MCP Server
Control Rymi with natural language. The @rymi/mcp package connects your AI assistant directly to your Rymi account. Create agents, inspect calls, validate configs, and go live without opening the dashboard or writing a single HTTP request.
Works with Claude Desktop, Claude Code, Cursor, Windsurf, or any MCP-compatible client.
Rymi is listed on the official MCP Registry as io.github.rymi-live/rymi-mcp, so clients that browse the registry can find it there.
How it works
You (natural language)
↓
AI client (Claude, Cursor, etc.)
↓ MCP tool calls
mcp.rymi.live ← you sign in (OAuth), or pass an API key
↓
Rymi APIThe MCP server translates your plain-English instructions into typed Rymi API calls. The AI figures out which tools to chain. You describe what you want.
Quickstart: Claude (no API key)
Add Rymi as a custom connector and sign in. Nothing to install, no key to paste.
Open claude.ai → Settings → Connectors
Add custom connector, name it Rymi, and paste this URL:
https://mcp.rymi.live/mcpConnect, and sign in to Rymi when the browser opens.
Paste the URL exactly, no trailing slash. OAuth identifies the server by that string, and …/mcp/ is a different one.
One-click install, where the client supports it:
- Add to Cursor
- Add to VS Code
- Claude Code:
claude mcp add --transport http rymi https://mcp.rymi.live/mcp, then/mcpto sign in
Steps for ChatGPT, Windsurf, Grok and other clients are in Connect Rymi to your AI client.
What you're approving
Signing in grants the connector everything you can do in Rymi, acting as you , including placing calls that spend credits, and (if you're an owner or admin) billing and API keys. The consent screen spells this out. Revoke any time from Settings.
Quickstart: API key
For clients without OAuth, CI, or scripts. Point at the same host and pass a secret key as a Bearer token.
{
"mcpServers": {
"rymi": {
"type": "http",
"url": "https://mcp.rymi.live/mcp",
"headers": {
"Authorization": "Bearer rymi_your_secret_key"
}
}
}
}// ~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"rymi": {
"type": "http",
"url": "https://mcp.rymi.live/mcp",
"headers": {
"Authorization": "Bearer rymi_your_secret_key"
}
}
}
}Get your API key from the Rymi dashboard → Settings → API keys. Publishable keys (sb_publishable_… and legacy rymi_pk_…) are rejected; they ship to browsers and never carry tool access.
Run it locally (npm)
Run the server on your own machine instead of the hosted endpoint. The @rymi/mcp package is a standalone MCP server you launch with npx. No clone, no build. It talks to the Rymi REST API directly and speaks stdio by default, which is what Claude Desktop, Claude Code, and Cursor expect for a local server.
{
"mcpServers": {
"rymi": {
"command": "npx",
"args": ["-y", "@rymi/mcp"],
"env": {
"RYMI_API_KEY": "rymi_your_secret_key"
}
}
}
}{
"mcpServers": {
"rymi": {
"command": "npx",
"args": ["-y", "@rymi/mcp"],
"env": {
"RYMI_API_KEY": "rymi_your_secret_key"
}
}
}
}The server reads your key from the RYMI_API_KEY environment variable (it exits immediately if it's missing).
Local server options
| Variable / flag | Effect |
|---|---|
RYMI_API_KEY | Required. Your Rymi secret key. |
RYMI_MCP_READONLY=1 | Restrict to read-only tools. Hides every mutating tool, including create_call, batch_call, and publish_agent. |
stdio only
@rymi/mcp speaks stdio. --transport http and RYMI_MCP_PORT were removed in 2.0.0: the hosted endpoint above serves HTTP, with OAuth and role-based tool gating the local server never had.
Mutating tools are on unless you turn them off
A local @rymi/mcp server exposes the full write surface, including publish_agent, create_call, and batch_call, by default, with RYMI_MCP_READONLY=1 as the single switch that hides them all at once. In shared or automated environments, run it with that flag set. The hosted endpoint instead filters by tenant role. See Mutating tools and what limits them.
Available tools
These are the tools your AI can invoke once connected. They cover the full lifecycle: discover → preview → build → attach a number → reach → observe → feed knowledge → audit → measure.
Agents
| Tool | What it does |
|---|---|
list_agents | Short summary per agent: id, name, persona, languages, STT/LLM/TTS models, publish and pause state. Call get_agent for prompts and the full config |
get_agent | Fetch a single agent's full config |
create_agent | Create a new voice agent. Supports supported_languages (multi-language), STT/LLM/TTS model selection, per-channel fallbacks, and custom endpoints |
update_agent | Modify an existing agent's config (same fields as create) |
clone_agent | Duplicate an agent (great for A/B variants) |
delete_agent | Permanently remove an agent |
apply_agent_changes | Dry-run a change-set. Validates without saving |
validate_agent_publish | Check if an agent passes all publish requirements |
preview_stack | Resolve the per-language model stack (STT/LLM/TTS) + blockers/warnings before saving |
enrich_company | Auto-generate a company description from a URL (write) |
generate_agent_draft | Generate a full agent config from a plain-English prompt (write) |
publish_agent | Make an agent live (write) |
get_pending_publish | On a phone line from Rymi, the latest held edit and its review status (pending, approved, rejected, stale) |
API tools (webhooks) & secrets
An agent's API tools (call_webhook bindings) are edited by name, so existing tools are never resent or lost. Edits are saved to the live agent and apply from its next call; each one is logged in list_agent_changes. Header values come back only as {{secrets.NAME}} references; any literal value shows as [redacted].
| Tool | What it does |
|---|---|
list_agent_tools | An agent's tool bindings, with header secrets redacted |
add_agent_tool | Add an API tool: https URL, method, purpose, fields, headers, timeout_ms (1000–10000) |
update_agent_tool | Change one API tool by name. Headers merge by name; omitted ones are kept |
remove_agent_tool | Remove one API tool by name |
list_tool_secrets | Secret names and the host each is pinned to (values are never returned) |
set_tool_secret | Store a secret for headers like x-api-key: "{{secrets.NAME}}". Only sent to its host (sensitive) |
Credential headers (authorization, x-api-key, *token*, *secret*) must use a {{secrets.NAME}} reference; a literal key is rejected.
Built-in tool switches
Built-in tools are switched on or off per agent. Connected-app tools (calendar, CRM lookup, Freshdesk tickets, WhatsApp, Telegram, shareable assets) are opt-in: connecting the app under Settings → Connectors makes them available, and an agent gets one only when you switch it on. Other agents are not affected. Changes are saved to the live agent and apply from its next call.
| Tool | What it does |
|---|---|
set_agent_tools | Switch tools on or off for one agent. Takes tool ids (lookup_customer, send_whatsapp_message, handoff_to_human, end_call, …) or the groups calendar (all five calendar tools) and tickets |
Example prompt: "Turn the calendar off for the Navratri invite agent and keep it on for Isha."
Public share link
One link per agent that anyone can open and talk to in the browser, billed to your workspace within the limits you set.
| Tool | What it does |
|---|---|
get_share_link | The agent's link: URL, limits, minutes used and remaining (link: null if none) |
set_share_link | Create the link or change it: enabled, minutes_limit (0–6000), max_call_seconds (60–1800), max_concurrent (1–10), calls_per_ip_hour (1–30). New links default to 30 min, 300 s per call, 2 at once, 3 per IP per hour (sensitive) |
regenerate_share_link | Issue a new URL; the old one stops working at once (sensitive) |
Discovery
| Tool | What it does |
|---|---|
list_llm_options | Full catalog of available LLMs and voices |
list_voices | Voices only, filterable by provider, model_id, gender or a query on label or name. Returns 50 a page (limit up to 500, offset) with total matches |
Calls (read-only)
| Tool | What it does |
|---|---|
list_calls | Browse all calls across the account |
list_active_calls | Calls currently in progress |
get_call | Details, participants, duration, and cost for one call |
get_call_summary | Post-call summary |
get_call_transcript | Full transcript |
get_call_recording | Recording metadata / playback URL |
get_call_queue_stats | Outbound queue statistics |
reprocess_call | Re-run post-call intelligence for a call |
list_calls_for_agent | Call history scoped to one agent |
Call control (write)
| Tool | What it does |
|---|---|
end_call | Hang up an in-progress call |
add_call_participant | Add a participant to a live call |
start_test_number_confirmation | Open WhatsApp confirmation for a shared-test-line destination. Returns a wa.me link and the exact message that phone must send. Does not confirm the number. Needs calls:write |
get_test_number_confirmation | Whether that number is confirmed for this user, plus any code still waiting. Read-only. Needs calls:read |
request_test_number_code | Send a 6-digit WhatsApp code to that number. Returns delivery, expiry, resend cooldown and entry_url, never the code. Needs calls:write |
submit_test_number_code | Return the secure page where the user types the code. Takes no code. Read-only. Needs calls:read |
A regular create_call through a personal assistant agent needs personal_call, and the owner's yes first. See Personal assistant calls.
create_call with test: true dials from the shared test number and returns phone_not_confirmed until get_test_number_confirmation reports confirmed: true. The request expires in 10 minutes. Each user can start 20 confirmations a day. One user at a time holds the pending request for a number. Confirmation is stored on the API key's user, the same user the test call checks. These tools cannot submit a code or mark the number confirmed. "Send code instead" stays in Studio, because a correct code can become a login number.
Example prompt: "Confirm +919876543210 for test calls, then place a test call to it from this agent once it is confirmed."
Enter a code securely
A verification code is a secret, and anything you type into a chat or a tool call stays in the transcript. So no Rymi tool accepts a code.
- The assistant calls
request_test_number_code. Rymi sends the code to that phone on WhatsApp. - The assistant gives you
entry_url. You can also get it again fromsubmit_test_number_code. - You open the link in a browser signed in to the same Rymi account and type the code there.
- The assistant calls
get_test_number_confirmationuntil it reportsconfirmed: true.
Do not paste the code into the chat, and do not let an assistant read it from your phone. A code lasts 10 minutes and allows 5 attempts. If request_test_number_code returns status: refused, reason says why and retry_after gives the seconds to wait. start_test_number_confirmation still works when codes are unavailable.
Both methods confirm a test-call destination only. They never change your login number or account, are not consent to future calls, and do not place a call. Ask before you call.
Task calls
Make a phone call from plain instructions through an opted-in agent (features.task_calls_enabled: true). See Use Rymi from your AI agent.
| Tool | What it does |
|---|---|
place_task_call | Call a number through an opted-in agent, do one task, and extract up to 20 return_fields. Places a real, billable call |
wait_for_call | Wait up to timeout_s (5 to 240, default 90) for a call and its post-call results, then return status, summary and fields. Returns done: false while the call runs |
Cards
Some tools come with an interactive card (an MCP App). In a host that renders MCP Apps, such as Claude, ChatGPT or VS Code, the card appears under the tool result. Hosts without MCP Apps support show the same JSON text result as before. Both the hosted server and the npm package serve the cards. Cards follow the host's light or dark theme, and the device setting when the host doesn't say.
| Card | Tools | What it shows |
|---|---|---|
Call card (ui://rymi/call-card.html) | get_call, create_call, place_task_call, wait_for_call | Live: status, an elapsed clock and the transcript as it happens. Finished: a recording player, sentiment, whether the goal was met and its score, the WhatsApp follow-up status, the summary, extracted fields and the full transcript |
Voice picker (ui://rymi/voice-picker.html) | list_voices | Every voice with a play button for its preview, a search box and a gender filter. When there are more voices than the first page, search and the filter ask list_voices again, so they cover every voice. Use sends your pick to the chat so the model sets that voice |
Agent card (ui://rymi/agent-card.html) | get_agent, create_agent, update_agent, clone_agent, validate_agent_publish, publish_agent | Whether the agent is live, a draft or has changes not yet live; who it answers as; its speech-to-text, LLM and voice models, with a play button for its voice; recording and consent settings; any publish blockers and warnings; and a link to the agent in Studio |
Calls list (ui://rymi/calls-list.html) | list_calls, list_calls_for_agent, list_active_calls | Each call's status, number, agent, time, how it ended, sentiment and duration. Live calls tick, and the card checks them with get_call every 5 seconds until they end. Click a row to read its summary. Load more fetches the next page |
Campaign report (ui://rymi/campaign-report.html) | get_campaign, get_campaign_report | The funnel from contacts to the campaign goal, goal rate, cost and cost per goal in credits, talk time in minutes, answer rate by hour with the best hour marked, outcomes, sentiment and top failure reasons. Refreshes every 15 seconds while the campaign runs, for up to 30 minutes. A link opens the campaign in Studio |
Campaign suggestions (ui://rymi/campaign-suggestions.html) | list_campaign_suggestions, accept_campaign_suggestion | Each suggestion's category, what to change and why, the numbers behind it, and the change itself: a policy change next to the campaign's current value, or the new agent setting. Accept calls accept_campaign_suggestion |
Usage and plan (ui://rymi/usage-plan.html) | get_usage_summary, get_plan | Voice minutes left, Studio AI and post-call intelligence quotas as percentage bars, any plan recommendation, and every plan with its monthly price and call discount. Change plan opens billing in Studio |
The cards fetch through the host with the same tools you can call: the call card polls get_call every 4 seconds while a call runs, stops after 20 minutes, and after the call ends waits up to 3 more minutes for the summary. Their only other requests load recordings and voice previews from Rymi storage and fonts from Google Fonts.
Some hosts draw a card but don't let it call tools. There the card shows the result it arrived with: a live call card or calls list stops its clocks and says it's a snapshot, and a finished call card still shows the transcript, outcome and recording, because get_call returns them.
Numbers, telephony & keys
| Tool | What it does |
|---|---|
list_numbers | All numbers on the account and their attached agent |
register_number | Register a number, optionally attaching it to an agent |
attach_number | Route an existing number to an agent |
remove_number | Remove a number from the account |
telephony_status | Whether a carrier is connected, and which |
list_telephony_numbers | Numbers available on the connected carrier |
get_phone_line | Your phone line from Rymi: status, KYC items still missing, call limits, release date when rental is unpaid |
request_phone_line | Start or resume a phone line from Rymi for IN or US and get the Studio onboarding link (write) |
list_publishable_keys | Browser-safe keys (prefixes only, never full secrets) |
Compliance: Do-Not-Call
| Tool | What it does |
|---|---|
list_dnc | List numbers on the Do-Not-Call registry |
check_dnc | Check whether numbers are blocked, without adding them (read-only) |
add_dnc | Add a single number to the registry (write) |
add_dnc_batch | Add up to 1000 numbers at once (write) |
remove_dnc | Remove a number, re-enabling outbound to it (write) |
record_contact_consent | Record voice consent (or a revoke) for a number, with the exact wording the person saw (write) |
Compliance settings
The workspace's operating country, callee-local calling hours, do-not-call check, voice consent, recording-notice default, and destination countries. Compliance is yours; these settings help. Blocked calls fail with destination_not_allowed, outside_calling_hours, or no_voice_consent. See Compliance settings.
| Tool | What it does |
|---|---|
get_compliance_settings | Read the settings, including every country preset (read) |
update_compliance_settings | Set operating_country ("IN", or null to clear it) and a partial rules object. A country fills every setting you have not already edited (write) |
Example prompt: "Set this workspace's operating country to India, and turn on hours and consent for single calls."
Accounts
Over OAuth, your role decides. A secret key lists a tool only when GET /v1/keys/self shows the tool's scope. These four ask for accounts:read or accounts:write. Those strings are not scopes you can put on a key, so a scoped key does not list them. A legacy key does. The REST routes use account_members:read, account_members:write, and workspaces:read. See Accounts.
| Tool | What it does |
|---|---|
list_accounts | Accounts you belong to, with your role (owner, admin, billing, or none) (read) |
list_account_members | People in an account and their roles (read) |
add_account_member | Add a person by email as owner, admin, or billing. An Owner can grant owner. An Admin can grant admin or billing (sensitive) |
remove_account_member | Remove a person. The last Owner stays (last_owner) (sensitive) |
Example prompt: "List my accounts, then invite arvind@example.com to Rymi as billing."
Workspaces
A connection starts in the workspace of the key or the person who signed in. To act somewhere else for one request, pass workspace (an id from list_workspaces) to any tool: every tool takes it. An account key may name any workspace in its account. A workspace key that names another workspace gets workspace_forbidden. To move the whole connection, send the Rymi-Workspace header on the hosted server, or set RYMI_WORKSPACE for the local package. See Workspaces.
| Tool | What it does |
|---|---|
list_workspaces | Workspaces you can act in, with account_id, account_name, account_role, and the effective role. current marks the one this request is in (read) |
create_workspace | A workspace in the current account. You need to be account Owner or Admin. It does not get the free monthly credit. parent must name a workspace in that account (write) |
update_workspace | Rename it, change operating_country, or set spend_cap_cents_monthly in credits on a workspace that is not the primary (null removes the cap) (write) |
delete_workspace | Delete an empty workspace: no agents, numbers, calls or campaigns. Account Owner only. The primary workspace returns cannot_delete_primary. workspace_not_empty includes blockers (sensitive) |
get_workspace_usage | Calls, minutes, and credits for one calendar month (UTC). Client and Viewer responses omit credits (read) |
list_workspace_members | People in a workspace and their roles (read) |
add_workspace_member | Add someone by email with role admin, editor, client, or viewer (default client). Someone without a Rymi account gets an invite email (sensitive) |
remove_workspace_member | Remove someone's access (sensitive) |
Example prompt: "Create a workspace called Acme Workspace in this account, cap it at 5,000 credits a month, and invite arvind@example.com as an editor."
Webhooks
| Tool | What it does |
|---|---|
list_webhooks | List registered webhook endpoints |
create_webhook | Register an endpoint for event delivery (write) |
update_webhook | Change an endpoint's URL, events, or secret (write) |
delete_webhook | Remove a webhook endpoint (write) |
Knowledge & history
| Tool | What it does |
|---|---|
list_knowledge_sources | RAG sources attached to an agent |
add_knowledge_source | Add a source from raw text or a URL |
delete_knowledge_source | Remove a source |
list_agent_changes | Recorded config-change history |
undo_agent_change | Revert a single change to its previous value |
Insight
| Tool | What it does |
|---|---|
get_usage_summary | Remaining voice minutes, Studio AI usage, post-call usage |
list_agent_templates | Published template catalog. Use a template's defaults to seed create_agent |
run_evals | Run the evaluation suite (synthetic or live) |
run_eval_suite | Run the eval suite across many agents at once and return the aggregate report |
list_eval_runs | Past evaluation runs for an agent |
get_eval_run | A single run with per-scenario scores |
Campaigns
See Campaigns for how campaigns, members and live lead intake fit together.
| Tool | What it does |
|---|---|
list_campaigns | Campaigns in the workspace, filterable by status |
get_campaign | One campaign with its policies and counters |
get_campaign_report | Outcomes, goal rate, cost and breakdowns |
list_campaign_attempts | Individual dial attempts and their outcomes |
list_campaign_suggestions | Improvement suggestions with their evidence |
list_contacts | Contacts in the workspace |
create_campaign | Create a campaign in draft: goal, calling window, retries, concurrency. Places no calls (write) |
import_campaign_contacts | Add contacts as members (JSON rows or CSV). Works on a running campaign (write) |
launch_campaign | Start dialing. Places real calls and incurs charges (sensitive) |
pause_campaign | Stop new dials (write) |
resume_campaign | Resume dialing (sensitive) |
accept_campaign_suggestion | Apply a suggestion as an agent draft or a campaign-policy patch (write) |
get_campaign_intake | The campaign's lead-intake settings and when it last received a lead |
set_campaign_intake | Create the lead-intake URL (shown once) or change assume_voice_consent / default_country (sensitive) |
rotate_campaign_intake | Issue a new intake URL; the old one stops at once (sensitive) |
disable_campaign_intake | Remove the intake URL (write) |
Billing & cost
| Tool | What it does |
|---|---|
estimate_call_cost | Estimate a custom stack's cost (stt_model/llm_model/tts_model + duration_seconds) before dialing (read-only) |
set_auto_recharge | Configure automatic balance top-ups (write) |
set_spend_alerts | Configure spend-threshold and low-balance alerts (write) |
get_plan | Current monthly plan, discount, renewal date, minutes left, the plans on sale, and whether the next plan would have cost less over the last 30 days (read-only) |
get_plan_studio_link | Studio billing link for upgrading, downgrading or cancelling a plan. Plan changes happen in Studio, never through MCP (read-only) |
Bug reports
| Tool | What it does |
|---|---|
report_bug | Report a bug with a title, description, optional steps, expected and actual behaviour, severity, area, and a linked call_id or agent_id. It goes to the Rymi team as pending_review before anyone works on it. See Bug Reports (write) |
list_bug_reports | The workspace's bug reports, newest first, with their status: pending_review, approved, in_progress, fixed (with fix_ref), rejected or duplicate. Filter by status (read-only) |
Multi-language agents
To build a bilingual agent, pass supported_languages (e.g. ["hi-IN", "en-US"]) to create_agent. Call preview_stack first to confirm the language set resolves to a valid model stack. Some languages may fall back to a different provider.
Example prompts
Once connected, talk to your AI client:
Report a bug
"The Hindi agent cut off mid-sentence on my last call. Report it as a bug."
The AI will chain: list_calls → report_bug (with the call_id)
Build an agent from scratch
"Create a receptionist agent for BrightSmile Dental. Warm and professional tone. Grab the company info from brightsmile.example.com."
The AI will chain: list_llm_options → enrich_company → generate_agent_draft → create_agent
Tweak an existing agent
"Switch my Support agent to the Charon voice and update it to GPT-4o."
The AI will chain: list_agents → apply_agent_changes → update_agent
Validate before going live
"Is my agent abc-123 ready to publish?"
The AI calls: validate_agent_publish → surfaces any missing fields or config issues
Call people who sign up on a form
"Make an outbound campaign for agent abc-123 that calls between 9 AM and 8 PM India time, and give me a URL my subscribe form can post to. The form collects consent and 10-digit numbers."
The AI chains: create_campaign → launch_campaign (after you confirm) → set_campaign_intake with assume_voice_consent: true and default_country: "IN", then hands you the URL once
Share an agent for testing
"Turn on a public link for agent abc-123 with 120 free minutes."
The AI calls: get_share_link → set_share_link with enabled: true, minutes_limit: 120
Make a phone call as one step of a task
"Call +919812345678 for Sharma Clinic using agent abc-123. Confirm tomorrow's 10 AM appointment and ask how many people are coming. Give me booked and party_size."
The AI finds an agent with features.task_calls_enabled: true via list_agents, then chains: place_task_call → wait_for_call, and repeats wait_for_call until it returns done: true
Voice agents creating voice agents
A deployed Rymi agent with this MCP wired in can respond to a caller in real time: "Set up a new booking agent for my dental clinic" → the agent calls create_agent directly. No human in the loop required.
Capability discovery
Before mutating an agent, ask the server what is allowed. Each of these is a read-only call, safe to chain from any prompt:
| Tool | Discovers |
|---|---|
list_llm_options | LLM, STT, TTS, and voice options the current tenant can use |
list_agents | Existing agents the connected key can see |
get_agent | The full config + runtime-affecting fields for one agent |
apply_agent_changes | Whether a proposed change-set would validate (dry-run, no save) |
validate_agent_publish | Whether an agent passes every publish requirement, with reason codes for any blockers |
list_calls_for_agent | Recent call history, including statuses and reason codes |
The AI should resolve "what can this agent do" by calling get_agent + validate_agent_publish rather than guessing from the prompt. Errors carry machine-readable codes you can branch on; treat the human-readable message as display copy and the code as the routing key.
Mutating tools and what limits them
The write tools have real-world effects: publish_agent goes live to end users, and create_call / batch_call dial real numbers and spend credits. create_call with test: true marks that call as a test and uses the shared test number when you omit from_number. The destination must be confirmed on WhatsApp first (start_test_number_confirmation, then get_test_number_confirmation). What restrains them depends on which server you're talking to.
Hosted (mcp.rymi.live), over OAuth: tools are filtered by your role. Account Owner maps to owner. Account Admin and workspace Admin map to admin. Anyone else maps to member.
| Your role | What's exposed |
|---|---|
owner / admin | Every tool, including account-scoped ones (billing, keys, numbers, DNC) |
member | Agent-scoped tools only; account-scoped writes are not registered at all |
Over a secret key, hosted and local both call GET /v1/keys/self once and list only tools whose scope the key holds. A legacy key lists every tool. Calling a billing write from that list still returns scope_missing. See API keys and scopes.
Local (npx @rymi/mcp): no OAuth role filter. The key's scopes still apply. RYMI_MCP_READONLY=1 hides every mutating tool on top of that. That flag is set by whoever starts the server.
Role is not a sandbox
Role filtering bounds the tool surface, not the credential. An API key or OAuth token remains valid against the REST API directly, where the same permissions apply as any other client. Treat MCP access as equivalent to handing over that credential. For a narrow job, use a workspace key with only the scopes that job needs.
Every tool carries MCP annotations, so a client can tell them apart before it asks you to approve a call. Tools that only read are marked readOnlyHint. Tools that delete, remove or end something (delete_agent, remove_number, end_call and the rest) are marked destructiveHint. The hints are advisory and don't replace the role and scope limits above.
Carrier connect/disconnect and publishable-key creation/revocation are intentionally not available over MCP. They enter credentials and change standing configuration, so do them from the dashboard.

