The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
docs: YOLO / approvals.mode off skip the MCP trust prompt, not the spend limit
Tekniumaa82d233cf24
diff --git a/website/docs/reference/mcp-config-reference.md b/website/docs/reference/mcp-config-reference.mdindex fce0e82eee0..401c573a5d4 100644--- a/website/docs/reference/mcp-config-reference.md+++ b/website/docs/reference/mcp-config-reference.md@@ -69,7 +69,7 @@ mcp_servers:| `sampling` | mapping | both | Server-initiated LLM request policy (see MCP guide) || `elicitation` | mapping | both | Server-initiated user-input requests. `enabled` (default `true`) and `timeout` in seconds (default `300`). Form-mode requests route through the approval surface; URL-mode is declined (see MCP guide) || `events` | bool | both | Experimental MCP events client with webhook delivery (default: `false`). Registers `list_event_types`, `subscribe_event` and `unsubscribe_event` for this server. Needs `platforms.webhook.extra.public_url` (HTTPS) and a running gateway. Hermes refreshes subscriptions only while it is connected to the server, so with `lazy: true` or `idle_timeout_seconds` a subscription can lapse until the next connection resubscribes it. See [MCP Events](../user-guide/features/mcp.md#mcp-events-experimental) |-| `trust` | string | both | Trust tier: `full` (default) or `untrusted`. On an `untrusted` server, every write-capable tool call (any tool without a `readOnlyHint: true` annotation) requires user approval through the standard approval surface before it runs. The prompt offers Allow Once, Session (that tool, for the rest of the chat) and Always (written to `command_allowlist` as `mcp:<server>/<tool>`; deleting the entry revokes it for every other chat, while the chat that answered Always keeps it until that chat ends); each covers one tool, never the server's others. `readOnlyHint` is a server-supplied *hint* — a lying server can at most skip approval for tools it claims are read-only, never gain extra access — so mark any server you don't fully control as `untrusted`. The same hint decides whether a call is transparently retried after the transport session expires mid-call: only `readOnlyHint: true` tools are replayed, while unannotated (write-capable) tools return an `outcome_uncertain` error — on a Streamable-HTTP server that expires idle sessions this means the first unannotated call after an idle period may fail and must be verified before re-invoking. Unrecognized values are treated as `untrusted` (fail-closed) A portable package can request `untrusted` for its own server in `plugin.json` (see [Portable Agent Plugins](../developer-guide/plugins/index.md#portable-agent-plugins-v1-packages)); a `config.yaml` entry with the same name always wins |+| `trust` | string | both | Trust tier: `full` (default) or `untrusted`. On an `untrusted` server, every write-capable tool call (any tool without a `readOnlyHint: true` annotation) requires user approval through the standard approval surface before it runs. The prompt offers Allow Once, Session (that tool, for the rest of the chat) and Always (written to `command_allowlist` as `mcp:<server>/<tool>`; deleting the entry revokes it for every other chat, while the chat that answered Always keeps it until that chat ends); each covers one tool, never the server's others. [YOLO](../user-guide/security.md#yolo-mode) and `approvals.mode: off` skip this prompt for the session they are on; a declared `spend_limit` still asks. `readOnlyHint` is a server-supplied *hint* — a lying server can at most skip approval for tools it claims are read-only, never gain extra access — so mark any server you don't fully control as `untrusted`. The same hint decides whether a call is transparently retried after the transport session expires mid-call: only `readOnlyHint: true` tools are replayed, while unannotated (write-capable) tools return an `outcome_uncertain` error — on a Streamable-HTTP server that expires idle sessions this means the first unannotated call after an idle period may fail and must be verified before re-invoking. Unrecognized values are treated as `untrusted` (fail-closed) A portable package can request `untrusted` for its own server in `plugin.json` (see [Portable Agent Plugins](../developer-guide/plugins/index.md#portable-agent-plugins-v1-packages)); a `config.yaml` entry with the same name always wins || `tool_costs` | mapping | both | What each tool spends, keyed by the MCP tool name: a non-negative number, a `[min, max]` range, or `{credits: <number or range>, note: "per animation"}`. The approval prompt names the cost and the chat's running total. Undeclared and zero-cost tools are unaffected. See [Tool costs and spend limits](../user-guide/features/mcp.md#tool-costs-and-spend-limits) || `spend_limit` | number | both | Per-chat budget for this server's costed tools (`0`/absent = no limit). A costed call that would take the chat past it asks first, even on a `trust: full` server, and is refused where nobody can answer (cron, `chat -q`) || `cost_unit` | string | both | Label shown with costs and totals (default: `credits`) |diff --git a/website/docs/user-guide/features/mcp.md b/website/docs/user-guide/features/mcp.mdindex 3a37fb2328f..15f811069dd 100644--- a/website/docs/user-guide/features/mcp.md+++ b/website/docs/user-guide/features/mcp.md@@ -1003,13 +1003,14 @@ mcp_servers:- The total follows the conversation when context compression or `/branch` continues it under a new session id. Starting a new chat (`/new` in the CLI, a new Desktop chat) starts a new total; on messaging platforms the total belongs to the chat, so it survives `/new` there. Each cron run has its own total. Restarting Hermes resets every total.- With `spend_limit`, a costed call that would take the chat past the limit asks first, even on a `trust: full` server that otherwise never prompts, and every further costed call in that chat asks again: that prompt offers only Allow Once and Deny, and an earlier Session or Always answer for the tool does not silence it. Where nobody can answer (cron jobs, `hermes chat -q`, a gateway with no approver), the call is refused without running and the model is told not to retry.- Below the limit, a `trust: full` server still runs costed tools without asking; on a `trust: untrusted` server every write-capable call already asks and the prompt now carries the cost. Tools annotated `readOnlyHint: true` stay unprompted unless the budget requires it.+- [YOLO](../security.md#yolo-mode) (`/yolo`, `--yolo`) and `approvals.mode: off` skip the `trust: untrusted` prompt for that session, but not the spend limit: below the limit a costed call runs without asking, and a call that would cross it still asks (or is refused where nobody can answer, such as a cron job with `approvals.mode: off`).- A malformed `tool_costs`, `spend_limit` or `cost_unit` skips the server with an error naming the key, so a budget you declared is never silently dropped. A `tool_costs` key the server does not offer (a typo, or Hermes' `mcp_<server>_` display prefix copied in) is named in a warning at connect time: use the server's own tool names.A [portable plugin package](../../developer-guide/plugins/index.md#portable-agent-plugins-v1-packages) can declare the same table for its server. If you add a same-name `mcp_servers` entry (for example to set `trust: full`), the package's `tool_costs`, `spend_limit` and `cost_unit` still apply unless your entry sets those keys itself; `spend_limit: 0` turns the package's limit off and `tool_costs: {}` drops its costs.## MCP Elicitation Support-MCP servers can ask the user for structured input mid-tool-call via the `elicitation/create` protocol (mcp Python SDK ≥ 1.11.0). Hermes routes **form-mode** elicitations through its existing approval surface — an interactive prompt in the CLI/TUI, or approval buttons on gateway platforms like Telegram and Slack — so the request reaches you wherever the session lives. Each answer covers that one request, so the prompt offers only Allow Once and Deny. **URL-mode** elicitations (where a server points you at an external URL) are declined as unsupported.+MCP servers can ask the user for structured input mid-tool-call via the `elicitation/create` protocol (mcp Python SDK ≥ 1.11.0). Hermes routes **form-mode** elicitations through its existing approval surface — an interactive prompt in the CLI/TUI, or approval buttons on gateway platforms like Telegram and Slack — so the request reaches you wherever the session lives. Each answer covers that one request, so the prompt offers only Allow Once and Deny. YOLO and `approvals.mode: off` do not answer these for you: the server is asking you a question, not asking permission. **URL-mode** elicitations (where a server points you at an external URL) are declined as unsupported.Elicitation is **enabled by default** per server. Configure it under the `elicitation` key:diff --git a/website/docs/user-guide/security.md b/website/docs/user-guide/security.mdindex 232a4eec9dc..f2fafdebf0b 100644--- a/website/docs/user-guide/security.md+++ b/website/docs/user-guide/security.md@@ -56,7 +56,7 @@ The full set of keys:|------|----------|| **smart** (default) | Use an auxiliary LLM to assess risk. Low-risk commands (e.g., `python -c "print('hello')"`) are auto-approved for that command only. Genuinely dangerous commands are auto-denied. Uncertain cases escalate to a manual prompt. || **manual** | Always prompt the user for approval on dangerous commands. |-| **off** | Disable all approval checks — equivalent to running with `--yolo`. All commands execute without prompts. |+| **off** | Disable all approval checks — equivalent to running with `--yolo`. All commands execute without prompts, and write calls on `trust: untrusted` MCP servers run without the trust prompt. An MCP server's declared spend limit still asks (see [YOLO mode](#yolo-mode)). |:::warningSetting `approvals.mode: off` disables all safety prompts. Use only in trusted environments (CI/CD, containers, etc.).@@ -91,6 +91,12 @@ When YOLO is active, Hermes shows two persistent visual reminders so it's hard tYOLO mode disables **all** dangerous command safety checks for the session — **except** the hardline blocklist (see below). Use only when you fully trust the commands being generated (e.g., well-tested automation scripts in disposable environments).:::+YOLO and `approvals.mode: off` also skip the [`trust: untrusted`](../reference/mcp-config-reference.md) MCP prompt: a write-capable tool on an untrusted server runs without asking, in that session only (another chat without `/yolo` is still asked). They do **not** skip:++- **An MCP server's spend limit.** A call that would take the chat past a declared [`spend_limit`](./features/mcp.md#tool-costs-and-spend-limits) still asks, and where nobody can answer (a cron job, `hermes chat -q`) it is refused. The limit is a budget you set, not a safety prompt.+- **Questions from an MCP server** ([elicitation](./features/mcp.md#mcp-elicitation-support)) and sampling requests, which ask you for input rather than for permission.+- **Payment-card fill prompts** from the [credential vault](./features/credential-vault.md#paying-and-filling-addresses), which hand a card to a page, and MCP event subscriptions, which install a standing trigger.+For destructive session slash commands (`/clear`, `/new` / `/reset`, `/undo`, `/quit --delete` — `/exit --delete` is an alias), the CLI also prompts for confirmation before running them. See [Slash Commands — Confirmation prompts for destructive commands](../reference/slash-commands.md#confirmation-prompts-for-destructive-commands).### Supervised-gateway lifecycle restriction