Write access in MCP
Write access in MCP means an MCP server exposes tools that change state in a connected system — creating, updating, pausing or deleting records — rather than only reading and reporting data.
Most early MCP servers were read-only: fetch rows, summarize a report, search documents. Genuinely useful, but every conversation ended the same way — "now go into the dashboard and do it." Write tools close that loop. The model proposes a change and then executes it in the source system through the platform's official API, so analysis and action happen in one place instead of two.
MCP describes this with tool annotations: readOnlyHint, destructiveHint, idempotentHint and openWorldHint. Clients use them to label tools in the interface and to decide when to ask for extra confirmation. It is worth being precise about what they are, though — they are declarations made by the server, not enforcement. They tell you what a tool claims to do; they do not stop it. Trust in the server still does the heavy lifting.
Responsible write access has a recognisable shape: least-privilege OAuth scopes; per-customer isolation so one tenant's tools can never reach another's account; explicit account scoping, which in advertising means using the correct login-customer-id when working under an MCC or agency manager account; a dry-run preview and an approval step before execution; caps on how large a single change can be; and an audit log. There is also an engineering reality — API quotas, validation errors and partial failures all show up on the write path, so a mature server returns readable errors instead of leaving half-applied changes behind.
Write access is the whole point of Opus Growth. Across a 221-tool catalogue you can create and pause campaigns, set budgets and bids, manage keywords and negatives, build audiences and push creatives on Google Ads (69 tools), Microsoft Advertising (30), TikTok Ads (25) and LinkedIn Ads (18), plus Google Tag Manager, GA4, Search Console and YouTube. Meta Ads (27 tools) is in final platform review and Business Profile (7) is waiting on Google's API quota approval, so neither is live yet. Every write is previewed as a dry run, applied only on your approval, scoped per customer, MCC-safe and logged.
Frequently asked questions
No. Many are read-only by design and can only fetch or summarize data. You can tell by looking at the server's tool list and its annotations — read-only tools are usually marked as such by the client.
Safety comes from the controls around it: narrow OAuth scopes, an approval gate on every write, per-account isolation, limits on how far a single change can go, and an audit log. With those in place the model proposes and you decide.
The failure modes are ordinary and expensive: the right change made in the wrong account, a budget read as 400 instead of 4,000, a pause applied far too broadly, or an audience deleted with no undo.
