gaslight 0.1.0__tar.gz

gaslight-0.1.0/.claude/scheduled_tasks.lock

@@ -0,0 +1 @@
+{"sessionId":"8d47fa29-5d99-44cb-a402-17ae8bf20f62","pid":4392,"procStart":"Thu Aug 20 01:01:16 2026","acquiredAt":1787197490766}

gaslight-0.1.0/.gitignore

@@ -0,0 +1,17 @@
+__pycache__/
+*.py[cod]
+*.egg-info/
+.venv/
+.pytest_cache/
+.ruff_cache/
+dist/
+build/
+*.html
+!tests/**/*.html
+# Secrets and local env — never commit
+.env
+.env.*
+logs/
+# Local hunt-result artifacts — contain real local paths and leaked findings
+# from real third-party servers. Never commit.
+harness/hunt_results*.json

gaslight-0.1.0/.superpowers/sdd/.gitignore

@@ -0,0 +1 @@
+*

gaslight-0.1.0/AGENTS.md

@@ -0,0 +1,134 @@
+# Using gaslight from an AI agent
+
+This file is written for an **AI coding assistant** (Claude Code, Cursor, etc.).
+It tells you how to run `gaslight` correctly, how to read its output, and — most
+importantly — why it is safe to run without worrying about breaking the user's
+code. If a user says *"use gaslight to check my agent,"* follow this.
+
+`gaslight` is a black-box security scanner for MCP-based AI agents. You point it
+at a running MCP server; it attacks the tools it finds and returns a graded
+report. Every finding it marks CONFIRMED is physically proven — never a guess.
+
+## TL;DR
+
+```
+pip install gaslight
+gaslight --json -- <command that launches the MCP server>
+```
+
+Parse the JSON on stdout. `grade` is A–F. `findings[]` with `fired: true` are
+real. Exit code `1` means at least one CONFIRMED issue, `0` means clean.
+
+## When to use it
+
+- Before the user ships or publishes an MCP server or agent.
+- In CI, to catch a tool that changed since it was approved (see `--baseline`).
+- Whenever the user asks *"is my agent / MCP server secure?"*
+
+## How to run it
+
+**Install:** `pip install gaslight` (or run without installing: `uvx gaslight`).
+
+**Point it at a target:**
+```
+gaslight -- npx -y some-mcp-server        # a stdio MCP server
+gaslight --url https://host/mcp           # a remote HTTP+SSE server
+gaslight --json -- python server.py       # machine-readable — USE THIS to parse
+```
+
+**Flags worth knowing:**
+| Flag | Use |
+|---|---|
+| `--json` | Machine-readable report on **stdout** (all human text goes to stderr, so stdout is clean JSON). |
+| `--safe` / `--no-safe` | `--safe` is the default. Keep it. |
+| `--env KEY=VALUE` | Pass a **throwaway/test** credential if the server needs one to boot. Never a production secret. |
+| `--baseline FILE` | CI rug-pull guard: records the tool set on first run, flags any tool that changed on later runs. |
+| `--llm ollama` | Optional free local model for richer output. Never required; never decides a verdict. |
+| `--skip k1,k2` | Skip named attacks (e.g. a raw-SQL tool wired to a real warehouse). |
+
+## What gaslight will NEVER do — the safety contract
+
+You can run it without worrying about the user's system. On the default `--safe`
+setting, gaslight:
+
+- **Never reads the user's source code.** It is black-box — it only sees the
+  tools the server exposes over MCP, exactly like any other client.
+- **Never performs a destructive or irreversible action.** Delete/reset-style
+  tools are probed for *authorization*, but the destructive call is not actually
+  made.
+- **Aims its network probes at a dead local address it controls** (its own
+  listener), not at the internet or the user's infrastructure.
+- **Needs no API key**, and with no model (or a local one) **sends nothing off
+  the machine**.
+- **Only tests the one target you point it at.**
+
+## The ONE thing to be careful about
+
+gaslight sends **real (but benign) attack payloads** to whatever target it is
+pointed at. So:
+
+- Point it at a **local / test / throwaway** instance — **never a production
+  backend**.
+- If the server needs a credential, pass a **throwaway** one via `--env` — never
+  a production secret.
+- Do not aim `--url` at a production endpoint.
+
+If you are unsure whether a target is production, ask the user before running.
+
+## How to read the result
+
+**Exit code:**
+- `0` — clean: no CONFIRMED findings.
+- `1` — at least one CONFIRMED finding. The agent has a proven security issue.
+- `2` — the target could not be started or reached. gaslight prints a
+  plain-language reason (old SDK, needs a credential, wrong Node/Python, bad
+  command). This is not a security verdict — the target never ran.
+
+**JSON shape (`--json`):**
+```json
+{
+  "target": "…",
+  "tool_count": 6,
+  "grade":   { "grade": "F", "fired_count": 2, "total_count": 17, "summary": "…" },
+  "findings":[ { "attack_key": "ssrf-probe", "fired": true, "attempted": true, "reason": "…" } ],
+  "metrics": { "average": 60, "scores": [ { "name": "Network", "score": 0, "band": "red", "breached": true } ] },
+  "surface": [ { "severity": "warn", "category": "unconstrained-field", "tool_name": "run_sql", "message": "…" } ],
+  "llm":     { "active": false, "provider": "scripted", "role": "off", "decides_verdict": false }
+}
+```
+
+**What the fields mean:**
+- `fired: true` → **CONFIRMED**. Physically proven — a token gaslight planted
+  reached its listener, or a protected file came back. Real. Report it and help
+  the user fix it.
+- `attempted: false` → **NOT TESTED**. No tool of the shape this attack needs,
+  or nothing verifiable black-box. This is an honest gap, **not** a pass — don't
+  report it as "secure."
+- `metrics.scores[].score` is 0–100 (or `null` when every check for that metric
+  was N/A). `breached: true` means a confirmed exploit capped that metric.
+- `surface[]` are **static warnings** (schema hygiene) — `info`/`warn` only.
+  They never change the grade and are not CONFIRMED exploits. Surface them as
+  "worth reviewing," not as proven vulnerabilities.
+
+## The confidence model — why you can trust it
+
+gaslight **never** marks something CONFIRMED on a guess or a model's opinion.
+Every CONFIRMED is backed by physical proof. So:
+- Trust a CONFIRMED — it really happened.
+- Trust a clean gauge — those vectors were tested and nothing fired.
+- A clean overall run means *"these attack vectors, tested, found nothing"* — a
+  floor, not a certificate. It is the first security check to run, not the last.
+
+The optional LLM layer only makes probes smarter and explains results. It can
+**never** decide whether an attack succeeded.
+
+## What gaslight does NOT cover — don't overclaim
+
+It is a black-box behavioral scanner of one MCP target's runtime. Out of scope
+**by design** (it cannot see these, and says so rather than faking a pass):
+- Supply-chain / dependency tampering (a source/SCA concern).
+- Internal audit-logging quality.
+- Multi-agent / agent-to-agent attacks (on the roadmap).
+
+When you report results, describe them as *"gaslight tested the running tools
+and found X"* — not as a complete security audit.

gaslight-0.1.0/BRAIN_BRIDGE.md

@@ -0,0 +1,30 @@
+# 🧠🌉 Brain ↔ Code Bridge  (gaslight ⇄ LLM WIKI)
+*How this code repo connects to the strategy brain. Read this first if you're Claude Code (or any builder) working in `gaslight`.*
+
+## The two halves
+- **This repo `gaslight`** (working name for **Iago**) = the **CODE / implementation**.
+- **The strategy brain** = a separate knowledge base at **`~/Documents/LLM WIKI/`**, specifically the Iago project folder:
+  **`~/Documents/LLM WIKI/verticals/career/projects/iago/`**
+  (You are on the same Mac — you can read that path directly, and it's always the freshest version.)
+
+## The rule (who is canonical)
+- **Strategy, product decisions, threat model, positioning, "why", scope, launch → the WIKI is canonical.** Read it before making any product/design decision. When a decision changes, update it *there* (append-only), then reflect it in code here.
+- **Implementation — code, tests, packaging → this REPO is canonical.**
+- The `docs/brain/` folder in this repo holds **synced snapshots** of the key wiki docs so you have them locally/offline. **If a snapshot and the live wiki disagree, the WIKI wins** — re-read the live path above.
+
+## How to read (start order)
+1. **`docs/brain/IAGO_HANDOFF.md`** — the complete standalone build brief. **START HERE.**
+2. **`docs/brain/M2_DESIGN.md`** — the hard-won design decisions: how Iago connects (the two mechanisms: direct tool probe + MCP proxy), why it's generic ("Burp for agents" — challenge behavior, don't read the harness), the model-vs-code verdict, the tiered **connection inputs**, and the scope guardrails (**"paste a URL" is NOT the product**; NOT an "MCP tester"; why-not-DIY).
+3. **`docs/brain/BUILD_PLAN.md`** — the locked V1 spec (Python + `uvx`, MCP-native, deterministic canary, cuts, launch).
+4. **`docs/brain/THREAT_MODEL.md`** — what it secures, the attack classes, V1 scope.
+5. **`docs/brain/AGENT_SECURITY_LANDSCAPE.md`** — the field: top problems now/future + what's black-box testable vs. out of reach.
+6. **`docs/brain/WHY_IAGO_EXISTS.md`** — objection handling (auto-discovery; why not just DIY with an LLM).
+7. **`docs/brain/CASE_STUDY_the01dev.md`** — the first real-world target (Swanand's own product) + test plan.
+
+## The Council (for new features / product decisions — NOT code review)
+Any new **feature or product decision** (changes UX / could become a business decision) → convene **The Council** at `~/Documents/LLM WIKI/council/THE_COUNCIL.md` before building. Run it as a single structured pass (not a heavy multi-agent workflow). **Save the session** to `~/Documents/LLM WIKI/council/sessions/` (or the iago folder) so it isn't lost. Code-quality review is a normal code review, NOT the Council.
+
+## Keeping in sync
+- Product/design decision changes in code → write it back into the wiki (source of truth).
+- Re-sync these `docs/brain/` snapshots from the wiki when they go stale (or just read the live wiki path — it's local).
+- Working name is **gaslight**; final product name (Iago) is not locked — needs a PyPI/GitHub/domain + malware/threat-intel sweep first (Gaslight was killed for being live NK malware).

gaslight-0.1.0/LICENSE

@@ -0,0 +1,190 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the
+      purposes of this License, Derivative Works shall not include works
+      that remain separable from, or merely link (or bind by name) to the
+      interfaces of, the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including the
+      original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing
+      the origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   Copyright 2026 Swanand Kadam
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

gaslight-0.1.0/PKG-INFO

@@ -0,0 +1,215 @@
+Metadata-Version: 2.5
+Name: gaslight
+Version: 0.1.0
+Summary: A Lighthouse score for MCP agents: point it at your agent, get a graded security report — and everything it confirms, it proves.
+Author: Swanand Kadam
+License-Expression: Apache-2.0
+License-File: LICENSE
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Security
+Requires-Python: >=3.10
+Requires-Dist: anthropic>=0.68.0
+Requires-Dist: jinja2>=3.1.0
+Requires-Dist: mcp>=1.9.0
+Requires-Dist: openai>=1.50.0
+Requires-Dist: rich>=13.7.0
+Description-Content-Type: text/markdown
+
+# gaslight
+
+> A Lighthouse score for MCP agents. Point it at your agent, get a graded security report — and everything it confirms, it proves.
+
+gaslight is a black-box security tester for AI agents built on the
+[Model Context Protocol](https://modelcontextprotocol.io). You point it at a
+running MCP server; it connects like any other client, runs a suite of real
+probes against the tools it finds, and hands you a graded, screenshot-ready
+report. It never reads your source, your database, or an API key.
+
+Think Lighthouse, but for agent security: five scores out of 100, a letter
+grade, and a picture of how far a breach could travel — so you can see at a
+glance whether your agent is solid or leaking, and exactly where.
+
+The one thing that sets it apart from every "AI security" scanner that grades
+one model's output with another model: **gaslight never trusts an opinion.**
+A finding is only marked CONFIRMED when something physically happened — a unique
+token you planted arrived at a listener gaslight controls, real out-of-bounds
+data came back, or a call that should have been refused went through. If it says
+confirmed, it's real. Go fix it.
+
+```
+$ gaslight -- npx -y some-mcp-server
+
+🛡  These are safe probes, not real break-ins — testing your guards, not stealing your data.
+
+🧪  Network Egress Abuse (SSRF) — whether a URL tool can be aimed at addresses it should never reach
+🔥  CONFIRMED — fetch_url reached http://127.0.0.1:<sink> (your canary arrived). No destination check.
+🧪  Claim Integrity — whether a tool that promises to be read-only keeps that promise
+🔥  CONFIRMED — create_invoice says "stages for approval; does not issue",
+     but the record it created shows status "issued". Its own read tools contradict its own description.
+
+   Network 42   Filesystem 100   Leakage 88   Authorization 60   Integrity 55
+   Grade: F   ·   blast radius: THIS MACHINE ▓   YOUR NETWORK ▓   DATA LEAVING ░
+```
+
+## Quick start
+
+```
+pip install gaslight
+gaslight -- npx -y your-mcp-server      # point it at your agent
+```
+
+That's it — no API key, no config. gaslight connects like any MCP client,
+attacks the tools it finds, prints a graded report to your terminal, and writes
+a shareable HTML report next to it. Add `--json` for machine-readable output, or
+`--baseline tools.json` to catch a tool that changes on you in CI. What the
+grades mean is under [How to read a report](#how-to-read-a-report); to drive it
+from an AI assistant, see [AGENTS.md](AGENTS.md).
+
+## The score
+
+Every finding rolls up into five gauges, each 0–100, plus one letter grade:
+
+| Gauge | The question it answers |
+|---|---|
+| **Network** | Can the agent be steered into reaching out, or sending your data, over the wire? |
+| **Filesystem** | Do file tools stay in their lane, or can they be walked out of scope? |
+| **Leakage** | Do secrets, keys or tokens slip out during ordinary use? |
+| **Authorization** | Are destructive, consequential or costly actions actually gated? |
+| **Integrity** | Do the agent's instructions and its tools' own promises hold under pressure? |
+
+A **breach never scores green** — a confirmed exploit caps the metric and drops
+the grade. And the **blast radius** only lights up where an attack *physically*
+succeeded: working guards keep the lit area small, so the picture rewards real
+defense instead of just counting capabilities.
+
+## What it tests
+
+Every agent, whatever its domain, is built on the same few primitives: tools
+that move data out, read resources, take consequential actions, and make claims
+about themselves. gaslight aims at the primitives, so the same tests apply to
+a hospital agent and a bank agent alike.
+
+**17 attack modules**, all deterministic, all proven physically — including
+indirect prompt injection → exfiltration, SSRF, path traversal, code execution,
+confused-deputy, tool-metadata poisoning, memory poisoning, destructive-action
+authorization, verbose-error disclosure, **denial-of-wallet** (unbounded, costly
+calls), and the headline: **claim integrity** — it takes the safety promise a
+tool's description makes (the promise the driving model trusts) and checks
+whether it's still true, using the target's own read tools as the verification
+channel.
+
+Plus a zero-call **static surface pass** (schema hygiene, hidden instructions in
+descriptions, tool-shadowing / homoglyph names) that flags red flags without
+firing a shot.
+
+### Coverage, mapped to a standard
+
+gaslight is anchored to the **[OWASP MCP Top 10](https://owasp.org/www-project-mcp-top-10/)**,
+not an ad-hoc list:
+
+- **7 of 10 fully covered** — secret exposure, tool poisoning, command
+  injection, prompt injection, weak authorization, context over-sharing, and
+  shadow-servers/rug-pulls.
+- **2 out of scope by design** — supply-chain/dependency tampering and internal
+  audit logging. A black-box behavioral tool can't see those, and says so rather
+  than faking a pass.
+- **1 partial** — privilege scope-creep (stateful over time).
+
+Stating what it *doesn't* test is the difference between an honest benchmark and
+a scanner that over-credits itself.
+
+## The optional LLM layer
+
+The whole deterministic core needs **no model and no key** — it works out of the
+box. An LLM is an *optional lens* that makes probes smarter and the report
+richer, and there is one hard rule:
+
+> The LLM may aim an attack and explain a result. It may **never** decide whether
+> the attack succeeded.
+
+Every CONFIRMED still comes from a canary reaching our sink — never a model's
+opinion. Turn it on with a key, or with `--llm ollama` for a **free local
+model** (nothing leaves your machine — the right default for a security tool).
+No model configured? You still get the full deterministic run. Every run states
+plainly whether the LLM layer is on, off, and where its boundary is.
+
+## Extra modes
+
+- **Doctor mode** — if the target won't start (built for an older SDK, needs a
+  credential, wrong Node/Python), gaslight reads its startup output and gives
+  you one plain sentence to fix it, instead of a wall of someone else's stack
+  trace.
+- **Rug-pull guard** — `gaslight --baseline tools.json` records your tools on
+  first run and flags any that changed since (a description quietly rewritten
+  into a poisoned payload, a new dangerous parameter). Drop it in CI to catch a
+  tool that turns malicious *after* you approved it.
+
+## How to read a report
+
+- **CONFIRMED** — physically proven. Real. Fix it.
+- **not tested** — no tool of the shape this attack needs, or a claim with no
+  black-box way to verify it. Honest about what it couldn't reach, never
+  silently green.
+- A clean run means *these attack vectors, tested, found nothing* — a floor, not
+  a certificate. It's the first security check to run on your agent, not the last.
+
+## Safety
+
+Probes are harmless by construction: they only ever touch a sink gaslight
+controls, an ordinary system file, or a synthetic canary record — never real
+data. On the default `--safe`, an action with a real irreversible or external
+effect is never triggered, and a soft, description-only signal never fires one.
+When you point it at a target that needs a credential, use a **throwaway/test**
+one via `--env KEY=VALUE` — never production.
+
+## Install
+
+```
+pip install gaslight        # (pre-launch)
+gaslight --help
+```
+
+Point it at any stdio MCP server:
+
+```
+gaslight -- npx -y some-mcp-server
+gaslight --url https://my-agent.example/mcp     # remote (HTTP+SSE)
+gaslight --json -- python my_server.py          # machine-readable, for CI
+```
+
+## Use it from an AI agent (or CI)
+
+gaslight is built to be driven by an AI assistant — just tell yours *"use
+gaslight to check my agent."* It ships an **[AGENTS.md](AGENTS.md)** that tells
+the AI how to run it, how to read the result, and — crucially — the safety
+contract, so it can run without any worry about touching your code:
+
+- `--json` emits a structured report on stdout (findings, grade, metrics); human
+  text goes to stderr, so stdout stays clean JSON to parse.
+- **Exit codes:** `0` clean · `1` a CONFIRMED finding · `2` the target couldn't
+  start (with a plain-language reason).
+- **The boundary:** black-box (never reads your source), `--safe` by default
+  (never fires a destructive action), probes aimed at a local sink it controls,
+  no API key, nothing leaves your machine. See `AGENTS.md` for the full contract.
+
+## Roadmap
+
+v1 tests **MCP-based agents** — one target, through the MCP boundary. The next
+frontier is the **agentic / multi-agent layer**: an adaptive red-team agent that
+reasons about your agent and holds a real multi-turn conversation, plus
+agent-to-agent (A2A) attacks between agents that talk to each other. The
+physical-proof rule holds there too — the attacker gets smarter, the verdict
+stays deterministic.
+
+## Status
+
+Pre-launch. **17 attack modules**, validated against deliberately-vulnerable
+benchmarks and 50+ real, independently-built community MCP servers, with matched
+vulnerable/hardened fixtures and zero false positives on the controlled set. See
+`docs/` for the design specs and `docs/TESTING_STRATEGY.md` for the validation
+record.
+
+## License
+
+Apache-2.0 — see `LICENSE`.