preste

Every agent run. Signed. Forever.

preste is the first agentic CLI that combines BYOK provider routing, three loop patterns, a local MCP server, and automatic Ed25519 cryptographic attestation on every single run — in a single MIT-licensed binary.

npm install -g @vauban-org/preste
preste setup   # connect your LLM in 60 seconds
preste agent run BUILDER --task "scaffold a /api/ping route"
# ✓ attestation saved → ~/.preste/attestations/b0c520a4....json

Why preste?

| Feature | preste | Hermes | OpenClaw | Claude Code | |---|:---:|:---:|:---:|:---:| | BYOK — any LLM provider | ✅ | ✅ | ✅ | ❌ | | Standalone — no server required | ✅ | ✅ | ✅ | ❌ | | OODA / ReAct / Plan loops | ✅ (3) | ❌ (1) | ❌ (1) | ❌ | | Ed25519 attestation — every run | ✅ | ❌ | ❌ | ❌ | | Local MCP server (IDE delegation) | ✅ | ❌ | ❌ | — | | Starknet on-chain anchor | ✅ | ❌ | ❌ | ❌ | | SKILL.md (agentskills.io standard) | ✅ | ❌ | ❌ | ❌ | | IETF draft conformance | ✅ | ❌ | ❌ | ❌ |

Quick start

npm install -g @vauban-org/preste
preste setup
preste agent run BUILDER --task "list the TypeScript files in src/ and count them"
preste attest list          # see your signed certificates
preste attest verify <runId>

Real output:

[BUILDER › OODA › deepseek-v4-flash]
  step 1  list_directory src/
  step 2  list_directory src/routes/
  step 3  ✓ done

There are 27 TypeScript files in packages/cli/src/.

✓ attestation saved  ~/.preste/attestations/b0c520a4....json  (kid=f33f00c0…)
  3.7 s  ·  3 steps  ·  3389 in / 207 out tokens

Installation

npm install -g @vauban-org/preste   # global
pnpm add @vauban-org/preste         # per-project

Requires Node.js ≥ 18.

Onboarding

preste setup

Interactive wizard. Tests connectivity before saving to ~/.preste/config.yaml.

| Provider | Default model | Notes | |---|---|---| | LiteLLM (sovereign) | deepseek-v4-flash | Self-host. Recommended. | | Groq | llama-3.3-70b-versatile | Free tier. Fast fallback. | | OpenAI | gpt-4o | Standard. |

Resolution order (first wins): ~/.preste/config.yaml → LITELLM_URL → GROQ_API_KEY → OPENAI_API_KEY.

Agent run

preste agent run <TYPE> --task "<description>"

Types: ARCHITECT · BUILDER · TESTER · SCRIBE · SYNERGY

Loop patterns (--pattern):

| Pattern | What it does | Pick when | |---|---|---| | ooda (default) | Structured function-calling loop | Files, shell, HTTP | | react | Text-based Thought → Action → Observe | Heavy reasoning | | plan | Upfront plan → sequential execution | Multi-step tasks |

# OODA — structured tool calls (default)
preste agent run BUILDER --task "scaffold a retry helper with exponential backoff"

# ReAct — reasoning-first
preste agent run ARCHITECT --task "review circuit-breaker in src/router/" --pattern react

# Plan-then-Execute
preste agent run TESTER --task "write edge-case tests for the quota manager" --pattern plan

# One-shot — no loop, just an LLM answer
preste agent run SCRIBE --task "summarise this doc" --one-shot --context "$(cat notes.md)"

# With episodic memory
preste agent run SYNERGY --task "weekly trend analysis" --recall --remember

| Flag | Default | Description | |---|---|---| | --task <text> | required | Task description | | --pattern ooda\|react\|plan | ooda | Loop pattern | | --context <text> | — | Extra context injected into system prompt | | --one-shot | false | Single LLM call, no tools | | --max-steps <n> | 20 | Max loop iterations | | --recall | false | Inject relevant past memories before run | | --remember | false | Save this run to episodic memory | | --no-attest | false | Skip attestation | | --json | false | JSON output |

Mobile observe + steer via sovereign relay

Pair your phone with a running preste --remote session through a zero-knowledge relay. Both sides dial OUT to the relay (NAT-friendly). The relay forwards opaque ciphertext frames PC ↔ phone and never decrypts them — X25519 + AES-256-GCM, ephemeral keys per session.

# 1. Start preste with the Vauban-hosted relay (or your own).
preste --remote --relay https://relay.preste.vauban.tech
#   📡 remote-control — observe + steer this session from your phone
#      http://0.0.0.0:3113/remote/stream
#      token: <hex>
#
#   🌍 relay — pair your phone from anywhere
#      scan the QR with preste.vauban.tech (PWA) :
#      [ASCII QR]
#      preste-pair://eyJ2IjoxLCJyZWxheVVybCI6Imh0dHBzOi8v...

Open preste.vauban.tech on your phone (install as a PWA — works iOS/Android/desktop), scan the QR, and observe the live transcript. Tap-to-veto each tool call. Voice STT mic to inject, TTS to read back the assistant.

Self-host the relay (production / regulated workloads) :

# 1-command docker-compose with Caddy auto-TLS :
git clone https://github.com/vauban-org/command-center
cd command-center/packages/preste-relay
cp docker-compose.example.yml docker-compose.yml
cp Caddyfile.example Caddyfile
# Edit Caddyfile : replace relay.example.com with your domain.
docker-compose up -d

Cert auto-provisioned by Let's Encrypt + rotated by Caddy. The relay itself : ~300 lines of pure Node http + crypto, MIT, no runtime deps.

See packages/preste-relay/README.md for the protocol spec + the K8s ingress patterns Vauban uses for the public relay.

Cryptographic attestation

Every preste agent run auto-generates a signed RunCertificate. No flag. No opt-in. No server.

# List all certificates
preste attest list

# Verify offline — no network, no Vauban infra
preste attest verify <runId>
# ✓ valid · owned by local key

# Anchor the Poseidon hash on Starknet
preste attest anchor <runId>                      # dry-run (computes hash only)
preste attest anchor <runId> --network sepolia    # real transaction (needs STARKNET_PRIVATE_KEY)

Certificate JSON (excerpt):

{
  "runId": "b0c520a458aefcee87654ac0e479279c",
  "agentId": "BUILDER",
  "status": "completed",
  "stepCount": 3,
  "pattern": "ooda",
  "provider": { "name": "litellm", "model": "deepseek-v4-flash" },
  "usage": { "inputTokens": 3389, "outputTokens": 207 },
  "signatureAlg": "Ed25519",
  "signaturePubkey": "f33f00c09dc3179f…",
  "signature": "3b6e8a…"
}

Anchor dry-run output:

{
  "runId": "b0c520a458aefcee87654ac0e479279c",
  "poseidonHash": "0x2eaaaa012a1beb050a5d6236cfe80b53daeb7577db9d458f030740cbe3e93fc",
  "dryRun": true
}

How it works:

1. First run generates ~/.preste/ed25519.key (chmod 600). Your key. Never leaves your machine.
2. Canonical JSON via JCS (RFC 8785) + Poseidon Merkle root over loop steps.
3. Signed with Ed25519. Cert saved to ~/.preste/attestations/<runId>.json.
4. Third parties verify with the embedded SPKI pubkey — zero Vauban infrastructure.

Conforms to IETF draft-vauban-skill-attestation-00.

Regulatory fit: Provides the tamper-evident traceability substrate for AI Act Art. 14 (human oversight) and EU AI Office guidance on high-risk AI. Auditable 5 years later from the JSON file alone.

MCP server — let your IDE delegate to preste

preste serve --mcp turns the binary into a local MCP server. Claude Code, Claude Desktop, Cursor, Continue.dev spawn it and call agent runs directly — each producing a signed RunCertificate.

preste serve --mcp               # stdio (IDE-spawned)
preste serve --http --port 3200  # HTTP (StreamableHTTP)

Wire into Claude Code (~/.claude.json):

{
  "mcpServers": {
    "preste": {
      "type": "stdio",
      "command": "preste",
      "args": ["serve", "--mcp"]
    }
  }
}

Tools exposed:

| Tool | Purpose | |---|---| | preste_agent_run | Execute OODA / ReAct / Plan agent | | preste_agent_models | List models on your LLM provider | | preste_skills_list | List installed SKILL.md skills | | preste_attest_verify | Verify Ed25519 RunCertificate | | preste_attest_list | Enumerate local attestations | | preste_attest_anchor | Poseidon hash + optional Starknet tx |

Skills (SKILL.md)

preste supports the agentskills.io standard.

preste skills install starknet-validator
preste skills run starknet-validator --task "check validator health"
preste skills validate ./my-skill/SKILL.md      # lint your own skills
preste skills create my-skill --tier verified   # scaffold from template

Tiers: official (Vauban-signed) · verified (community-reviewed) · unverified

Marketplace (P9)

Cryptographically signed skill marketplace. Every package ships a proof.json ProofClaim ; install refuses any package whose signature does not verify (no --no-verify opt-out). Full reference : docs/skills-marketplace.md.

# Search the marketplace.
preste skills search "starknet" --registry https://skills.vauban.tech

# Install a specific version (slug@version).
preste skills install [email protected] --registry https://skills.vauban.tech

# Publish your own (after staging SKILL.md + script.ts + proof.json).
preste skills publish --path ./my-skill --registry https://skills.vauban.tech

Bearer for publish lives in ~/.preste/auth.yaml (key registry_token). Registry URL resolution : --registry > PRESTE_SKILLS_REGISTRY env > https://skills.vauban.tech.

Persona

preste persona set --name "Senior Rust Engineer" --tone concise --language en
preste persona set --domain "distributed systems" --domain "ZK proofs"
preste persona show
preste persona reset

Stored at ~/.preste/persona.yaml, injected into every agent system prompt.

Episodic memory

preste memory recall "circuit breaker implementation"
preste memory store --content "Decided to use sliding window for rate limiting" --type decision

# Or during agent run
preste agent run BUILDER --task "..." --recall --remember

Memory lives at ~/.preste/memory/episodic.jsonl. Recall uses RRF-scored retrieval (cosine + recency + importance).

Session memory

preste uses three-stage memory within a conversation session:

| Stage | Scope | Storage | |---|---|---| | Working memory | Last 6 turns verbatim | In-process (configurable) | | Episodic | Older turns, LLM-compacted | In-process snapshot | | Long-term | Compaction summaries across sessions | Brain (optional, BYOK) |

In-session commands:

/compact              # compress older turns now (automatic at 70% context fill)
/save                 # persist the session snapshot
preste cmd-chat --resume <snapshot>  # restore a prior session

Enabling Brain long-term storage (optional):

# ~/.preste/config.yaml
brain:
  api_key: bk_...       # from brain.vauban.tech
  brain_id: <uuid>      # auto-detected on first run if omitted

When configured, compaction summaries are archived to Brain after each /compact. The next session restores them as ## Long-term context in the system prompt.

Without Brain config: identical behaviour to 0.24 -- zero errors, zero prompts.

Commands reference

| Command | Description | |---|---| | preste setup | Interactive BYOK onboarding | | preste agent run <TYPE> | Run OODA / ReAct / Plan agent | | preste agent list | List agent types | | preste agent models | List models on the configured provider | | preste skills list | List installed skills | | preste skills install <name> | Install from CC backend or marketplace (with --registry) | | preste skills publish | Publish a signed skill package to the marketplace (P9) | | preste skills search <query> | Search CC backend or marketplace (with --registry) | | preste skills run <name> | Execute a local skill | | preste skills validate <path> | Lint a SKILL.md | | preste skills create <name> | Scaffold a new skill | | preste attest verify <runId> | Verify cert offline | | preste attest list | List all local certs | | preste attest anchor <runId> | Poseidon hash + Starknet anchor | | preste serve --mcp | Local MCP server (stdio) | | preste serve --http | Local MCP server (HTTP) | | preste persona show\|set\|reset | Manage agent persona | | preste memory recall <query> | Search episodic memory | | preste memory store | Write to episodic memory | | preste cmd-chat --resume <snap> | Restore session from snapshot | | /compact (in-session command) | Compress older turns | | /save (in-session command) | Persist session snapshot |

Environment variables

| Variable | Description | |---|---| | LITELLM_URL | Sovereign LiteLLM base URL | | LITELLM_API_KEY | LiteLLM API key (optional) | | LITELLM_MODEL | Model override. Default: deepseek-v4-flash | | GROQ_API_KEY | Groq API key | | OPENAI_API_KEY | OpenAI API key | | CC_SERVICE_URL | CC server URL (for --server mode) | | CC_SERVICE_TOKEN | Bearer token for CC server (cc:read scope) | | STARKNET_PRIVATE_KEY | Private key for preste attest anchor --network | | STARKNET_ACCOUNT | Starknet account address | | PRESTE_SKILLS_REGISTRY | Skill marketplace URL (default https://skills.vauban.tech) | | PRESTE_HOME | Override ~/.preste storage prefix (config, skills, attestations) | | PRESTE_ALLOW_MAINNET_RECEIPT | Lift the Sepolia-only gate for zkpay receipts ; set to literal "1" only (operator burden per ADR-ECO-022 v2) |

Programmatic API

import { runAgentRun, runAttestVerify } from '@vauban-org/preste';

const result = await runAgentRun({
  agentType: 'BUILDER',
  task: 'scaffold a health-check route for /api/ping',
  pattern: 'ooda',
  maxSteps: 10,
});

console.log(result.content);
// → "Here is the health-check route..."

const verify = await runAttestVerify({ runIdOrPath: result.runId });
console.log(verify.valid); // true — Ed25519 signature verified

Architecture

Single esbuild bundle (~423 KB), no runtime server dependency.

• Loop: @vauban-org/agent-sdk AgentLoop (minimal-loop, function-calling)
• Attestation: node:crypto Ed25519 + Poseidon felt252 (starknet.js)
• MCP: @modelcontextprotocol/sdk Stdio + StreamableHTTP transports
• Config: YAML (yaml lib), Zod-validated schemas
• Memory: JSONL append-only at ~/.preste/memory/episodic.jsonl

Exit codes

| Code | Meaning | |---|---| | 0 | Success | | 1 | Validation failure (invalid cert, agent error) | | 2 | Strict mode failure | | 3 | Fatal (network error, unrecoverable I/O) |

Contributing

PRs welcome. pnpm install && pnpm build && pnpm test

License

MIT — Vauban · Command Center