forma-sdk 2.1.0

FORMA — AI agent compliance SDK. Zero-config tracking, EU AI Act/DPDP/RBI compliance, cryptographic audit trails.

FORMA SDK

Make non-compliance impossible — in one line of code.

FORMA doesn't just record what your AI did; it blocks non-compliant decisions before they execute. Add enforce=[...] and PII leaks, prompt injections, and policy violations are stopped at runtime — then every decision is cryptographically signed and mapped to RBI, DPDP, EU AI Act, and ISO 42001.

---

Install

pip install forma-sdk

Import name is trustlayer:

import trustlayer as tl

Quickstart — runtime enforcement (zero-touch)

import trustlayer as tl

# One call. Enforcement is process-wide on every LLM/tool call — no decorators,
# no context managers, no edits inside your business logic.
tl.init(
    api_key="tl_live_YOUR_KEY",   # from forma.2bd.net → Settings
    human_sponsor="[e-mail]",
    enforce=["rbi_ml_risk", "dpdp"],   # ← blocks violations BEFORE they run
)

# Your existing code is unchanged.
result = run_underwriting_model(application)

Now, before any LLM or tool call executes:

• An Aadhaar / PAN / card number in the prompt → blocked (DPDP)
• "Ignore previous instructions…" / jailbreaks → blocked
• "Auto-approve without review" → blocked (RBI human-review rule)

A blocked action raises ComplianceViolation, and every decision (allow / warn / block) lands in your signed audit trail. The gate runs locally in ~1 ms — no network hop in your request path.

Multiple agents — tl.register() / agents= (still zero edits)

Map your existing functions to agents without touching their bodies. Each gets its own discrete signed run plus its own enforce / risk_level / compliance / approval config:

import trustlayer as tl
from my_app import approve_loan, validate_kyc, detect_fraud

tl.init(
    api_key="tl_live_YOUR_KEY",
    agents={
        "kyc-validator":  validate_kyc,                                        # bare callable
        "loan-approval":  {"fn": approve_loan, "enforce": ["rbi_ml_risk", "dpdp"],
                           "risk_level": "HIGH"},                              # per-agent config
        "fraud-detector": {"fn": detect_fraud, "enforce": ["dpdp"]},
    },
)

# Equivalent if you register after init:
tl.register("loan-approval", approve_loan, enforce=["rbi_ml_risk"], risk_level="HIGH")

# Call them exactly as before — each is governed and attributed independently.
result = approve_loan(application)

register / agents= work for sync and async def functions and stay correct under 20+ concurrent agents (scope is tracked with contextvars, so threads and asyncio tasks never cross-attribute). They must run before the target is called (top of your entrypoint).

from trustlayer import ComplianceViolation

try:
    result = approve_loan(application)
except ComplianceViolation as e:
    print(e.reason)   # "PII detected in LLM prompt: Aadhaar number. ..."

Zero-config tracking (no enforcement)

import trustlayer as tl
tl.init(api_key="tl_live_YOUR_KEY")

# Every OpenAI / Anthropic / LangChain / LiteLLM call is now auto-captured —
# tokens, cost, latency, anomaly score — with no other code changes.

Human approvals (EU AI Act Article 14 / RBI human review)

import trustlayer as tl

tl.init(
    api_key="tl_live_YOUR_KEY",
    require_approval_when=lambda ctx: ctx.get("amount", 0) > 1_000_000,
    approval_message="Loan above ₹10L requires human review.",
)

# Your existing function runs unchanged — high-stakes calls pause for sign-off.
result = approve_loan(application)

The decision pauses in your FORMA Approvals inbox until a human approves or rejects. Fail-closed: a timeout or unreachable API is treated as a denial — a skipped review never silently passes.

---

Policy packs

Pack	Region	Enforces
dpdp	India	Aadhaar / PAN / card / phone blocked from prompts & tool args; consent-bypass blocked
rbi_ml_risk	India Banking	auto-approval-without-review blocked; fund-disbursal routed to approval
eu_ai_act	EU	human-oversight bypass (Art. 14) & logging suppression (Art. 12) blocked; PII protection
iso42001	Global	concealing AI involvement blocked

---

Manual step control (advanced)

Optional. Auto-capture already records every LLM call. Use this only to add non-LLM steps (DB queries, HTTP calls, custom logic) to the timeline — the one case where with tl.using() + tracker.tool_call() is the right tool.

import openai
import trustlayer as tl

tracker = tl.init(api_key="tl_live_YOUR_KEY", auto_capture=False)

def my_agent(task: str) -> str:
    with tl.using("my-agent"):
        with tracker.llm_call(label="generate", model="gpt-4o", prompt=task) as step:
            resp = openai.chat.completions.create(
                model="gpt-4o",
                messages=[{"role": "user", "content": task}],
            )
            step.prompt_tokens     = resp.usage.prompt_tokens
            step.completion_tokens = resp.usage.completion_tokens
            return resp.choices[0].message.content

# For tools:  with tracker.tool_call("send_email", {"to": addr}): ...

Kill switch

Pass kill_switch=True to tl.init(). When you trigger a kill from the dashboard, the next LLM/tool call raises KillSwitchTriggered in under ~100 ms.

CI/CD compliance gate

# Blocks the deploy (exit 1) if any agent is below the threshold
trustlayer gate --pre-deploy --fail-below 80

---

tl.init() parameters (all 24)

Parameter	Description
api_key	FORMA API key (or FORMA_API_KEY env).
api_url	Backend URL override.
agent_name	Display name for ambient auto-captured runs.
human_sponsor	Accountable human on the audit record.
auto_capture (True)	Auto-patch OpenAI/Anthropic/LiteLLM/LangChain; False = manual llm_call/tool_call.
kill_switch (False)	Start global process-level kill-switch watcher.
gate (True)	Pre-check every auto-captured call through the compliance gate.
compliance	Frameworks for scoring/reports (EU_AI_ACT, DPDP, RBI_MRM, ISO_42001).
ambient_runs (True)	Create real signed runs for auto-captured calls.
enforce	Policy packs blocked locally pre-execution (dpdp, rbi_ml_risk, eu_ai_act, iso42001).
agents	Zero-edit multi-agent map. {"name": fn} or {"name": {"fn": fn, "enforce": [...], "risk_level": "HIGH", ...}}.
purpose	Plain-English purpose stored in the identity passport.
risk_level ("MEDIUM")	LOW | MEDIUM | HIGH | CRITICAL.
authorized_actions	Allow-list of tool names; gate blocks others.
drift_threshold (0.15)	Behavioural drift threshold for the passport.
max_cost_usd	Hard per-run cost cap; over-budget runs marked failed.
require_approval_when	Predicate ctx→bool; True pauses for human approval.
approval_message / approval_title	Text in the Approvals inbox.
approval_timeout (3600)	Seconds to wait for a decision.
approval_poll_interval (5)	Seconds between inbox polls.
approval_via ("forma")	Approval channel.
timeout (10)	HTTP timeout seconds.
max_retries (2)	Transport retry attempts.
fail_closed (False)	Block when the gate is unreachable (default fails open).

tl.register(name, fn, **cfg) / agents={...}

Bind an existing function to an agent with no code edits. register re-points the function's name in its defining module to a governed wrapper (and also works as a decorator). agents={...} on tl.init() does the same for a whole map. Both accept the same per-agent keyword arguments as tl.using() below (unknown keyword → TypeError) and support sync + async def.

tl.register("loan-approval", approve_loan, enforce=["rbi_ml_risk"], risk_level="HIGH")
# decorator form:
@tl.register("loan-approval", enforce=["rbi_ml_risk"])
def approve_loan(application): ...

tl.using() parameters (all 15) — advanced manual scoping

tl.using() is the per-agent scoping primitive that register / agents= are built on. Use it directly only to scope an arbitrary block or to instrument manual tracker.tool_call() / tracker.llm_call() steps. It works as a context manager, a sync decorator, or an async decorator.

Parameter	Description
name	Agent name (positional).
enforce	Policy packs (per-scope override).
compliance	Frameworks for scoring/reports (per-scope override).
risk_level	Risk level (per-scope override).
human_sponsor	Accountable human (per-scope override).
purpose	Plain-English purpose (per-scope override).
authorized_actions	Allow-list of tool names (per-scope override).
max_cost_usd	Per-run cost cap (per-scope override).
require_approval_when	Approval predicate (per-scope override).
approval_message / approval_title	Approvals inbox text (per-scope override).
approval_timeout	Seconds to wait (per-scope override).
approval_poll_interval	Poll interval seconds (per-scope override).
approval_via	Approval channel (per-scope override).
version	Agent version label.

kill_switch, gate, auto_capture, ambient_runs, drift_threshold, and agent_name are tl.init() only — passing them to tl.using() raises TypeError.

---

Node.js / TypeScript

npm install forma-sdk

import * as forma from "forma-sdk";

forma.init({ apiKey: "tl_live_YOUR_KEY", humanSponsor: "[e-mail]" });

const { decision } = await forma.gate("agent-id", {
    actionType: "llm_call",
    prompt: userPrompt,
});

Note: The Node SDK is coming soon (not yet published to npm). Runtime enforcement (enforce) and approvals are Python-only today. The Node package will provide init + gate first; enforcement parity is on the roadmap.

---

Dashboard & docs

• Dashboard: forma.2bd.net
• Developer guide: forma.2bd.net/developer-guide
• Generate a tailored snippet: forma.2bd.net/snippet-generator

License

MIT — see LICENSE.