RootSign

Tamper-evident provenance logging for production AI agents.

What is RootSign?

RootSign is a Providex AI product — the agent capture layer of the Providex AI Agent Accountability Platform.

When AI agents take actions in production — calling tools, hitting APIs, writing to databases — there is no built-in audit trail. If something goes wrong (a wrong refund, a leaked PII record, a malformed deployment), there is no way to prove what the agent did, in what order, on whose authorization, or whether the record has been tampered with after the fact.

RootSign solves this. Each agent action is captured as an Action record containing a SHA-256 hash of the previous action — a cryptographic hash chain that makes the record tamper-evident. Modify any record after the fact and rootsign verify detects it.

Compliance-grade audit trails. Zero changes to your agent code.

Status

v0.2.0. pip install rootsign → a verified hash chain in under a minute: no Docker, no database, no plumbing. LangGraph + CrewAI integrations, a framework-agnostic MCP proxy, rootsign verify CLI, PII redaction, human-in-the-loop checkpoints, opt-in decision capture (PRD-19 / ADR-008), and opt-in SDK micro-batching are all shipping.

Phase	Scope	Status
0	Data model + storage + ingest handler	✅ Complete
1	Python SDK — @rootsign.trace, LangGraph + CrewAI + MCP proxy, rootsign verify CLI, redaction, HiTL checkpoint, decision capture, micro-batching	✅ v0.1.5
1.5	Zero-dependency onboarding — JSONL default backend, rootsign.init() facade	✅ v0.2.0
2	Hosted ingest backend + compliance dashboard	Planned
3	Policy enforcement + incident workflow	Planned
4	Cross-platform governance	Planned

Quickstart

pip install rootsign

No extras, no database, no Docker. Three RootSign calls around your ordinary agent code:

import asyncio, rootsign

rootsign.init(agent="invoice-agent", risk_tier="high")        # 1. once, at startup

@rootsign.trace()                                            # 2. per tool
async def send_invoice(customer_id: str, amount: float) -> str:
    return "sent"

@rootsign.trace()
async def log_payment(customer_id: str, amount: float) -> str:
    return "logged"

async def main():
    async with rootsign.session(objective="invoice ACME") as ctx:   # 3. per run
        await send_invoice("acme-corp", 1500.00)
        await log_payment("acme-corp", 1500.00)
    print(ctx.session_id)

asyncio.run(main())

Then verify the chain — the session lives in ~/.rootsign/sessions/<session_id>.jsonl:

$ rootsign verify --local ~/.rootsign/sessions/f758a636-7bcd-4f96-8940-eff7d80e760a.jsonl
VALID ✓  —  2 records, chain intact
  Session:  f758a636-7bcd-4f96-8940-eff7d80e760a

Exit code is 0 for VALID and 1 for TAMPERED, so this drops into CI or a cron audit. Change any character in any record and the verifier names the broken link:

$ rootsign verify --local ~/.rootsign/sessions/f758a636-7bcd-4f96-8940-eff7d80e760a.jsonl
TAMPERED ✗  —  chain broken at record #1
  Detail:   self_hash mismatch
  Session:  f758a636-7bcd-4f96-8940-eff7d80e760a
WARNING: This session log may have been tampered with.

That's the whole surface: init() → session() → @trace / wrap_tools() → verify. init() is synchronous and does no I/O, so it's safe at module scope and inside a running event loop (notebooks, FastAPI startup); the agent record is get-or-created on the first session() entry, keyed on (name, environment). Re-running your script never re-registers.

A runnable version of the above is examples/quickstart-jsonl.

Framework integrations

The three RootSign calls don't change — you only swap in the wrapper for your framework's tool list.

pip install rootsign[langgraph]     # or [crewai], or [mcp]

import rootsign
from langchain_core.tools import tool
from langgraph.prebuilt import ToolNode

@tool
def send_invoice(customer_id: str, amount: float) -> str:
    """Send an invoice to a customer."""
    return "sent"

rootsign.init(agent="invoice-agent", risk_tier="high", framework="langgraph")

async def run_graph():
    async with rootsign.session(objective="invoice ACME") as ctx:
        tool_node = ToolNode(rootsign.wrap_tools([send_invoice]))
        # ...build and run your graph as normal

Python 3.11 or 3.12 recommended. RootSign itself supports 3.11+, but the [crewai] extra currently lags on 3.13/3.14 wheels. If you hit No matching distribution found for crewai, switch to Python 3.12 and reinstall.

See docs/framework-support.md for the version matrix and integration notes. A full runnable LangGraph example (ReAct agent, three instrumented tools, OpenAI-backed) lives in examples/langgraph-invoice-agent.

Production backend (PostgreSQL / TimescaleDB)

The default JSONL backend is a single-process, append-only writer — right for local development, evaluation, and single-process jobs. Switch to Postgres when you outgrow it:

Switch to postgres when you need	Why JSONL can't
Multiple writer processes on one audit trail	No cross-process file locking (ADR-011) — concurrent writers are out of contract
Cross-process human-in-the-loop — rootsign approve from another terminal, or a web UI	Needs a shared store the poll loop can read; JSONL HiTL is an inline TTY prompt only
Queries across sessions — "every action this agent took last week"	JSONL is one file per session, no index
The Phase 2 hosted dashboard	Reads from the store, not from laptops

pip install 'rootsign[postgres]'
rootsign-admin start-db   # docker run timescale/timescaledb:latest-pg16
rootsign-admin init       # alembic upgrade head
export ROOTSIGN_BACKEND=postgres

start-db wraps a single docker run so you don't need to clone the repo. If you have cloned it, docker-compose up -d db is the equivalent developer path. Both reuse the same rootsign-timescaledb container name and rootsign_pgdata volume — pick either, not both.

Your application code does not change. The same init() / session() / wrap_tools() above now writes to Postgres, and sessions are verified by id instead of by path:

$ rootsign verify 660e8400-e29b-41d4-a716-446655440001
VALID ✓  —  3 records, chain intact
  Session:  660e8400-e29b-41d4-a716-446655440001

Advanced: the explicit API

init() is a convenience over the real seams, which stay public, tested, and documented. Use them when one process drives several agents, or when you want to own the DB session and its transaction:

import rootsign
from rootsign import LocalIngestClient, register_agent
from rootsign.database import AsyncSessionLocal

agent = await register_agent(
    name="my-invoice-agent", owner="platform-team",
    environment="production", risk_tier="high", framework="langgraph",
)

async with AsyncSessionLocal() as db:
    client = LocalIngestClient(db=db)
    async with rootsign.session(agent_id=agent.agent_id, client=client) as ctx:
        tools = rootsign.wrap_tools([send_invoice], ctx=ctx, client=client)
        # ...run your graph
    await db.commit()          # the caller owns the commit on this path

Explicit arguments always win over the ambient session — mixing the two is safe, and passing ctx=/client= never consults the implicit context. See ADR-012.

Decision capture (opt-in)

Record the why before each tool call — foundational for Phase 2 session replay. Off by default; opt in deliberately with ROOTSIGN_CAPTURE_DECISIONS=true.

import os
os.environ["ROOTSIGN_CAPTURE_DECISIONS"] = "true"

async with rootsign.session(objective="invoice ACME") as ctx:
    # Record what the agent decided before calling the tool.
    await ctx.record_decision(
        selected_action="send_invoice",
        reasoning_summary="Amount within policy; recipient verified.",
        confidence=0.97,
    )
    tools = rootsign.wrap_tools([send_invoice])
    await tools[0].ainvoke({"customer_id": "acme", "amount": 1500.0})
    # The Action record now carries decision_id linking it to the reasoning above.

Depth controls how much reasoning is persisted, via ROOTSIGN_REASONING_DEPTH:

Value	What's stored
minimal	selected_action + confidence only
summary (default)	+ reasoning_summary truncated to 500 chars
full	+ reasoning_summary truncated to 10,000 chars + alternatives_considered

Calling ctx.record_decision() when the flag is off is a silent no-op — safe to ship in capture-on and capture-off environments without conditionals at the call site. One Decision links to one Action; the pending slot is single and cleared after the next tool call consumes it. Decisions are not in the hash chain (ADR-008) — verify_chain is unchanged.

Quickstart — CrewAI

CrewAI integration is the same shape — wrap the tool list at construction time.

pip install rootsign[crewai]

import rootsign
from crewai import Agent
from crewai.tools import tool

@tool("send_invoice")
def send_invoice(customer_id: str, amount: float) -> str:
    """Send an invoice to a customer."""
    return "sent"

rootsign.init(agent="invoice-crew", risk_tier="high", framework="crewai")

async def run_crew():
    async with rootsign.session(objective="send invoices") as ctx:
        agent = Agent(
            role="Invoicing assistant",
            goal="Send invoices",
            tools=rootsign.wrap_crewai_tools([send_invoice]),
        )
        # ...run your crew as normal

Tested against CrewAI 0.28, 0.40, and 1.x (see CI matrix).

Quickstart — MCP proxy (any framework)

Instead of a per-framework adapter, RootSign can intercept at the Model Context Protocol layer. Point your agent's MCP client at the RootSign proxy and every tools/call becomes a tamper-evident ACTION_RECORD — any MCP-compatible agent is instrumented with zero framework code.

pip install rootsign[mcp]

import rootsign
import uvicorn
from rootsign.mcp.proxy import create_proxy_app

rootsign.init(agent="mcp-proxied-agent", risk_tier="high")

async def serve_proxy():
    async with rootsign.session(objective="proxy MCP tool calls"):
        app = create_proxy_app(
            upstream_url="http://your-mcp-server:8001/mcp",
            # require_approval=True  # gate every proxied call on human approval
        )
        # A uvicorn-compatible ASGI app. Point the agent's MCP_SERVER_URL
        # here; tools/call is recorded and forwarded, other methods
        # (initialize, tools/list, …) pass through unchanged.
        await uvicorn.Server(uvicorn.Config(app, host="0.0.0.0", port=8000)).serve()

require_approval=True gates every proxied tool call on a human decision — the same HiTL flow as @rootsign.trace, pausing before the call reaches the upstream server. See ADR-010.

RootSign can also run as an MCP server — exposing the audit log itself as a read-only data source so an "auditor agent" can list sessions, pull a session's hash chain, verify integrity, and read approval records in-context:

from rootsign.mcp.server import create_server_app

app = create_server_app()   # ASGI app; mounts the MCP server at /mcp
# uvicorn rootsign.mcp.server:app --port 8001

Four read-only tools (list_sessions, query_session_chain, verify_session_chain, get_approval_records) over the existing store — no new tables.

Human-in-the-loop checkpoint

High-risk actions can be gated on a human decision. Pass require_approval=True to @rootsign.trace and the SDK blocks the tool from running until someone approves it via the CLI.

import rootsign

@rootsign.trace(
    require_approval=True,
    timeout_seconds=300,   # 5 minutes
)
async def wire_transfer(account: str, amount: float) -> str:
    # This runs ONLY after a human approves.
    return execute_transfer(account, amount)

When wire_transfer(...) is called, the SDK inserts an ACTION_RECORD with authorization_status='pending' and waits. An operator approves (or rejects) from another terminal:

$ rootsign approve --list
Pending approvals (1):
  <action-id>  wire_transfer  session=<session-id>  submitted=<timestamp>

$ rootsign approve <action-id> --reason "Verified with customer"
✓  Action <action-id> approved.

The decorated function returns normally. Rejection (--reject) raises HiTLRejectedError; a 5-minute timeout raises HiTLTimeoutError and the action's authorization status becomes 'timed_out' (a terminal forensic state distinct from 'human_rejected').

The cross-process flow above needs the Postgres backend — rootsign approve runs in a different process than your agent. On the default JSONL backend, require_approval=True prompts inline on the terminal instead; a headless run raises HiTLUnsupportedBackendError on the tool's first call, before any work happens, naming the fix.

See ADR-007 for the design rationale (poll loop, timeout semantics, race tolerance).

PII redaction

RedactionConfig runs before hashing, so stored input_hash / output_hash values carry no PII signal. Three ready-to-use configs:

from rootsign import StandardPIIConfig, FinancialPIIConfig, HealthcarePIIConfig

# Standard: email, phone, US SSN, credit card, UK NI number
redaction = StandardPIIConfig()

tools = rootsign.wrap_tools([send_invoice], redaction_config=redaction)

FinancialPIIConfig adds account / routing / IBAN patterns; HealthcarePIIConfig adds MRN / NPI / DOB. Each accepts extra_rules={...} for domain-specific patterns without subclassing. See ADR-006.

Micro-batching (opt-in)

BufferedIngestClient wraps any ingest client and buffers ACTION_RECORDs in memory, flushing asynchronously — so a long, tool-heavy pipeline doesn't pay a per-call ingest round-trip. Enable it with ROOTSIGN_BUFFERED=true (the factory wraps the transport for you), or wrap explicitly:

from rootsign import BufferedIngestClient, LocalIngestClient

async with BufferedIngestClient(LocalIngestClient(db=db)) as client:
    async with rootsign.session(agent_id=agent_id, client=client) as ctx:
        ...  # session() flushes the buffer before SESSION_CLOSE

# ROOTSIGN_BUFFERED=true also applies to the facade path — `init()` never
# wraps the transport implicitly (ADR-012), so buffering stays a deliberate
# opt-in.

Only auto-authorized actions are buffered; HiTL, decision, and session records pass through synchronously, so approvals and hash-chain ordering are never deferred. See ADR-009.

Performance

Instrumentation overhead is designed to be negligible. The LangGraph tracer's per-call overhead is benchmarked over 1,000 instrumented tool calls against a mock ingest client — isolating interception cost from the datastore:

Metric	Per-call overhead
p99	~0.3 ms
mean	~0.23 ms
median	~0.23 ms

That is ~15× under the 5 ms p99 budget enforced by the regression test test_p99_overhead_under_5ms (ADR-004). Reproduce it yourself — no database required:

ROOTSIGN_SKIP_DB_BOOTSTRAP=1 python -m pytest \
    tests/performance/test_langgraph_benchmarks.py -m benchmark -s

The -m benchmark marker keeps the performance suite opt-in. Run it without --cov: coverage instrumentation roughly doubles the measured overhead and would not reflect production numbers. Figures above are indicative (dev laptop, Python 3.12); your absolute numbers will vary, but the budget assertion runs in CI-representative conditions.

Architecture

• @rootsign.trace wraps a tool callable and emits an ACTION_RECORD envelope per call. LangGraph BaseTool and CrewAI tools are detected automatically.
• MCP proxy — create_proxy_app intercepts MCP tools/call at the protocol layer, so any MCP-compatible agent is instrumented without a framework adapter (ADR-010).
• rootsign.init() stores config with no I/O; rootsign.session() resolves the backend lazily on first entry and publishes (ctx, client) in a ContextVar, which is how wrap_tools / @trace / the MCP proxy find them without arguments (ADR-012).
• JsonlIngestClient is the default transport: append-only JSONL under ~/.rootsign, no dependencies (ADR-011). LocalIngestClient is the Postgres in-process path; a HttpIngestClient for the hosted backend lands in Phase 2. BufferedIngestClient optionally wraps any of them for async micro-batching (ADR-009).
• Hash chain is per-session: each Action carries prev_action_hash so reconstructing the chain detects any after-the-fact modification.
• HiTLCheckpoint is an async poll loop that opens its own DB session per cycle — see ADR-007 for the loop-binding rationale.
• Storage is either the JSONL writer (default, zero-dependency) or PostgreSQL 16 + TimescaleDB 2.14, where the actions table is a hypertable and the chain stays intact across chunks. Both use the same frozen canonical hash formula (ADR-001) — rootsign verify gives the same verdict either way.

What's next

• Phase 2 cloud backend — HttpIngestClient + hosted compliance dashboard. Drop-in replacement for LocalIngestClient; BufferedIngestClient already removes the per-call round-trip latency it would otherwise add.
• Web UI for HiTL — approve/reject pending actions from a browser instead of the CLI.
• AutoGen integration — same duck-typing shape as CrewAI.

Watch the GitHub Issues for the active roadmap.

Contributing

We welcome contributions. See CONTRIBUTING.md for development setup, coding standards, and the PR process. By submitting a contribution, you agree to the CLA.

Have a question, an idea, or feedback from using RootSign? Start a thread in GitHub Discussions — that's the best place for design feedback, use-case questions, and feature ideas. For reproducible bugs and concrete feature requests, open a GitHub Issue.

License

Apache License 2.0 — see LICENSE and NOTICE.

Security

To report a vulnerability, see SECURITY.md. Do not open a public GitHub issue.