EMILIA Protocol
आधिकारिककिसी नामित मानव से ऑफलाइन-सत्यापन योग्य अनुमोदन आवश्यक है, इससे पहले कि कोई AI एजेंट कोई अपरिवर्तनीय कार्रवाई करे — भुगतान जारी करना, रिकॉर्ड बदलना, डिप्लॉय करना। दो-व्यक्ति नियम, Ed25519 ट्रस्ट रसीदें, IETF-प्रारूपित, Apache-2.0।
EMILIA Protocol MCP के साथ आप क्या कर सकते हैं?
- Gate a protected MCP tool — Ask your assistant to wrap an existing tool like
sendWirewith@emilia-protocol/scanso it refuses calls lacking a valid authorization receipt. - Verify authorization receipts offline — Have your assistant check a Trust Receipt's integrity and authenticity locally using
@emilia-protocol/verify, with no backend or API key required. - Run the payment approval example — Ask your assistant to execute the bundled MCP payment server, demonstrating how a payment is refused without approval and bound to exact amount, currency, vendor, and destination.
- Test the authority freeze — Instruct your assistant to run the emergency authority freeze example, blocking new reservations and preventing older ones from entering after an epoch change.
- Issue a demo receipt — Ask your assistant to generate a sample Trust Receipt via
npx @emilia-protocol/issue demo, fully offline, to explore the protocol's evidence format.
दस्तावेज़
EMILIA Protocol
Let your agent prepare the payment. Control what it can release.
An agent prepares an $82,000 supplier payment. Someone approves it. Then the amount or bank destination changes. The earlier approval must not release the changed payment.
EMILIA Gate checks authority before a protected tool runs. The open protocol makes the resulting evidence independently checkable under the verifier's own trusted keys and rules. Keep your existing identity provider, agent framework, and business systems.
Run the payment example
With Node.js 20.19 or newer and npm installed:
git clone https://github.com/emiliaprotocol/emilia-protocol.git
cd emilia-protocol
npm ci
FAST=1 node examples/mcp/payment-server.mjs
The example refuses a call without approval, binds approval to the amount, currency, vendor, and destination, admits the matching call once, and refuses changed payment details, replay, and forged evidence. See the example and its limits.
This is a local demonstration: generated signing keys, in-memory consumption, and a mock payment tool. It does not perform a real human ceremony or move money. Production needs enrolled credentials, durable shared state, and Gate on every path to the protected provider credential. A valid approval does not establish that the bank details are legitimate.
Prefer a browser? Try a passkey on a sample payment. That separate demo shows receipt integrity with your platform authenticator; no payment is sent.
Protect one tool you already use
Start at the service that holds the real provider credential, not only inside the agent's process. The owner defines the allowed work and when a fresh approval is required. An agent can work inside those limits without a person approving every call.
- MCP or HTTP: Gate Starter walks through one covered action.
- Hugging Face smolagents: wrap an existing tool.
- GitHub: Merge Gate binds a check to the proposed merge. The check must be required and alternate merge paths closed.
Protocol proves. Gate prevents on the paths the deployment completely mediates. Gate cannot constrain a bypass path. If a provider's result is unknown, the production lifecycle preserves that uncertainty for reconciliation rather than blindly retrying.
AI systems and repository reviewers: start with AI_CONTEXT.md. Current machine-readable evidence, provenance, assumptions, and exclusions are published at EMILIA-REPO-CONTEXT-v1. Archived or staged documents do not establish current implementation or IETF status. Public due-diligence evidence and claim boundaries: DUE_DILIGENCE.md.
Engineering evidence, not architecture claims
EMILIA ships a security case that reviewers can execute. The current repository resolves 35 security claims over 264 hashed evidence files, verifies 20 Tamarin lemmas across two composed Dolev-Yao models — 17 all-traces obligations and 3 exists-trace reachability witnesses — and preserves 8 deliberately weakened variants that produce concrete attack traces when load-bearing checks are removed. The live same-team conformance corpus contains 21 suites and 340 current vectors. Separately, an externally authored Rust verifier is pinned to the frozen 16-suite/164-vector bundle and a 359-case hostility campaign. The broader suite contains 10,700+ automated tests across 650+ files.
Production JavaScript and JSDoc surfaces are compiler-checked with TypeScript
checkJs; the secure app has its own compatibility compiler project, while
declarations and the public TypeScript SDK are checked in strict mode. This is
complete configured production type-check coverage, not a claim that the
repository was converted wholesale from JavaScript to TypeScript or that every
JavaScript project has TypeScript's strict option enabled.
Each security claim names the enforcement path, positive and negative vectors, language coverage,
formal scope or explicit gap, assumptions, exclusions, and evidence hash. Start with the
human-readable evidence map, then inspect the
resolved security case or run npm run check:security-case.
AEB-1: test the evidence-to-effect boundary
The open AEB-1 Consequence Admission Conformance
pack tests the composed CAID/AEC path at the last control point before
a consequential action: native verification, relying-party acceptance,
exact-action binding, required CAID matching, required AEC evidence
satisfaction, local authorization, atomic reservation, INVOKING custody,
separate provider-outcome and observed-effect truth, no-blind-retry behavior,
and authenticated reconciliation.
Read the consequence-admission boundary for the exact division of responsibility across the composed CAID/AEC path, the direct native path, AEB custody, and provider outcome evidence.
npx @emilia-protocol/verify aeb-conformance --reference
It is format-neutral and self-run. A passing report is self-attested conformance evidence, not an audit, certification, production-deployment claim, or permission to execute an action.
Under AEB-07, posted on 2026-09-25 as an individual Internet-Draft and not adopted by any working group, CAID is used only when independently encoded actions must be joined, and AEC is used only when local policy requires several evidence legs. A separate 26-case synthetic corpus models that direct-native lifecycle over AuthZEN/COAZ-MCP, AP2, OAuth Transaction Token, and local signed-mandate adapter results:
npm run conformance:composition:consequence-admission
The corpus runner is a standalone lifecycle model with its own in-memory store
and admission logic. It does not execute the shipped @emilia-protocol/verify
or @emilia-protocol/gate code, so it is not evidence for those packages,
which have their own test suites. It is also not evidence that the named
native protocols conform to AEB-07.
For a focused executable proof of the repository's Gate path, run:
npm run proof:gate:reference
This command exercises local examples and focused service boundaries with generated keys, in-memory state, and mock provider behavior. It is useful local proof, not evidence of a real human, external bank, production deployment, or one end-to-end production integration.
Identity is not a job description
Identity says who or what is calling. Policy says what is generally allowed. Neither defines the finite job an autonomous worker may perform now: its mission, material-action limits, budget, required evidence, expiry, delegation rules, and exception path.
EMILIA keeps those questions separate:
| Layer | Question |
|---|---|
| Identity | Who or what is present? |
| Policy | What is generally allowed? |
| Authority | What exact work may this agent perform under this mandate? |
Credentials grant reach. Authority defines the job. Not every action needs a human; every consequential action needs valid authority.
At the foundation, EP Core still exposes three interoperable objects: a Trust Receipt carries attributable evidence, a Trust Profile represents structured trust state, and a Trust Decision records the relying party's policy-evaluated result. The authority-control-plane layers add exact action binding, finite mandates, admission, consumption, and outcome evidence without collapsing those objects into one claim.
Set the mandate once. Let the agent work.
The customer defines the mission, limits, evidence requirements, expiry, and exception rules. Local code may narrow that authority; it cannot invent or widen it. Gate binds each executable request to the mandate, reserves the covered authority before provider entry, permits one admitted provider attempt for that authorization instance inside the shared durable authority domain, and escalates only when authority is missing, stale, exhausted, or too narrow.
The bundled MCP examples exercise a policy profile requiring approval at the boundary. They generate demonstration signing keys and invoke mock tools. The payment example binds all four declared material fields; the other examples demonstrate narrower resource bindings. They do not capture a real human decision or contact a provider:
node examples/mcp/payment-server.mjs # release_payment — refuses without a receipt
node examples/mcp/github-admin.mjs # delete_repo — refuses without a receipt
node examples/mcp/prod-deploy.mjs # deploy_production — refuses without a receipt
The deeper composition demo executes a CAID-bound delegated payment through Gate's real bounded-capability path, then verifies the signed execution certificate offline:
npm run demo:receipt-program
It deliberately includes no blockchain or simulated zero-knowledge claim. See the receipt-program architecture for the production state and trust requirements.
Start with one declared consequential MCP action. Install the exact local runtime, then create its Gate Starter and run the bounded four-case check:
npm install --save-exact @emilia-protocol/mcp-guard@0.6.0
npx @emilia-protocol/scan@0.5.0 protect ./tools.json --action sendWire --apply --verify
# after reading emilia/authority-map.html and action-control.manifest.json
npx @emilia-protocol/scan@0.5.0 protect ./tools.json --action sendWire --reviewed \
--crossing-profile ccs-wang-draft08-v13
The generated local check uses explicitly ephemeral demo state and proves only
the stated synthetic missing, exact-match, mutation, replay, and unscanned-tool cases.
The reviewed command creates a privacy-bounded handoff; it does not activate Gate.
Production requires a durable provenance ledger,
a shared atomic consumption store, pinned keys, and the wrapper on every path to the real
provider credential. See
examples/mcp/ and /mcp.
Stop new actions without pretending to undo old ones
Gate's emergency authority freeze blocks new reservations and prevents older reservations from entering after the covered control domain's epoch changes. If provider entry happened first, the attempt remains consumed and needs reconciliation. Restoring authority does not revive an old reservation.
This requires complete mediation and authoritative shared state. It does not stop computation, reverse an effect, or instantly reach disconnected domains. See the control-domain implementation and limits.
Try it in 30 seconds
# Issue a receipt offline — no API key, no backend needed
npx @emilia-protocol/issue demo
# Add EMILIA to Claude / Cursor / Cline
npx -y @emilia-protocol/mcp-server
Try a platform passkey on a sample payment → Sign an illustrative $82,000 payment, change its amount, and see the integrity check fail. Your authenticator may use biometrics or a device PIN. The page also offers a separate software simulation. Neither mode sends a payment or establishes production authorization.
Verify any receipt in your browser — paste it in, nothing is uploaded.
How it works — one authority lifecycle

Run it yourself:
node examples/crash-test.mjs— fully offline, no API key.
[ MANDATE ] [ EXACT WORK ] [ VERIFY ] [ RESERVE + ENTER ] [ RECONCILE ]
mission, limits canonical action pinned native one admitted preserve provider
evidence, expiry + occurrence evidence provider entry and effect truth
Mandate. The authority source defines finite work. It can be a customer-signed operating program, bounded capability, required human decision, quorum, or a relying-party composition of native evidence.
Exact work. Gate binds method, origin, callee, target, occurrence, and every material field into the canonical executable object. Intent, a prompt, or ticket text is not that object.
Verify, reserve, and enter. Native artifacts remain native. The relying party pins trust and mapping profiles, evaluates the complete evidence requirement, makes the separate local authorization decision, and reserves the covered authority before the credential-owning adapter enters the provider.
Fresh human authority when required. A policy can require a WebAuthn/passkey decision bound to the exact action and deterministic display hash. This narrows the “what you saw is what you signed” gap; it does not prove comprehension, wisdom, legality, or outcome.
For enterprise deployments, Gate can additionally require an independently verified Authorization Server confirmation bound to that exact human evidence, the same exact action, the identity snapshot the AS actually observed, and the intended Resource Server key. The snapshot time and relying-party maximum age are explicit: a fresh token cannot make stale directory data current. The AS leg is evidence under customer-pinned trust; it never authorizes by itself, proves instantaneous employment standing, or turns the agent orchestrator into an authority.
Truthful result. Admission is not execution, and execution is not effect. A signed record can be
verified offline; provider and observer evidence remain separate. A lost response becomes
INDETERMINATE, which is a state to reconcile—not permission to retry. A remedy is a new authorized
action and never rewrites the old result.
Why developers use it
Start by mapping the work locally, then protect one declared action surface with the MCP server or thin SDK wrapper. The scanner proposes a reviewable map; the owner defines the mandate; Gate owns the provider credential and enforces the exact action on the covered path. No scan proves complete mediation, and discovery alone grants no authority.
# langchain-emilia — wrap any LangChain tool with an EP gate
from langchain_emilia import EmiliaGateClient
gate = EmiliaGateClient(base_url="https://www.emiliaprotocol.ai", api_key="...")
safe_tool = gate.wrap(your_destructive_tool)
pip install langchain-emilia # PyPI
npm install @emilia-protocol/verify # npm
The agent receives the ability to perform bounded work, not a standing credential it can reinterpret.
Why enterprises need it
Agent processes restart and models change. The customer's mandate, consumption state, revocation, uncertainty, and work history must survive outside them. EMILIA keeps that durable authority state at the customer's boundary while accepting foreign proof through pinned adapters.
The managed Gate and Assurance Plane add mandate operations, integrations, evidence operations, re-performance, support, and service levels around the open protocol. The customer retains control of authority, trust roots, credentials, policy, and portable evidence.
The standard
EMILIA Protocol is open and Apache-2.0. Its standards work is published as a portfolio of individual Internet-Drafts. A published Internet-Draft is not an RFC, an adopted working-group item, or IETF endorsement; Datatracker is authoritative for revision and status.
One consequence-boundary surface
The current published
AEB-07,
posted on 2026-09-25 as an individual Internet-Draft and not adopted, makes AEB
the composition point after a native decision. OAuth, AuthZEN, COAZ, AP2, and
local systems keep ownership of their credentials, operation mappings, and
authorization decisions. AIMS (draft-ietf-wimse-aims) is an Informational
WIMSE working-group document that profiles existing standards such as WIMSE
and OAuth; it does not itself issue credentials or decisions. AEB applies the
native decision at the protected provider boundary: it binds the final action
when needed, derives one native replay identity per grant, reserves before
provider entry, refuses a second attempt at the same action while an earlier
one is in flight or uncertain, and keeps an uncertain result locked until
authenticated reconciliation.
CAID-04 is used when independently encoded representations must be compared. It is not a mandatory second mapping when the consequence-owning PEP already derives and enforces a current decision over the final operation. AEC-06 is used when the relying party requires several evidence legs. Authorization Receipts, Human Authorization Binding, and Authority Introduction remain available profiles for deployments that need them; they are not prerequisites for every AEB integration. Architecture-03 remains the navigation document.
The consequence-admission guide states
the implementation boundary and the cases where AEB is unnecessary.
The direct native handoff profile
is a repository implementation profile that shows how an existing permit
reaches Gate without a second CAID or AEC layer. AEB-07 specifies what such a
gateway handoff attests and how the boundary verifies it, but leaves the
encoding to deployment pins; it cites a pinned snapshot of this profile as one
informative reference encoding. The exact submitted -07 bytes and their
publication record are retained in
standards/staged/NEXT-AEB-07.
The complete active portfolio is 26 Datatracker records: 21 sole-authored records and five coauthored records, each with its own scope, revision history, and recorded maintenance status. See the standards guide, portfolio, and machine-readable status inventory.
| IETF Internet-Drafts | Current local snapshot paths: status inventory · sole-authored posted inventory · authoritative live status: IETF Datatracker |
| Cross-language verifiers | JavaScript · Python · Go — all three proven to agree on adversarial conformance vectors, every push (npm run conformance). A consistency check across one team's ports, not clean-room independent implementations. Separately, an externally authored from-spec Rust implementation (source public) passes the pinned 16-suite/164-vector bundle and the pinned 359-case hostility campaign under an evaluator-controlled rebuild from an immutable source tree. Its checked-in construction evidence remains implementer-signed, not third-party-attested (signed statement); strict clean-room acceptance waits for the corrected third-party-attested manifest and independently pinned attestor key. |
| Formal-model evidence | 26 bounded TLA+ safety properties held in their configured state spaces; this is not implementation refinement or an unbounded proof · 35 Alloy facts, 32 assertions across four models · two composed symbolic Dolev-Yao models covering challenge, CAID, two approvals, issuer and authority pins, registry view, revocation, consumption, execution, and six dedicated claim boundaries. Twenty Tamarin lemmas verify — 17 all-traces obligations and 3 exists-trace witnesses; eight deliberately weakened variants produce concrete attack traces when load-bearing checks are removed (formal/tamarin/). |
| MCP distribution | npm package @emilia-protocol/mcp-server · official Registry publication is tracked separately in MCP-REGISTRY.md; aggregator listings are not inferred from either state |
| License | Apache-2.0 |
Three same-team reference ports (JS / Python / Go) agree across all 21 suites and 340 vectors. Separately, an externally authored Rust implementation rebuilt from a pinned public source tree passes the pinned 16-suite/164-vector clean-room bundle and a 359-case hostility campaign, re-run in its own CI lane on every change. The newer AEC acceptance and four-outcome resolution suites are not attributed to Rust. That is external interoperability evidence, not strict clean-room construction acceptance; the aggregate CI case records the strict acceptance count as zero pending independent attestation. See CONFORMANCE.md, or verify a receipt yourself at emiliaprotocol.ai/verify.
The authority stack
| Layer | What it does |
|---|---|
| Mandate | Defines mission, limits, evidence, expiry, delegation, and exception rules. |
| CAID / exact action | Compares material meaning when authorization and execution use independently encoded representations; it does not authorize. |
| AEC | When required, evaluates whether independently verified and matched evidence satisfies the relying party's multi-leg requirement; it does not authorize. |
| AEB / Gate | Applies the native or local authorization decision at the consequence boundary, reserves the covered authority, and controls provider entry. |
| Outcome evidence | Keeps invocation, provider response, observed effect, and uncertainty distinct. |
Proof points
| Metric | Value |
|---|---|
| Automated test cases | 10,700+ across 650+ files; all platform-applicable cases must pass |
| TLA+ safety properties | 26 bounded invariants held in the configured state space; not an implementation-refinement or unbounded proof — see PROOF_STATUS.md |
| Alloy relational assertions | 35 facts + 32 assertions across four models — verified in CI |
| Red-team cases cataloged | 86 — RED_TEAM_CASES.md |
| Release security status | Repository security checks pass; every Strix finding on the audited changes is remediated with regression coverage and its review thread resolved |
| Conformance (7/7) | node conformance/ep-conformance-test.js https://www.emiliaprotocol.ai |
| Cross-language conformance | 340 vectors · 21 suites: receipts · device signoffs · four-outcome resolution · multi-party quorum · revocation · Outcome Binding (semantic + real-crypto) · Authority Document/Proof issuer join · time-attestation · trust-receipt (x2 profiles) · provenance · evidence-record · canonicalization · boundary · AEC acceptance · currency · initiator-attestation · consumption-proof · witness · timestamp-proof (RFC 3161). JS / Python / Go verifiers agree (node conformance/run.mjs). The external Rust baseline remains 164 vectors / 16 suites. See CONFORMANCE.md. |
| Handshake create p95 | 575ms at 50 VUs — PERFORMANCE_PROOF.md |
Cryptographic longevity (with explicit deployment boundaries)
Evidence meant to be verified years later must outlive the algorithms it was signed under. EP ships four bounded capabilities for that, each with an exact boundary that is part of the claim:
- Hybrid signatures (EP-RECEIPT-HYBRID-v1). Ed25519 and ML-DSA-65 over the same
canonical bytes, with the required algorithm set committed into the signed
bytes so stripping a leg breaks the surviving signature. The capability is
opt-in at deployment; once an approved dual signer is registered and policy
permits its PQ leg, an unpinned Gate posture resolves to dual issuance by
default. Otherwise it stays classical-only with a named reason. v1 verifiers
refuse hybrid receipts cleanly rather than accepting one leg. The external
signer contract and AWS KMS adapter are implemented, but no live AWS signing
call, production key, relying-party verification, or ML-DSA FIPS validation
is claimed. See
conformance/hybrid-receipts/andlib/pq-custody-aws-kms.ts. - SCITT Signed Statement profile (EP-SCITT-STATEMENT-v1). A complete RFC 9943 Signed Statement shape for EP receipts, including the CWT Claims protected header. Boundary: no Transparency Service has accepted an EP statement; external registration is a separate, gated step and none has been performed. See EP-RECEIPT-SCITT-PROFILE.md.
- Re-attestation (EP-EVIDENCE-REATTESTATION-v1). Evidence signed under an aging algorithm can be re-anchored under a current one before the old one weakens. Boundary: re-attestation must precede compromise; it cannot repair evidence after the fact.
- FIPS deployment mode (EP-FIPS-MODE-v1). Runs classical operations through an operator-supplied FIPS 140-3 validated provider, with the ML-DSA path gated behind an explicit unvalidated-implementation acknowledgment. Boundary: this earns "FIPS-based algorithms, with a validated-provider deployment mode" and depends on the operator's provider and declared certificate boundary; it is not a blanket compliance claim, and nothing here is FIPS validated. See FIPS-MODE.md.
The stack-wide hybrid program (every internal signature surface) is mapped in pq-hybrid-program.md and is not complete; until it is, no blanket claim about the whole stack is made.
Core protocol objects
| Object | What it is |
|---|---|
| Authority program / bounded capability | A finite mandate with explicit scope, budget or units, expiry, delegation, and consumption rules. |
| CAID | A canonical identifier for one material action under a named mapping profile; matching is not authorization. |
| Evidence requirement and AEC result | The relying party's pinned rule and its SATISFIED, UNSATISFIED, or INDETERMINATE evaluation. |
| AEB admission and custody record | The executor-side record of authorization, reservation, provider entry, and reconciliation state. |
| Authorization and outcome evidence | Portable native or EP artifacts that retain their exact issuer, scope, and claim boundary. |
Quickstart
- Install the exact local guard runtime, then run
npx @emilia-protocol/scan@0.5.0 protect ./tools.json --action sendWire --apply --verify, replacingsendWirewith one exact declared consequential tool. - Review the generated Authority Map, action manifest, material fields, selected boundary, and named blind spots.
- Run the separate
--reviewed --crossing-profile <launch-profile>command to create the owner-only handoff and unsealed Lab workspace from those unchanged bytes. - Install Gate on the path that owns the provider credential and durable consumption state.
- Define the operating mandate and any fresh-human or quorum exception rules.
- Run the refusal, exact-action, replay, timeout, and reconciliation cases before enabling enforcement.
90-second demo · Quickstart · Agent walkthrough · IETF Draft · Discord
What EP is — and is not
EMILIA is authority infrastructure for autonomous work, not an identity system, wallet, reputation score, settlement rail, or universal policy engine.
- Is: a control plane for finite operating mandates, exact-action verification, durable admission state, truthful uncertainty, and portable evidence on covered executor paths.
- Is not: a replacement for OAuth/OIDC, workload identity, or policy engines. Those remain native inputs under the relying party's pins.
- Is not: a requirement that a human approve every action. A mandate may permit automatic work inside finite bounds and demand fresh authority only at the edge.
- Is not: proof that an admitted action executed successfully or caused the intended effect.
- Is not: proprietary protocol control. The core is Apache-2.0 and the Internet-Drafts are individual submissions, not RFCs or IETF endorsement.
See CONFORMANCE.md · SECURITY.md · THREAT_MODEL.md · GOVERNANCE.md · Neutrality Covenant