Stoffel AI Agent Implementation

Use this skill when asking an AI coding agent to build, modify, deploy, or debug a Stoffel application. The goal is to give the agent enough context to choose the right MPC backend, model secret inputs correctly, and validate the smallest working program before expanding the app.

Give the agent Stoffel docs access

Before a large implementation task, install the Stoffel skills and connect the live docs MCP server so the agent can use current docs instead of training-data guesses. Install the skills:

npx skills add https://docs.stoffelmpc.com

List the published skills first if you need to choose a narrower playbook:

npx skills add https://docs.stoffelmpc.com --list

Connect the Mintlify-hosted search MCP server:

npx add-mcp --name stoffel-docs --transport http https://docs.stoffelmpc.com/mcp

For Claude Code directly:

claude mcp add --transport http stoffel-docs https://docs.stoffelmpc.com/mcp

Programmatic discovery endpoints:

https://docs.stoffelmpc.com/.well-known/agent-skills/index.json
https://docs.stoffelmpc.com/.well-known/mcp
https://docs.stoffelmpc.com/.well-known/mcp/server-card.json

Mandatory portability contract

Before changing files or dependencies, the agent must:

Discover and confirm the project root from its current working directory and repository markers such as Stoffel.toml, Cargo.toml, or .git. It must not invent or require a machine-specific path such as /workspace/....
Select Stoffel and related project dependencies from public, reproducible sources in this order: the current crates.io release, then the official GitHub repository pinned to a full immutable commit SHA when the needed change is not published.
Never use a floating branch or require a local path, sibling checkout, or other external filesystem checkout in the default app implementation.
Use a local Stoffel checkout only when the user explicitly requests framework development. Mark that route nonportable and keep it separate from the default public-dependency implementation.
Stop and report the unavailable dependency and attempted public sources if no suitable public dependency exists. Never silently fall back to a local checkout.

Keep the implementation prompt version-agnostic. Concrete versions belong in installation docs or the generated project manifest.

Start with the application trust architecture

For client-owned private input, the input owner’s device or process is the Stoffel MPC client. It submits directly through the client protocol to the separately deployed MPC service. An application backend may manage public metadata, authorization, session configuration, non-sensitive receipts, lifecycle, and explicitly authorized opened aggregates, but it must not receive or persist participant plaintext. A backend gateway that receives raw input is a separate, weaker trust model. Do not introduce it implicitly or use it to work around missing browser/client support. Before asking an agent to write code, require it to answer:

Which values are secret and public.
Who owns each plaintext input and where that plaintext exists before protection.
Which participant-owned process executes the Stoffel client protocol.
Which components are forbidden from receiving, logging, queueing, caching, analyzing, or persisting plaintext.
What the application control plane may receive and persist.
Who receives each output.
Whether each output is an opened value, client-output share, public commitment, curve-encoded value, or signature-related artifact.
Whether the requested participant runtime supports direct client-protocol submission.
Which backend is selected and why.
Which command proves the program is valid.
Which command proves local MPC works.
Which command proves the production private-data path bypasses the application service.
Which artifacts and configs are required before deployment.
Which discovered project root the agent will modify.
Which public dependency source satisfies the portability contract.
Whether framework development was explicitly requested; otherwise local checkout use is forbidden.

Do not begin implementation while the plaintext location, submission process, forbidden components, persistence allowlist, output recipients, participant runtime support, or dependency source is unresolved. For source snippets, the validation command is usually:

stoffel check path/to/program.stfl

For docs changes, also run:

npx mintlify validate
npx mintlify broken-links

Backend selection context

Give the agent the backend decision in terms of value representation and output boundary. Use HoneyBadgerMPC when:

the program computes over secret integers, fixed-point values, or field-compatible shares;
the main cost questions are multiplication count, multiplication depth, comparisons, and reveal boundaries;
public parameters can stay public while private inputs remain ordinary MPC shares;
the outside system consumes an opened result or client-output shares from private computation.

Use AVSS when:

the program needs public commitments to secret shares;
the backend must use a curve that matches an external verifier or protocol;
public transcript bytes must become curve-field challenges;
the outside system consumes a commitment, curve-encoded value, opened scalar response, or signature-related artifact.

If the task is ordinary private arithmetic over application values, start with HoneyBadgerMPC. Use AVSS only when the protocol needs committed scalar shares, curve compatibility, or threshold-cryptography artifacts.

Prompt template

Build the smallest StoffelLang program for this task before expanding it.

Secret values:
- ...

Public values:
- ...

Input owners and plaintext locations:
- client 0 owns ...; plaintext exists in ...
- client 1 owns ...; plaintext exists in ...

Participant client path:
- participant runtime: native Rust / participant-side Tauri Rust / supported browser client / other
- process that loads bindings and submits each client slot: ...
- separately deployed coordinator/MPC endpoints: ...

Application control plane:
- allowed public metadata/session fields: ...
- persistence allowlist: ...
- non-sensitive receipt shape: ...
- components forbidden from plaintext: API, logs, traces, database, cache, queue, analytics, crash reports, ...

Runtime capability gate:
- direct client protocol support: verified / unsupported / unverified
- if unsupported or unverified: stop or choose an explicitly participant-controlled sidecar; do not add a plaintext backend gateway

Output boundary:
- opened value / client-output share / public commitment / curve-encoded artifact / signature-related artifact

Backend:
- honeybadger or avss:<curve>
- reason for choosing it

Project and dependency portability:
- discovered project root and the repository markers used to confirm it
- dependency source: current crates.io release / official GitHub repository at a full immutable commit SHA
- public-source availability confirmed: yes/no
- framework-development exception explicitly requested: yes/no
- if yes, describe the separate nonportable local-checkout route; if no, do not use local or sibling path dependencies
- if no suitable public source is available, stop and report the attempted sources

Cost and safety constraints:
- keep public constants, transcript bytes, hashes, and encodings public until a secret operation needs them
- avoid secret × secret multiplication unless required
- use public weights/constants when they are not private inputs
- minimize openings and reveal only the intended output boundary
- for AVSS signing flows, never persist or reuse per-signature nonces unless the protocol explicitly requires and protects that state

Validation:
- run `stoffel check ...`
- run `stoffel build --program-info`
- run a local MPC smoke with documented inputs and label it as a trusted fixture harness
- verify control-plane schemas and persistence contain no participant private fields or generic private payloads
- run a plaintext-canary check across service requests, logs, traces, database, cache, queues, analytics, crash reports, and receipts
- verify participant clients submit directly to the separately deployed MPC service
- if deployment is requested, produce bytecode, bindings, topology, coordinator/client config, and operator handoff fields
- if editing docs, run `python scripts/sync_developer_skills.py`, `npx mintlify validate`, and `npx mintlify broken-links`

HoneyBadgerMPC agent guidance

Ask the agent to inspect circuit shape before optimizing code:

Count secret × secret multiplications.
Separate multiplication count from multiplication depth.
Keep public weights, thresholds, normalization factors, and lookup tables public when possible.
Treat comparisons, bit decomposition, and nonlinear functions as expensive until measured.
Use MpcOutput when a client should receive result shares instead of opening the result to the host application.
Size preprocessing for the largest expected input shape if the program runs on a network deployment.

Useful APIs and concepts:

ClientStore.get_number_clients()
ClientStore.take_share(...)
ClientStore.take_share_fixed(...)
Mpc.has_capability("client-output")
MpcOutput.send_to_client(...)
Share.add, Share.mul, Share.mul_scalar
opened values via .reveal(), Share.open(...), or open_fixed() where appropriate

AVSS agent guidance

Ask the agent to keep the cryptographic transcript explicit:

Choose the curve based on the external verifier or protocol.
Keep transcript bytes public until they become a scalar challenge with Crypto.hash_to_field(..., Mpc.curve()).
Use get_commitment(0) when the outside system needs a public commitment or public key point.
Persist long-lived key shares with LocalStorage when the workflow requires stable key material.
Do not persist or reuse per-signature nonces unless the protocol explicitly requires and protects that state.
Minimize openings; open scalar responses only when the protocol boundary requires it.

Useful APIs and concepts:

Share.random()
LocalStorage.exists(...)
LocalStorage.store(...)
LocalStorage.load(...)
LocalStorage.load_share(...)
secret_key.get_commitment(0)
Mpc.curve()
Crypto.hash_to_field(...)
Crypto.point_to_sec1(...)
MpcOutput.send_to_client(...)

Common corrections to give the agent

Do not treat client, app, backend, gateway, and participant-side Tauri Rust as interchangeable roles.
Do not send participant plaintext through an application endpoint in the default client-owned-input architecture; a gateway that does this is a separately approved degraded-trust design.
Do not use local .with_client_input(...) fixture injection as production private-data-plane evidence.
Do not compensate for unsupported browser/WASM submission by silently adding a plaintext backend gateway.
Do not use AVSS for generic private arithmetic unless the task needs commitments or curve-compatible artifacts.
Do not reveal intermediate secret values just to make the program easier to write.
Do not turn public transcript material into secret shares earlier than necessary.
Do not assume a curve selector applies to HoneyBadgerMPC.
Do not stop after writing code; run the relevant validation command and report the real output.

Overview

App Development Skills

Give the agent Stoffel docs access

Mandatory portability contract

Start with the application trust architecture

Backend selection context

Prompt template

HoneyBadgerMPC agent guidance

AVSS agent guidance

Common corrections to give the agent

See also

​Give the agent Stoffel docs access

​Mandatory portability contract

​Start with the application trust architecture

​Backend selection context

​Prompt template

​HoneyBadgerMPC agent guidance

​AVSS agent guidance

​Common corrections to give the agent

​See also

Give the agent Stoffel docs access

Mandatory portability contract

Start with the application trust architecture

Backend selection context

Prompt template

HoneyBadgerMPC agent guidance

AVSS agent guidance

Common corrections to give the agent

See also