Skip to content

Abe™ · free · open source · Act, Block, Escalate

Let your agent act on its own, without the 2 a.m. surprise.

Put Abe before your agent acts. Most agent actions don't need AI judgment. Abe handles the obvious ones locally for free. When Abe can't decide safely, Flow can.

Abe checks a proposed action against your policy and returns ACT, BLOCK or ESCALATE, with an immutable record of why. It runs entirely on your machine. No account, no API key, no network, no model.

Open source on GitHub · Read the spec

Abe, a friendly traffic light giving a thumbs up
< 1 msMedian evaluation, no external calls.
0Network calls, accounts or keys required. No call-home.
L3Every record passes FJP-CONF v0.1. 160 conformance checks, offline.

Before and after Abe

Stop paying AI to approve the obvious.

Without a gate
  • Every action either runs unchecked or goes to a model for a yes or no
  • You pay for judgment on calls a rule could make
  • No record of why the agent acted
With Abe
  • Obvious calls decided on your machine in under 1 ms, free
  • Only escalations reach Flow: about $5 per 1,000 actions if 1 in 10 escalates
  • A tamper-evident record of every decision, FJP-CONF Level 3

Three answers. Nothing else.

You built the agent. Don't build the judgment system too.

ACT

The policy found no reason to stop or escalate. Execute exactly as evaluated. It does not claim the action is optimal.

BLOCK

A rule clearly forbids it: over a hard limit, blocked vendor, unauthorized agent. Do not execute.

ESCALATE

Rules conflict, evidence is missing, stakes or irreversibility are high, or judgment is required. Hold. Not an error.

BLOCK beats ESCALATE beats ACT. Bad input or a broken rule returns ESCALATE / EVALUATION_FAILURE. Abe never silently acts.

01

Install and run your first check

$ pip install abe-ai    # or: npm install abe-ai
$ abe init              # writes abe-policy.yaml + request.json
$ abe check request.json

Decision: ESCALATE
Reason: FINANCIAL_THRESHOLD_EXCEEDED
Rules: purchase_review_limit
Record: jgr_01M3Q4…
02

Put it in front of the action

PYTHON
from abe import Abe

abe = Abe("abe-policy.yaml")
r = abe.check(
  action={"type": "wire_transfer",
          "amount": 50000},
  context={"agent_id": "finance-agent"})

if r.decision == "ACT":
    execute()
TYPESCRIPT
import { Abe } from "abe-ai";

const abe = new Abe({
  policy: "./abe-policy.yaml" });
const r = await abe.check({
  action: { type: "purchase",
            amount: 12500 },
  context: { agent_id: "procurement-agent" }
});

if (r.decision === "ACT") execute();

Any other language: abe serve runs a local HTTP sidecar on 127.0.0.1 (POST /v1/check), also available as a Docker image. MCP agents: abe mcp exposes fjp_check_action; add --resolver flow (Python) to send escalations to Flow.

03

Write the policy you already have in your head

version: "0.1"
defaults: { unmatched: ESCALATE }          # fail closed
rules:
  - id: purchase_hard_limit
    when: { all: [ { field: action.type,   op: eq, value: purchase },
                   { field: action.amount, op: gt, value: 100000 } ] }
    decision: BLOCK
  - id: purchase_review_limit
    when: { all: [ { field: action.type,   op: eq, value: purchase },
                   { field: action.amount, op: gt, value: 10000 } ] }
    decision: ESCALATE
    reason_code: FINANCIAL_THRESHOLD_EXCEEDED
evidence_requirements:
  - { id: refund_evidence, when: { field: action.type, op: eq, value: refund },
      require: [evidence.customer_verified, evidence.original_transaction] }
judgment_required:
  - action.type: terminate_employee
SectionDoes
rules11 operators (eq gt in contains exists…) with all / any / not
authorizationWhich agent may take which action types → AUTHORIZATION_FAILED
evidence_requirementsRequired facts before acting → EVIDENCE_MISSING with the missing fields
risk · irreversibilityLevels with thresholds. Callers can raise risk, never lower it.
confidenceAgent-supplied confidence minimums, recorded as uncalibrated
judgment_requiredAction classes that always need judgment
04

Keep a record an auditor can verify

Every call returns an FJP Judgment-Grounded Record: the decision, every matched rule, the exact policy hash, the request hash and a SHA-256 record hash, optionally Ed25519-signed with a local key. Records never change. Outcomes and resolutions are appended as linked records.

decisionESCALATE · FINANCIAL_THRESHOLD_EXCEEDED
policy.hashsha256:7152f8e4… (which rules existed when the agent acted)
falsifierAn OUTCOME record reports status executed before a RESOLUTION with decision ACT is linked
record_hashsha256:684f5e90… · signature ed25519 (optional)
$ abe validate-record jgr.json --public-key abe-signing-key.pub.pem
VALID  jgr.json  (hash verified, signature verified; FJP-CONF v0.1 L0-L2)
05

When Abe escalates, escalate to Flow.

Rules handle the obvious. When Abe says ESCALATE, hand that one action to Flow Resolver. ACT and BLOCK never leave your machine. Only escalations are sent, redacted, and only they cost anything: about $0.03 to $0.05 each.

$ pip install abe-flow

from abe_flow import FlowResolver
abe = Abe("abe-policy.yaml",
            resolver=FlowResolver(api_key=os.environ["FLOW_API_KEY"]))

Python CLI, MCP or sidecar (including the Docker image): add --resolver flow and set FLOW_API_KEY. No code.

$ pip install "abe-ai[mcp,flow]"

{"mcpServers": {"abe": {"command": "abe",
  "args": ["mcp", "--policy", "/abs/abe-policy.yaml", "--resolver", "flow"],
  "env": {"FLOW_API_KEY": "fjp_live_..."}}}}
Abe saysFlow ResolverResult
ACT / BLOCKnot calledUnchanged. A resolver can never loosen a BLOCK.
ESCALATE: human approvalnot calledStays with the human.
ESCALATE: anything elsejudgeREACH → ACT · SKIP → BLOCK · WAIT / RESEARCH_FIRST → ESCALATE, with Flow's reason and falsifier linked to Abe's record.
Flow unreachableStays ESCALATE. Abe never fails because Flow is down.
Abe acts, blocks or escalates. Flow Resolver judges the escalations. Abe is useful without Flow. With roughly 1 in 10 actions escalating, Flow costs about $5 per 1,000 agent actions.
06

Prove it conforms

$ abe conformance --bench
160/160 checks passed
RESULT: conforms to FJP-CONF v0.1 Gate profile, Level 3
performance: p50 0.34 ms  p95 0.44 ms

Third-party implementations can run the same published fixtures and claim “FJP Compatible”. Apache-2.0.