groundcrew

Deterministic state oracle and semantic action codec for computer-use agents.

Quick Start · How It Works · CLI Reference · GitHub Action · vs. Alternatives · Contributing

---

Why

Computer-use agents act on real software: they write files, call APIs, run scripts. But how do you know what they actually did vs. what they were supposed to do?

Screenshot-based LLM judges give you a visual approximation at best. They miss side effects — the extra file written, the config silently overwritten, the database row changed. And they cannot replay, diff, or audit what happened.

groundcrew inverts the architecture: instead of watching from the outside, it snapshots the filesystem before and after every action and produces a content-addressed ActionReceipt — a tamper-evident record of exactly what changed. No guessing. No LLM judge. Just a deterministic diff.

groundcrew capture --root . --verb write --target config.json --run "agent.py"
# → ActionReceipt: 3 files added, 1 modified, diff stored in .groundcrew/receipts.db

---

How It Works

flowchart LR
    A[Agent declares\nActionSpec\nverb · target · params] --> B[Oracle captures\nStateSnapshot BEFORE\nSHA-256 of file tree]
    B --> C[Agent runs\nthe action]
    C --> D[Oracle captures\nStateSnapshot AFTER]
    D --> E[SnapshotDiff\nadded · removed · modified]
    E --> F[ActionReceipt\nspec + before_id + after_id + diff]
    F --> G[ReceiptStore\nSQLite persistence]

Core primitives:

• FileState — a content-addressed snapshot of a single file: path, size, SHA-256.
• StateSnapshot — a content-addressed snapshot of a directory tree. ID = SHA-256[:16] of sorted file states.
• SnapshotDiff — the structural delta between two snapshots: added, removed, modified files. .added and .removed are list[FileState]; each element has a .path attribute (relative path string). .modified is list[tuple[FileState, FileState]] (before, after).
• ActionSpec — a semantic, content-addressed action description: (verb, target, params). ID = SHA-256[:16] of the spec. The same action on the same target always produces the same ID.
• ActionReceipt — binds an ActionSpec to a before-snapshot ID, after-snapshot ID, and SnapshotDiff. Stored permanently as an audit trail.
• ReceiptStore — SQLite-backed store. Save receipts, retrieve by ID, list history.

Snapshots are computed by walking the directory tree with os.walk, hashing each file with SHA-256, and content-addressing the whole collection. This is purely Python standard-library code — no kernel hooks, no elevated privileges, no platform-specific APIs required.

---

Features

Feature	Details
Content-addressed snapshots	Same file tree always produces the same snapshot ID
Deterministic diffs	Added, removed, and modified files — no approximation
Semantic action codec	ActionSpec is portable, content-addressed, version-robust
Tamper-evident receipts	ActionReceipt binds intent to effect, stored permanently
SQLite receipt store	Single-file persistence, no server, works offline
Rich terminal output	Color diff tables, receipt summaries
JSON output	Machine-readable for downstream automation
Markdown output	Ready-to-paste audit reports
FastAPI REST server	/capture, /receipt/{id}, /receipts, /diff/{id}
MCP server	Model Context Protocol integration for Claude and other agents
91 tests	Comprehensive test suite covering all layers

---

Quick Start

pip install groundcrew            # core library + CLI
pip install "groundcrew[api]"     # + FastAPI REST server
pip install "groundcrew[mcp]"     # + MCP server

import json, pathlib, tempfile
from groundcrew import Oracle, ActionSpec, ReceiptStore

# Use a temporary directory so nothing is left behind in your cwd
with tempfile.TemporaryDirectory() as tmpdir:
    # Declare what you're about to do
    spec = ActionSpec(verb="write", target="config.json", params={"key": "value"})

    # Capture before/after state around the action
    with Oracle(tmpdir, spec) as oracle:
        pathlib.Path(tmpdir, "config.json").write_text(json.dumps({"key": "value"}))

    receipt = oracle.record(spec)
    print(receipt.diff.changed_paths)   # {'config.json'}
    print(receipt.id)                   # content-addressed ID

    # Persist for auditing (db lives inside the temp dir — cleaned up automatically)
    store = ReceiptStore(f"{tmpdir}/.groundcrew/receipts.db")
    store.save(receipt)
    # tmpdir and all its contents are removed when the `with` block exits

---

CLI Reference

groundcrew [--db PATH] COMMAND [OPTIONS]

Command	Description	Key options
capture	Snapshot before/after a shell command	--root DIR, --verb VERB, --target TARGET, --run CMD
diff RECEIPT_ID	Show the SnapshotDiff for a stored receipt	—
log	List all stored receipts	—
status	Show database info	—
watch DIRECTORY	Watch a directory for unexpected changes	--interval SECS, --max-checks N, --allow PATH

Global options:

Option	Default	Env var
--db PATH	.groundcrew/receipts.db	GROUNDCREW_DB

Examples:

# Capture what an agent script does to the current directory
groundcrew capture --root . --verb run --target agent.py --run "python agent.py"

# Show what changed
groundcrew diff abc123de

# List all receipts
groundcrew log

# Status
groundcrew status

---

GitHub Action

Add groundcrew auditing to your CI pipeline:

# .github/workflows/groundcrew.yml
name: groundcrew audit
on: [push, pull_request]

jobs:
  audit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: sandeep-alluru/groundcrew@main
        with:
          root: .
          db: .groundcrew/receipts.db

---

vs. Alternatives

	groundcrew	Screenshot judges	AgentSight	OSWorld verifiers
Verification method	Filesystem diff	Vision LLM	eBPF syscall trace	Per-task custom code
Deterministic	Yes — content-addressed	No — probabilistic	Partial	Yes (per app)
No per-app code	Yes	Yes	Yes	No — 33 apps manually
Production runtime	Yes	Yes	Linux-only	VM/sandbox only
Audit trail	SQLite receipts	None	Log files	None
Action codec	Portable ActionSpec	None	None	None
Open source	MIT	N/A	MIT	Research
Python package	Yes	N/A	No	No

groundcrew is not a replacement for security-layer tools like AgentSight. It is specifically designed for agent developers who need a simple, deterministic record of what their agent changed on disk — suitable for testing, auditing, and CI/CD gating.

---

Claude / MCP integration

groundcrew ships a Model Context Protocol server that lets Claude and other MCP-compatible agents record and query action receipts directly:

# Start the MCP server
groundcrew-mcp

# In your Claude Code project's .claude/settings.json:
{
  "mcpServers": {
    "groundcrew": {
      "command": "groundcrew-mcp"
    }
  }
}

Once connected, Claude can call groundcrew/capture_state, groundcrew/get_receipt, and groundcrew/list_receipts as tools. See docs/mcp.md for the full tool schema.

---

OpenAI integration

groundcrew exposes a FastAPI REST server compatible with OpenAI's function-calling format. The tool definitions are in tools/openai-tools.json and the full API spec is in openapi.yaml.

# Start the REST server
uvicorn groundcrew.api:app --reload

# Pass to Codex CLI or any OpenAI-compatible agent
codex --tools tools/openai-tools.json "Capture what this script does to the filesystem"

Endpoints: GET /health, POST /capture, GET /receipt/{id}, GET /receipts, GET /diff/{id}. See docs/openai.md for details.

---

Case Studies

See how teams are using groundcrew in production:

• Cryptographic Deployment Receipts for SOC2-Compliant AI Agents
• Line-Level Audit Trails for AI-Modified Legal Contracts

---

Repository structure

groundcrew/
├── src/
│   └── groundcrew/
│       ├── snapshot.py       # FileState, StateSnapshot, SnapshotDiff
│       ├── codec.py          # ActionSpec, ActionReceipt (content-addressed)
│       ├── oracle.py         # Oracle context manager, capture(), ReceiptStore
│       ├── report.py         # print_receipt(), print_diff(), to_json(), to_markdown()
│       ├── cli.py            # Click CLI (capture, diff, log, status)
│       ├── api.py            # FastAPI REST server
│       └── mcp_server.py     # MCP server
├── tests/
│   ├── test_snapshot.py      # StateSnapshot, SnapshotDiff unit tests
│   ├── test_codec.py         # ActionSpec, ActionReceipt unit tests
│   ├── test_oracle.py        # Oracle context manager, ReceiptStore tests
│   ├── test_report.py        # Formatter tests
│   ├── test_cli.py           # CLI subprocess integration tests
│   ├── test_cli_runner.py    # Click CliRunner tests
│   └── test_api.py           # FastAPI TestClient tests
├── examples/
│   └── demo.py               # Standalone demo script
├── docs/                     # MkDocs documentation
├── tools/
│   └── openai-tools.json     # OpenAI function-calling tool definitions
├── assets/
│   ├── hero.png              # README hero image
│   └── logo.png              # Project logo
├── action.yml                # GitHub Action
├── openapi.yaml              # OpenAPI 3.1 spec
├── pyproject.toml            # Package metadata + dependencies
└── CONTRIBUTING.md           # Contribution guide

---

GitHub Topics

Suggested topics for discoverability:

ai-agents computer-use state-oracle action-codec filesystem-diff verification observability mcp openai llm-tools audit-trail ci-cd python

---

---

Stay Updated

Subscribe to The Silence Layer — weekly dispatches on production AI infrastructure, new releases, and the failure modes that production AI systems don't surface until it's too late.