.json        100 anchors (data)
├── src\icyloom\
│   ├── cli.py            doctor / paths / version / config / serve / ...
│   ├── config.py         Pydantic models of the single configuration
│   ├── paths.py          path resolution (no hard-coded path)
│   ├── logsetup.py       logs on stderr, UTF-8 forced
│   ├── db.py             single database: schema, WAL, additive migrations
│   ├── normalize.py      normalisation + cache key
│   ├── cache.py          LLM cache (TTL, LRU)
│   ├── upstream.py       multi-upstream routing
│   ├── pool.py           HTTP pool: drain -> recycle -> retry
│   ├── breaker.py        circuit breaker with real re-arming
│   ├── gateway.py        OpenAI/Anthropic HTTP gateway
│   ├── admin.py          admin API + metrics + console + /assets/logo.svg
│   ├── dashboard.py      metrics aggregate (read-only snapshot)
│   ├── static\           web console (HTML + SVG, self-contained) + logo.svg
│   ├── mcpserver.py      single FastMCP instance + 13 tools
│   ├── toolproxy.py      tool relay + cache (deny-by-default, epoch) + loop guard
│   ├── ledger.py         single economic journal
│   ├── tokenizer.py      local token counter (tiktoken or estimator)
│   ├── migrate.py        additive migration (sources read-only)
│   ├── runner.py         combined launcher
│   ├── tray.py           system tray icon (Windows / macOS / Linux)
│   ├── autostart.py      start-with-session (registry / XDG / LaunchAgent)
│   └── context\          drawer, anchors, invariants, verification, sandbox...
├── tools\
│   ├── seed_demo.py      demo data set (real traffic, isolated stack)
│   ├── spread_demo.py    deterministic time-spreading of the demo events
│   ├── campaign.py       hermetic functional campaign (88 checks)
│   ├── check_js.py       JavaScript syntax check of the console
│   ├── fake_mcp_server.py  fake downstream MCP server (relay target)
│   ├── run_fake_upstream.py fake OpenAI/Anthropic upstream
│   └── proof_tools.py, report_live.py, report_tools.py, diag_norm.py, ...
├── scripts\
│   └── install-windows.ps1  one-shot install: venv, package, tray, clients
└── tests\                 test suite, end-to-end proof, correctness harness, real-server test

Project map

Everything points at everything else:

Document

What it answers

README.md

What IcyLoom is, how to run it, what it measures

docs/INSTALL.md

How to install it, configure it, start it, wire the MCP clients

docs/TRAY.md

How the system tray icon behaves, and how to make it start with your session

PROGRESS.md

What is proven, what is not, what is next

MERGE_PLAN.md

How IcyCache and Loom were merged into one project

RESEARCH.md

The research behind the technical choices

ICYLOOM_PROMPT.md

The original brief this project answers

LICENSE

MIT

docs/logo.png

The mark (source: src/icyloom/static/logo.svg)

docs/screenshots/

The eight console views, 1440 × 960

The console is the living version of this README: it reads the same v_token_ledger the ledger section describes, through the same GET /api/dashboard route.

License

MIT — see LICENSE.

Copyright (c) 2026 Malik Belkacemi.

Ownership and license

Project origin and evidence · License


Original project and documentation by Malik Belkacemi · MLKBLK-IT · © 2026

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      MCPGate aggregates multiple MCP servers into a single unified endpoint, enabling centralized tool management with granular filtering, automatic namespacing, and observability. Features a real-time web dashboard and optional PostgreSQL-backed audit trails for monitoring and controlling AI tool access across local and remote deployments.
      Apache 2.0
    • A
      license
      A
      quality
      B
      maintenance
      Semantic tool-routing gateway for MCP servers that cuts 70-90% of context tokens, with RBAC, API-key auth, response caching, hot-reload config, and a real-time observability dashboard.
      3
      13
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Self-hosted MCP gateway that applies deterministic, compiled policy to tool discovery, invocation, and outbound data flow, with no model in the enforcement path. Every decision emits a hash-chained receipt sealed with Ed25519 and verifiable using public keys only.
      Apache 2.0