1Claw — HSM-Backed Secret Management for AI Agents

1Claw is a cloud HSM-backed secrets manager. Agents access API keys, passwords, and credentials at runtime without them ever entering the conversation context. Secrets are encrypted with keys that never leave the HSM.

API base URL: https://api.1claw.xyz Shroud (TEE proxy): https://shroud.1claw.xyz MCP endpoint: https://mcp.1claw.xyz/mcp Dashboard: https://1claw.xyz Docs: https://docs.1claw.xyz

When to use this skill

• You need an API key, password, or credential to complete a task
• You want to store a newly generated credential securely
• You need to share a secret with a user or another agent
• You need to rotate a credential after regenerating it
• You want to check what secrets are available before using one
• You need to sign or simulate an EVM transaction without exposing private keys
• You want TEE-grade key isolation for transaction signing (use Shroud at shroud.1claw.xyz)

Setup

Option 0: Self-enrollment (new agents)

If you don't have credentials yet, self-enroll — this sends them to your human:

# curl
curl -s -X POST https://api.1claw.xyz/v1/agents/enroll \
  -H "Content-Type: application/json" \
  -d '{"name":"my-agent","human_email":"human@example.com"}'

# TypeScript SDK (static method, no auth needed)
import { AgentsResource } from "@1claw/sdk";
await AgentsResource.enroll("https://api.1claw.xyz", {
  name: "my-agent",
  human_email: "human@example.com",
});

# CLI (no auth needed)
npx @1claw/cli agent enroll my-agent --email human@example.com

The human receives the Agent ID + API key by email. They then configure policies for your access.

Option 1: MCP server (recommended for AI agents)

Add to your MCP client configuration. Only the API key is required — agent ID and vault are auto-discovered.

{
    "mcpServers": {
        "1claw": {
            "command": "npx",
            "args": ["-y", "@1claw/mcp"],
            "env": {
                "ONECLAW_AGENT_API_KEY": "<agent-api-key>"
            }
        }
    }
}

Optional overrides: ONECLAW_AGENT_ID (explicit agent), ONECLAW_VAULT_ID (explicit vault).

Hosted HTTP streaming mode:

URL: https://mcp.1claw.xyz/mcp
Headers:
  Authorization: Bearer <agent-jwt>
  X-Vault-ID: <vault-uuid>

Option 2: TypeScript SDK

npm install @1claw/sdk

import { createClient } from "@1claw/sdk";

const client = createClient({
    baseUrl: "https://api.1claw.xyz",
    apiKey: process.env.ONECLAW_AGENT_API_KEY,
});

Option 3: Direct REST API

Authenticate, then pass the Bearer token on every request.

# Exchange agent API key for a JWT (key-only — agent_id is auto-resolved)
RESP=$(curl -s -X POST https://api.1claw.xyz/v1/auth/agent-token \
  -H "Content-Type: application/json" \
  -d '{"api_key":"<key>"}')
TOKEN=$(echo "$RESP" | jq -r .access_token)
AGENT_ID=$(echo "$RESP" | jq -r .agent_id)

# Use the JWT
curl -H "Authorization: Bearer $TOKEN" https://api.1claw.xyz/v1/vaults

Alternative: 1ck_ API keys (personal or agent) can be used directly as Bearer tokens — no JWT exchange needed.

Authentication

Agent auth flow

1. Human registers an agent in the dashboard or via POST /v1/agents with an auth_method (api_key default, mtls, or oidc_client_credentials). For api_key agents → receives agent_id + api_key (prefix ocv_). For mTLS/OIDC agents → receives agent_id only (no API key).
2. All agents auto-receive an Ed25519 SSH keypair (public key on agent record, private key in __agent-keys vault).
3. API key agents exchange credentials: POST /v1/auth/agent-token with { "api_key": "<key>" } (or { "agent_id": "<uuid>", "api_key": "<key>" }) → returns { "access_token": "<jwt>", "expires_in": 3600, "agent_id": "<uuid>", "vault_ids": ["..."] }. Agent ID is optional — the server resolves it from the key prefix.
4. Agent uses Authorization: Bearer <jwt> on all subsequent requests.
5. JWT scopes derive from the agent's access policies (path patterns). If no policies exist, scopes are empty (zero access). The agent's vault_ids are also included in the JWT — requests to unlisted vaults are rejected.
6. Token TTL defaults to ~1 hour but can be set per-agent via token_ttl_seconds. The MCP server auto-refreshes 60s before expiry.

API key auth

Tokens starting with 1ck_ (human personal API keys) or ocv_ (agent API keys) can be used as Bearer tokens directly on any authenticated endpoint.

MCP Tools Reference

list_secrets

List all secrets in the vault. Returns paths, types, and versions — never values.

get_secret

Fetch the decrypted value of a secret. Use immediately before the API call that needs it. Never store the value or include it in summaries.

put_secret

Store a new secret or update an existing one. Each call creates a new version.

delete_secret

Soft-delete a secret. Reversible by an admin.

describe_secret

Get metadata (type, version, expiry) without fetching the value. Use to check existence.

rotate_and_store

Store a new value for an existing secret, creating a new version. Use after regenerating a key.

get_env_bundle

Fetch an env_bundle secret and parse its KEY=VALUE lines as JSON.

create_vault

Create a new vault for organizing secrets.

list_vaults

List all vaults accessible to you. No parameters.

grant_access

Grant a user or agent access to a vault path pattern.

share_secret

Share a secret via link, with your creator, or with a specific user/agent.

Targeted shares (creator/user/agent) require the recipient to explicitly accept before access.

simulate_transaction

Simulate an EVM transaction via Tenderly without signing. Returns balance changes, gas estimates, success/revert status.

submit_transaction

Submit an EVM transaction for signing and optional broadcast. Requires intents_api_enabled.

REST API Quick Reference

Base URL: https://api.1claw.xyz. All authenticated endpoints require Authorization: Bearer <token>.

Auth (public — no token required)

Auth (authenticated)

Vaults

Secrets

Agents

Policies (Access Control)

Sharing

Intents API (requires intents_api_enabled)

Audit

Billing

Chains

Other

SDK Method Reference

All methods return Promise<OneclawResponse<T>>. Access via client.<resource>.<method>(...).

OpenAPI spec for custom SDKs

The API spec is published as an npm package for generating clients in any language:

npm install @1claw/openapi-spec

Ships openapi.yaml and openapi.json. Use with any OpenAPI 3.1 codegen tool:

# TypeScript
npx openapi-typescript node_modules/@1claw/openapi-spec/openapi.yaml -o ./types.ts

# Python
openapi-generator generate -i node_modules/@1claw/openapi-spec/openapi.yaml -g python -o ./oneclaw-py

# Go
oapi-codegen -package oneclaw node_modules/@1claw/openapi-spec/openapi.yaml > oneclaw.go

SDK also re-exports generated types: import type { ApiSchemas } from "@1claw/sdk".

Supported Chains

Default chain registry (query GET /v1/chains for live list):

Use chain names (e.g. "base", "sepolia") or numeric chain IDs in transaction requests.

Access Control Model

Agents do not get blanket access. A human must create a policy to grant an agent access to specific secret paths.

• Path patterns: Glob syntax — api-keys/*, db/**, ** (all)
• Permissions: read, write (delete requires write)
• Conditions: IP allowlist, time windows (JSON)
• Expiry: Optional ISO 8601 date

If no policy matches → 403 Forbidden. Vault creators always have full access (owner bypass).

Vault binding and token scoping

Agents can be restricted beyond policies:

• vault_ids: Restrict the agent to specific vaults. If non-empty, any request to a vault not in the list returns 403.
• token_ttl_seconds: Custom JWT expiry per agent (e.g., 300 for 5-minute tokens).
• Scopes from policies: JWT scopes are derived from the agent's access policies. If an agent has no policies and no explicit scopes, it has zero access.

Set via dashboard, CLI (--token-ttl, --vault-ids), SDK, or API.

Customer-Managed Encryption Keys (CMEK)

Enterprise opt-in feature (Business tier and above). A human generates a 256-bit AES key in the dashboard — the key never leaves their device. Only its SHA-256 fingerprint is stored on the server.

• Enable: POST /v1/vaults/{id}/cmek with { fingerprint }
• Disable: DELETE /v1/vaults/{id}/cmek
• Rotate: POST /v1/vaults/{id}/cmek-rotate (server-assisted, batched in 100s)
• Secrets stored in a CMEK vault have cmek_encrypted: true in responses

Agents reading from a CMEK vault receive the encrypted blob. The CMEK key is required to decrypt client-side. This is designed for organizations with compliance requirements — the default HSM encryption is already strong.

Intents API

When intents_api_enabled = true (set by a human):

1. Agent gains transaction signing via the Intents API (keys stay in HSM)
2. Agent is blocked from reading private_key and ssh_key secrets directly (403)

Default signing key path: keys/{chain}-signer. Override with signing_key_path.

Replay protection (Idempotency-Key)

Include an Idempotency-Key: <unique-string> header on POST /v1/agents/{id}/transactions. The server SHA-256 hashes the key and caches the result for 24 hours. Duplicate submissions with the same key return the cached response instead of re-signing and re-broadcasting. If two concurrent requests share a key, one returns 409 (retry after a moment).

Server-side nonce serialization

When nonce is omitted from a transaction request, the server resolves it automatically via eth_getTransactionCount (pending) and serializes concurrent callers with SELECT FOR UPDATE. This prevents two in-flight submissions from the same agent+chain+address from receiving the same nonce. You can still pass an explicit nonce to override.

signed_tx field gating

GET endpoints (/v1/agents/{id}/transactions and /v1/agents/{id}/transactions/{txid}) redact the signed_tx field by default to reduce exfiltration risk. To include it, pass ?include_signed_tx=true. The initial POST response always includes signed_tx for the originating caller.

Transaction guardrails

Human-configured, server-enforced limits on what the Intents API allows:

Agents cannot modify their own guardrails. Violations return 403 with a descriptive error.

Shroud per-agent LLM proxy

When shroud_enabled = true (set by a human), the agent's LLM traffic is routed through Shroud (shroud.1claw.xyz) for secret redaction, PII scrubbing, prompt injection defense, threat detection, and policy enforcement inside a TEE.

shroud_config is an optional JSON object that lets humans fine-tune the proxy behavior per agent:

Basic settings

Threat detection settings

Multi-layered detection for prompt injection, command injection, social engineering, and data exfiltration attempts:

unicode_normalization object:

command_injection_detection object:

social_engineering_detection object:

encoding_detection object:

network_detection object:

filesystem_detection object:

SDK:

await client.agents.create({
    name: "my-agent",
    shroud_enabled: true,
    shroud_config: {
        pii_policy: "redact",
        injection_threshold: 0.8,
        allowed_providers: ["openai", "anthropic"],
        max_requests_per_day: 1000,
        daily_budget_usd: 10.0,
        enable_secret_redaction: true,
        // Threat detection
        unicode_normalization: { enabled: true, normalize_homoglyphs: true },
        command_injection_detection: { action: "block", strictness: "default" },
        social_engineering_detection: { action: "warn", sensitivity: "medium" },
        encoding_detection: { action: "warn", detect_base64: true },
        network_detection: { action: "warn", blocked_domains: ["pastebin.com"] },
        filesystem_detection: { action: "log" },
        sanitization_mode: "block",
        threat_logging: true,
    },
});

await client.agents.update(agentId, {
    shroud_enabled: true,
    shroud_config: { pii_policy: "block", injection_threshold: 0.9 },
});

CLI:

1claw agent create my-agent --shroud
1claw agent update <agent-id> --shroud true
1claw agent update <agent-id> --shroud false

MCP: When shroud_enabled is true, the agent can send LLM requests through shroud.1claw.xyz. The Shroud proxy enforces the agent's shroud_config policy automatically — no client-side changes needed.

Share with Your Human

Agents can share secrets back with the human who created or enrolled them. Use recipient_type: "creator" — no email or user ID needed.

Via MCP:

share_secret(secret_id: "...", recipient_type: "creator", expires_at: "2026-12-31T00:00:00Z")

Via SDK:

await client.sharing.create(secretId, {
    recipient_type: "creator",
    expires_at: "2026-12-31T00:00:00Z",
    max_access_count: 5,
});

The human sees the share in their Inbound shares and accepts it. This is the primary pattern for agents that discover or generate credentials and need to report them to their human.

Fleet Patterns

When many agents operate in the same organization:

• Vault organization: Use a shared vault with path-scoped policies (e.g. agents/{name}/**) or per-agent vaults for strict isolation.
• Bulk provisioning: Use the authenticated POST /v1/agents endpoint with a human API key to create many agents, or stagger self-enrollment calls to respect the 10-min per-email cooldown.
• Vault binding: Set vault_ids on each agent to restrict JWT scope beyond what policies allow.
• Token TTL: Shorten to 5 min for ephemeral tasks (token_ttl_seconds: 300), keep default 1h for long-running agents.
• Transaction guardrails: Apply tx_max_value_eth, tx_daily_limit_eth, and tx_allowed_chains to all Intents API agents.
• Monitoring: Filter the audit log by agent ID to track per-agent activity. Use billing usage to monitor org-wide consumption.

Security Model

• Credentials are configured by the human, not the agent. The MCP server reads them from env vars.
• The agent never sees its own credentials. The MCP server authenticates on the agent's behalf.
• Access is deny-by-default. Even with valid credentials, only policy-allowed secrets are accessible.
• Secret values are fetched just-in-time and must never be stored, echoed, or included in summaries.
• Agents cannot create email-based shares (prevents unsolicited email sharing).
• Intents API is opt-in. When enabled, raw key reads are blocked.
• Transaction guardrails are human-controlled and server-enforced.
• Token revocation: DELETE /v1/auth/token (or SDK auth.logout()) revokes the current Bearer token; revoked tokens return 401.
• Request body limit: 5MB max; larger requests return 413.

Error Handling

All error responses include a detail field with a human-readable message.

Best Practices

1. Fetch secrets just-in-time. Call get_secret immediately before the API call that needs the credential.
2. Never echo secret values. Say "I retrieved the API key and used it" — never include raw values in responses.
3. Use describe_secret first to check existence or validity before fetching the full value.
4. Use list_secrets to discover available credentials before guessing paths.
5. Rotate after regeneration. If you regenerate an API key at a provider, immediately rotate_and_store the new value.
6. Use grant_access for vault-level sharing — creates a fine-grained policy with path patterns.
7. Use share_secret for one-off sharing — creates a time-limited, access-counted share link.
8. Simulate before signing. Always use simulate_first: true (default) or call simulate_transaction before submit_transaction.
9. Check list_vaults before creating. Avoid creating duplicate vaults.
10. Handle 402 gracefully. Billing/quota errors should be surfaced to the user, not retried.

Billing Tiers

Overage methods: prepaid credits (top up via Stripe, deducted per request) or x402 micropayments (per-query on-chain payments on Base).

Audit, org, security, chain, billing, and auth endpoints are free and never consume quota.

Links

• Dashboard: 1claw.xyz
• Docs: docs.1claw.xyz
• Status: 1claw.xyz/status
• API: https://api.1claw.xyz
• SDK: @1claw/sdk on npm
• OpenAPI Spec: @1claw/openapi-spec on npm
• MCP Server: @1claw/mcp on npm
• CLI: @1claw/cli on npm
• GitHub: github.com/1clawAI
• Support: ops@1claw.xyz