mcpsonar 0.9.1__tar.gz

mcpsonar-0.9.1/.gitignore

@@ -0,0 +1,19 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+.pytest_cache/
+.ruff_cache/
+
+# Virtual envs
+.venv/
+venv/
+
+# Local config (users copy from config.example.yaml)
+config.yaml
+config.yml
+
+# SQLite databases
+*.db
+*.sqlite
+*.sqlite3

mcpsonar-0.9.1/CHANGELOG.md

@@ -0,0 +1,400 @@
+# Changelog
+
+All notable changes to MCPSonar. Versions follow [SemVer](https://semver.org/); each release
+is a Git tag `vX.Y.Z`, which is also how the GitHub Action is pinned.
+
+## [0.9.1] — 2026-11-04
+
+No change to the code.
+
+### Added
+- Published on PyPI: `pip install mcpsonar`, `uvx mcpsonar`. The release workflow uploads
+  the wheel and sdist it already checked, through trusted publishing (no stored token), from
+  a `pypi` environment that needs the maintainer's approval.
+- `examples/rugpull-demo/`: a local MCP server that changes its tool, and a script that
+  shows MCPSonar catching it.
+
+### Changed
+- README, banner and repository description lead with rug-pull and tool-poisoning
+  detection; uptime and latency monitoring are described after them.
+- Install instructions use PyPI instead of the Git URL.
+
+## [0.9.0] — 2026-10-23
+
+### Added
+- **`mcpsonar demo --scenario NAME`**: deterministic demos told step by step in the
+  terminal: `healthy`, `poisoning`, `rug-pull`, `latency-regression`, `ci-gate`, or `all`.
+  - **Real engine:** the local demo servers are steered before each step (latency, clean or
+    poisoned tools) and the real engine runs on them: the monitor's check pipeline, the
+    `risks` view, `scan` with a lockfile and SARIF.
+  - **Self-checking:** each scenario asserts MCPSonar's reaction; exit 1 if one does not
+    behave, so `--scenario all --auto` is an end-to-end smoke test (run by the test suite).
+  - **Pacing:** Enter between steps on a terminal; `--auto` for no pauses; `--json`.
+  - **`--dashboard`:** the dashboard follows every step live (checks go through the
+    scheduler, like "Check now").
+  - **Guides:** `docs/demo/`: overview, scenarios, speaker guide, troubleshooting.
+- **`mcpsonar demo --servers`**: only the three demo MCP servers, with their URLs and
+  one-line `curl` commands to rug-pull or slow them down. The control endpoint also accepts
+  a query string (`/__demo?tools=poisoned`), so the same line works in every shell.
+- **Onboarding docs:**
+  - `docs/quickstart.md`: install, scan and pin, monitor, rug pull, review, CI gate and
+    doctor in about ten minutes, all local, every command run while writing it;
+  - `docs/faq.md`;
+  - `docs/troubleshooting.md`, including a verified procedure for a forgotten admin
+    password.
+- **`scan --no-color`**, like every other command.
+- **Slack and Discord notifications** (`type: slack` / `discord`): one message per change,
+  not per check.
+  - **What is sent:** down or degraded, back up with how long it was out, and new findings.
+    Episodes from 0.8 mean a slow period is one message.
+  - **Hostile text:** server-controlled text is stripped of invisible characters and
+    escaped: no `<!channel>` or `@everyone`, no disguised links. URLs must be https and
+    are masked.
+  - **Link to the instance:** `MCPSONAR_PUBLIC_URL` adds it to each message.
+- **Email notifications** (`type: email`): the same rules, through your SMTP server.
+  - **TLS:** required (STARTTLS that must succeed, or implicit TLS on 465). Plain SMTP only
+    for a relay on localhost.
+  - **Outbound policy:** checked before connecting; the connection goes only to an allowed
+    address, and the certificate is verified against the host name.
+  - **Password:** masked in every output and restored on edit. It can be a secret
+    reference.
+  - **Tested** against a scripted SMTP server with a certificate generated for the test:
+    STARTTLS, implicit TLS, a server without STARTTLS, an untrusted certificate and a
+    blocked destination.
+- **Dashboard:**
+  - **Incidents in History:** each run of non-up checks with its start, worst status,
+    duration (or "ongoing") and first reason. It comes from the same function as
+    `mcpsonar history`, now shared with the insights API (`"incidents"`).
+  - **Filters:** down, degraded, to review, findings, latency, auth required, paused. Only
+    the filters that match something are shown, and the choice is remembered in the
+    browser.
+  - **When the API can't be reached:** a banner says so, keeps the last data, and retries.
+    A 401 goes to the login page. Before, the page stayed on "Loading…".
+  - **Phones:** no more horizontal scrolling at 375 px (the top bar, tiles and actions
+    wrap).
+  - **Accessibility:**
+    - every text color reaches WCAG AA (4.5:1) in both themes; the light theme's green,
+      amber, teal and faint text were below it;
+    - per-server buttons have accessible names ("History · calculator");
+    - the heartbeat strip has a text equivalent ("Last 40 checks: 37 up, 3 down");
+    - the tool count is a real button.
+- **Accounts and roles:** `viewer`, `operator`, `admin`; admins manage accounts
+  (`/api/users`). The first account and every account from earlier versions is an admin;
+  the last admin can be neither demoted nor deleted. See
+  [Accounts and roles](docs/configuration.md#accounts-and-roles).
+- **Your account:** change your password (your other sessions end), list your sessions
+  (with their address and browser) and end any of them, keep your API tokens private. An
+  admin resetting a password also ends that person's sessions and tokens.
+- **Audit log** (`/api/audit`, admins): logins, accounts, tokens, server changes, imports,
+  exports with secrets and every approval of tool changes. No credentials. It is
+  append-only in the database itself (triggers refuse UPDATE and DELETE).
+- **Dashboard for teams:**
+  - **Per role:** each role sees only the controls it can use. Operators see stdio as
+    "admins only"; viewers get read tokens only. The server still enforces every rule.
+  - **Account:** change your password, see your sessions (browser, address) and sign them
+    out.
+  - **Users** (admins): add accounts, change roles, reset passwords, delete accounts.
+  - **Audit** (admins): the log, newest first, with what each approval covered.
+- **Trust Report**, for one server, several or the whole instance, over N days (default 30):
+  - **Contents:** availability (share of checks answered, incidents, time not up, the
+    longest incident, latency p50/p95); the pinned tool definitions with SHA-256
+    fingerprints; every tool change in the window with its review status and **who
+    approved it, when** (audit log); the lint findings on the current tools, labeled as
+    heuristics.
+  - **Format:** one self-contained page (no script, no external resource), styled for
+    printing to PDF. It gives no score and no verdict: it records observations.
+  - **Where:** `GET /api/trust-report` (html or json, any role), the dashboard (top bar,
+    History), and `mcpsonar trust-report` (renders the instance's JSON with the same code).
+- **Schema version 4:** user roles, session ids, the audit log. It is migrated
+  automatically, with a backup.
+- **Recovery for the generic webhook:** `on_recovery` (default on) sends one POST when a
+  server is back up from an `on` status. `CheckEvent` gains `previous_status` and
+  `previous_status_since` (live events only).
+
+### Changed
+- **Security: stricter API tokens.**
+  - **Lower of scope and role:** a token now acts with the lower of its scope and its
+    owner's current role, and never as an admin.
+  - **Owner gone:** a token whose owner no longer exists stops working.
+  - **No longer possible with a write token:** adding or changing a stdio server's process,
+    and exporting with secrets. These are admin actions and need a dashboard session.
+  - **Visibility:** tokens are listed and revoked by their owner (admins see all).
+- **The way back after a latency shift is found sooner.** After a reported shift, the next
+  shift is searched relative to the new level. Before, a short regression followed by a fix
+  could hide the return to normal, because the strongest split of the window was the
+  already-reported first change.
+- **Quieter new databases:** creating a new database no longer logs each migration step at
+  INFO; upgrading an existing one still does.
+
+## [0.8.1] — 2026-10-22
+
+### Fixed
+- **Dashboard stuck on "Loading…" after a latency shift.** Since 0.8.0 a `latency_shift`
+  carries `since`; `/api/status` failed (500) as soon as any server's last anomaly was a
+  shift, and the whole dashboard stopped loading. Upgrade if you run 0.8.0.
+- **Heartbeat strip of new servers looked empty:** the padding slots shared a CSS class with
+  the empty-state box, so servers with fewer than 40 checks showed no bars.
+- **The "Anomalies" tile counts each episode once:** checks marked `ongoing` no longer add
+  to it (they still show in the strip, with "(ongoing)" in the tooltip).
+- History is ordered deterministically when two checks share a timestamp (coarse clocks
+  on Windows): insertion order wins. A test depending on it failed intermittently on the
+  Windows CI runner.
+
+## [0.8.0] — 2026-10-21
+
+### Observability
+- **Where the time goes.** Every check now records:
+  - `timings_ms`: `handshake`, `discovery` (every `tools/list` page), `probes` and `total`
+    (connection set-up included);
+  - `tool_latency_ms`: the latency of each probed tool call.
+- **Unchanged meaning:** `latency_ms` is still the handshake, so baselines, anomaly detection
+  and dashboards behave as before. The new fields are additive in the API, webhooks and the
+  database, through an automatic migration.
+- **Prometheus:** new `mcpsonar_check_duration_ms`, `mcpsonar_phase_latency_ms{phase}` and
+  `mcpsonar_tool_latency_ms{tool}`. No existing metric was renamed.
+- **CLI:** `mcpsonar history` shows the median of each phase and of each probed tool.
+- **Response drift.** For each probed tool, MCPSonar remembers the *shape* of the answer
+  (error or not, content types, structure of `structuredContent`, order of magnitude of the
+  size) and reports a `response_drift` anomaly once when it changes.
+  - Response content is never stored.
+  - Opt-in `drift: hash` (with `drift_ignore`) also catches value changes on deterministic
+    probes; `drift: off` disables it per probe.
+  - The status is not affected. See [Response drift](docs/configuration.md#response-drift).
+
+### Packaging and developer experience
+- **`mcpsonar completion bash|zsh|fish|powershell`**, generated from the CLI's own parser,
+  so it always matches the installed version; no new dependency.
+- **`doctor` checks the remote setup:** whether `MCPSONAR_TOKEN` is accepted (scope,
+  expiry), the instance version against the CLI's, and it refuses to send the token over
+  plain HTTP to a non-local instance. `/api/auth/status` reports the version only to an
+  authenticated caller, and the validity of a bearer token sent with it.
+- **Checked packages.** A CI `package` job builds the wheel and sdist, checks their metadata
+  and contents (`scripts/check_package.py`, `twine check --strict`), installs the wheel into
+  a clean environment, runs it, and runs it through `uvx`. Releases attach both files.
+- **Metadata:** Documentation, Changelog and Issues URLs; README links are absolute, so they
+  work outside GitHub too; the sdist no longer ships the screenshots (7.3 MB to 241 KB).
+
+### Secret references
+- Header and stdio `env` values accept `${env:NAME}` and `${file:PATH}`, resolved at every
+  check and never stored, exported or shown resolved. `${file:...}` is re-read on every
+  check, so a rotated secret needs no restart.
+- **Allowlisted reads**, because configurations also come from the API and shared files:
+  variables named `MCPSONAR_SECRET_*` or listed in `MCPSONAR_SECRET_ENV`; files inside
+  `/run/secrets` or `MCPSONAR_SECRETS_DIR` (after resolving symlinks and `..`).
+- **Errors:** a malformed or forbidden reference is refused by the API (400) and by config
+  loading; an unresolvable one makes the check `down` with the reason, never a value.
+- **Display:** references are shown as-is in the dashboard, the API and exports, so an
+  export with references is complete and safe to share.
+- See [Secret references](docs/configuration.md#secret-references).
+
+### Storage
+- **Versioned schema.** The database records its schema version (`PRAGMA user_version`)
+  and is migrated forward on start by numbered steps, each in its own transaction.
+  Databases from v0.5, v0.6 and v0.7 are upgraded automatically; tests restore databases
+  created by each released tag and check that the migrated schema equals a new one.
+- **Backup before migrating:** `mcpsonar.db.schema<N>.bak` is written next to the database.
+- **No silent downgrade:** a database written by a newer mcpsonar is refused (exit 2).
+- **WAL mode** with a separate read-only connection: reads no longer wait for writes.
+  Back up with SQLite's online backup, not by copying the `.db` file alone (see
+  [Backups](docs/configuration.md#backups)).
+- **Atomic writes.** Each write runs in one transaction, one writer at a time. Before, a
+  multi-statement write (replacing a tool snapshot) could be committed halfway by another
+  task sharing the connection.
+
+### Anomaly detection
+- **Episodes.** A run of slow checks is one episode: the first `latency_spike` is notified,
+  the following ones carry `ongoing: true` (kept in the history, not notified), and the
+  first normal check reports a new **`latency_recovered`** anomaly. Generic webhooks with
+  `on_anomaly` now fire once per episode plus the recovery, instead of on every slow check.
+  `anomaly.degrade` keeps the server degraded for the whole episode.
+- **Baseline poisoning.** Slow checks are left out of the baseline while they are a
+  minority, so repeated bursts no longer inflate it and hide milder spikes.
+- **Sustained regression vs. burst.** A latency shift now needs most of the new segment
+  to sit at the new level: a burst shorter than 8 checks was reported as a shift before.
+  Once reported, the new level becomes the baseline (with `since`, when it started).
+- **Cold start.** From the 8th check a provisional baseline flags blatant spikes only
+  (≥ 2× and 1.5× the sensitivity); the insights API reports `samples` and `provisional`.
+- **Zero MAD.** When over half the latencies are identical, the mean absolute deviation
+  stands in for the MAD.
+- A reproducible scenario corpus (`tests/corpus/anomaly_scenarios.py`) pins all of this.
+
+### Reports and SARIF
+- `mcpsonar report` for one server or every server in `--config`, in **Markdown** (for an
+  issue or PR comment), **JSON** or **SARIF 2.1.0**. It includes findings with each rule's
+  documentation, lockfile drift (`--lock`, read-only) and unreachable servers. Exit 0/1/2
+  like `scan`.
+- `mcpsonar scan --sarif FILE` and a `sarif` input on the GitHub Action. Upload the file with
+  `github/codeql-action/upload-sarif` and findings become **code scanning alerts**, pointing
+  at the tool's line in the lockfile.
+- What the SARIF carries:
+  - every rule, with its help text, `security-severity`, `precision` (the rule's confidence)
+    and OWASP MCP Top 10 tags;
+  - lockfile drift as the rule `tool-definition-changed`;
+  - stable `partialFingerprints` so alerts follow the same issue across runs.
+  - Checked against the official SARIF 2.1.0 schema.
+- `risks --json` also carries each finding's confidence and OWASP categories (same function
+  as the reports).
+
+### Detection quality
+- New adversarial corpus (`tests/corpus/detection.yaml`), run in CI: positive and negative
+  cases for every lint rule, plus documented known limitations run as strict expected
+  failures.
+- Every rule is documented in [docs/detection-rules.md](docs/detection-rules.md) with its
+  severity, **confidence** (precision), OWASP MCP Top 10 mapping, examples and what to
+  review. The registry (`lint.RULES`) is the single source of each rule's severity.
+- Tuned against the corpus and against the tools of the 24 watched public servers: 8 false
+  positives on 42 live tools became 0, and the corpus has 0 misses and 0 false positives
+  (3 documented limitations).
+  - `prompt-injection` ignores CLI flags and parameters (`--ignore`, `override=true`).
+  - `sensitive-path`: private keys, `.env` files and MCP client configs count only when the
+    text asks to read or send them. "Add this server to your mcp.json" is setup, not an
+    attack.
+  - `exfiltration` needs data (conversation, files, keys, …) heading to the destination, so
+    "Send a POST request to https://api…" no longer matches.
+  - `coercive-instruction` needs "always / first / also": "You must provide a city name" and
+    API ordering notes no longer match.
+  - `hidden-unicode` ignores the zero-width joiner inside emoji sequences.
+  - `fake-system-tag` also catches `[SYSTEM]` / `[IMPORTANT]`.
+  - `encoded-blob` ignores long URLs.
+  - `long-description` threshold raised from 1,500 to 4,000 characters (real descriptions
+    reach ~3,300).
+
+## [0.7.0] — 2026-10-21
+
+### Added
+- CLI commands `tools`, `risks`, `config validate|show|export`, `doctor` and `version`
+  (plus `--version`). They work without a running instance, against a URL, `--stdio` or a
+  server name from a YAML file. Every command supports `--format json`, `NO_COLOR` /
+  `--no-color`, results on stdout and diagnostics on stderr, and stable exit codes (0/1/2).
+  See [docs/cli.md](docs/cli.md).
+- A rule registry (`lint.RULES`): every finding's rule ID now has a title, an explanation
+  and a remediation.
+- **API tokens** for the CLI and automation: created and revoked in the dashboard (new **API
+  tokens** dialog). They are scoped *read* (GET only) or *write*, can expire, are stored
+  hashed and shown once, and can't manage tokens. They are sent as
+  `Authorization: Bearer mcps_…`.
+- Remote-mode CLI commands that work against a running instance (`MCPSONAR_URL`,
+  `MCPSONAR_TOKEN`), through the same API as the dashboard and never through its database:
+  - `status`;
+  - `servers list|show`;
+  - `check NAME|--all`, run by the instance;
+  - `diff NAME`: changes marked pending, SUSPICIOUS or APPROVED; never approves anything;
+  - `history NAME --last 24h`: uptime, latency percentiles, baseline, shift, incidents;
+  - `servers add|edit|pause|resume|remove`. `--header-from-env` keeps credentials out of
+    shell history. `edit` changes only the given options and keeps credentials and probe
+    arguments. `remove` asks for confirmation on a terminal and needs `--yes` otherwise.
+- `GET /api/servers/{id}/tool-changes` includes the lint `findings` of each new definition.
+
+### Protocol
+- **MCP Python SDK 2.x (`mcp>=2.3,<3`) and spec revision 2026-07-28.** One client speaks both
+  protocol eras: `server/discover` for 2026-07-28 servers, falling back to `initialize` for
+  2024-11-05 … 2025-11-25 servers. Servers that only speak 2026-07-28 used to be reported
+  *down*. 5 of the 24 public servers in `examples/popular-servers.yaml` already speak it.
+- New validator for 2026-07-28. `spec_version` reports the negotiated revision.
+- Unchanged:
+  - **Tool fingerprints:** identical to mcp 1.x, with golden values tested and 42/42 tools
+    checked on live servers. Stored baselines and lockfiles are re-normalized, so annotation
+    keys 2.x no longer represents don't count as changes.
+  - **401/403/429 handling:** *auth required*, *credentials rejected*, *rate limited* +
+    `Retry-After`.
+  - **Latency continuity:** HTTP times only the RPC that succeeded; stdio includes process
+    startup, as before. A 401/403/429 on the discover probe ends negotiation at once, so
+    protected servers cost one request per check, as before.
+- `httpx` is replaced by `httpx2`, the stack mcp 2.x uses, so the outbound policy covers one HTTP
+  stack. On Windows and macOS, TLS uses the operating-system trust store, so corporate CAs work
+  without extra configuration.
+- "down" details now include the HTTP status when there was one (e.g. `(HTTP 503)`).
+
+### Changed
+- The dashboard's Tools view, `mcpsonar scan` and the new commands share one inspection
+  service. As a result `scan` now has an overall time budget too.
+- One-shot commands log only warnings; `run` and `demo` are unchanged.
+- `scan` also accepts `--format json`.
+
+### Fixed
+- `--stdio` with quoted paths on Windows (e.g. `"C:\Program Files\nodejs\npx.cmd"`): the
+  quotes were kept as part of the path, so the command could not be found.
+
+## [0.6.0] — 2026-10-21
+
+Security hardening release. Details, evidence and tests for each item:
+[docs/engineering-audit.md](docs/engineering-audit.md). Upgrade notes:
+[docs/configuration.md](docs/configuration.md#upgrading-from-v05).
+
+### Security
+- **Critical:** blocked a drive-by remote command execution. A page on any website could POST
+  YAML to `/api/import` on an `--no-auth` instance (including `mcpsonar demo` and the watch
+  compose) and make MCPSonar launch an arbitrary stdio command. State-changing requests from
+  other sites or origins are now refused, and with `--no-auth`, stdio servers can't be added
+  over the API unless `--allow-stdio-api` is given.
+- `Host` allowlist (`--allowed-host`; loopback-only by default with `--no-auth`) against
+  DNS rebinding.
+- First-admin setup requires a one-time token printed in the log (or
+  `MCPSONAR_SETUP_TOKEN`); concurrent setups can no longer create two admins.
+- Server-side sessions, revoked on logout and stored as hashes; `Secure` cookie over HTTPS
+  (`--cookie-secure`).
+- Login throttling; PBKDF2 raised to 600k iterations and moved off the event loop (old hashes
+  are upgraded at the next login).
+- Credentials (headers, `env`, URL keys and userinfo, webhook and push URLs) are masked in API
+  responses, the dashboard and exports, and scrubbed from check details, webhook and Kuma
+  messages, logs and `scan` output. Editing a server keeps its stored credentials.
+- Outbound policy enforced at connect time: cloud metadata and link-local addresses blocked;
+  `MCPSONAR_BLOCK_PRIVATE_NETWORKS`, `MCPSONAR_OUTBOUND_ALLOW` / `_DENY`. Configured headers are
+  no longer forwarded on cross-origin redirects; at most 5 redirects.
+- Security headers (no framing, CSP, nosniff, no-referrer); 1 MiB body limit; bounded history
+  queries; optional `MCPSONAR_METRICS_TOKEN` for `/metrics`; http(s)-only URLs.
+- The container runs as an unprivileged user with no capabilities and has a HEALTHCHECK.
+- `pyjwt>=2.15` and `certifi>=2026.10.19` (transitive dependencies with known advisories, GHSA-6q48-6c7w-9gw2 for certifi); CI runs `pip-audit`.
+
+### Fixed
+- `tools/list` pagination is followed. Only the first page used to be checked, pinned and
+  linted.
+- A stdio server that never answers can no longer hang its check loop: session read timeouts
+  plus an overall check budget.
+- Probes never call a tool the server doesn't list, and skip tools annotated destructive
+  unless `allow_destructive: true`; new per-probe `timeout_seconds`.
+- Duplicate tool names are reported; reordering a schema's `required` list is no longer a
+  "tool change".
+- Concurrent checks of the same server (Check now + loop) could record tool changes twice.
+- Editing a server in the dashboard dropped its probes' arguments.
+
+### Changed
+- `GET /api/export` is masked by default (`?include_secrets=true` for a full export); import
+  refuses a masked export.
+- Everyone signs in once more after upgrading (sessions moved server-side).
+- CI runs on Linux, Windows and macOS and smoke-tests the Docker image; Dependabot enabled.
+
+## [0.5.0] — 2026-10-14
+
+First tagged release.
+
+### Added
+- **Rug-pull detection**: every server's tool definitions are pinned on first sight; added,
+  removed or changed tools are recorded and (by default) keep the server *degraded* until a
+  human reviews the diff in the dashboard.
+- **Tool-poisoning lint**: hidden Unicode, instruction overrides, fake `<IMPORTANT>` blocks,
+  "don't tell the user", credential paths, exfiltration, tool shadowing, suspicious
+  parameters.
+- **Anomaly detection** with a learned per-server baseline: latency spikes (robust z-score),
+  level shifts (change point) and flapping.
+- `mcpsonar scan`: one-shot scan with a tool **lockfile** for CI, plus a GitHub Action.
+- `mcpsonar demo`: dashboard against four scripted demo servers.
+- `examples/popular-servers.yaml` + `docker-compose.watch.yml`: watch 24 hosted MCP servers.
+- Validators for MCP spec 2024-11-05, 2025-03-26 and 2025-06-18 (besides 2025-11-25).
+
+### Changed
+- OAuth-protected servers (401/403 to an anonymous client) are **up · auth required**
+  instead of down; rejected credentials are down; 429 is degraded (rate limited).
+
+### Fixed
+- Latency was inflated when several servers were checked at once (a TLS context was rebuilt
+  on every check).
+- Pinned `mcp>=1.28,<2`: the 2.x SDK is a breaking rewrite and broke every check on a fresh
+  install.
+- Source files contained literal zero-width and bidi-override characters (in the lint's own
+  regex and two test fixtures). They are now written as Unicode escape sequences, and CI
+  rejects any invisible character in tracked files.
+- ruff pinned to an exact version so CI and local linting agree.
+- Pausing, editing or deleting a server while one of its checks was in flight could hang
+  forever: the MCP SDK can turn a cancellation into an ordinary error, which was recorded as
+  "down" and the check loop kept running. Cancellation is now always propagated.

mcpsonar-0.9.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Alessandro Fraleoni
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.