Find my list of website visitors from July and tell me how many rows it has.Here the assistant lists your lists to find the right one, then reads its row count. Most changes need an id, so expect the assistant to look one up before it acts. When a client connects, the server also sends short setup notes on how lists, campaigns and the Smart Actions feed fit together; not every client reads them.
Until October 2026 the server exposed a small set of natural-language agent tools instead, one per feature area, such as
prospect_search, campaigns and inbox. Those tools no longer exist, and calling one answers that your tool list is outdated. If your client still shows them, refresh the Fuse connector in your client’s settings (reconnecting alone may not reload the list), or remove Fuse and add it again.The tools
The tools are grouped here by area; the Tool reference lists every one of them with a one-line description. The names are the server’s own and the list grows as Fuse adds features, so treat the examples as a guide and your client’s tool list as the source of truth. The server doesn’t announce changes to the list, so refresh the server in your client to pick up tools added since you connected. In some clients, ChatGPT among them, reconnecting alone doesn’t reload the list.What isn’t exposed
- Tools that take a password or an API key are left out, so credentials are never typed into a chat app. Connect or test an SMTP mailbox and save a Deep Research provider key in the Fuse app instead.
- Tools Fuse hasn’t approved for assistants yet. Each tool is added to the MCP server on purpose, after a review, so a new feature can reach the Fuse app before it reaches your assistant.
- Nothing that changes billing, plans or payments is exposed, and neither is removing team members, deleting users, accounts or teams, transferring ownership, or managing Fuse API keys.
- Some jobs have no tool yet. Nothing uploads a file (the import tools take rows the assistant already has), and nothing places a call. Use the app, or the REST API for uploads.
Deep Research
Deep Research runs take minutes, not seconds, so two tools handle them asynchronously:deep_research_startbegins a run and returns arunIdimmediately.deep_research_statusreports progress for arunIdand returns the report when the run completes.
deep_research_status in the meantime. Any prospect list the run produces also appears in your lists in the Fuse app, and the report is available there too.
Credits
MCP calls spend the same credit balance as the app and the API, and a tool costs what the same action costs in the app: searches, saves, enrichment, smart columns, signal agents and Deep Research runs bill as described in Credits & billing. Reads such as listing lists, reading rows and checking status are free. Most tools that charge say so in their descriptions.get_credit_prices returns current prices for searches, saves, enrichment, email validation and campaign sends, and signal agents, smart columns and scrapes state their prices in their own tools.
save_people_to_list saves up to 25 people per call, found by search_people or named in the chat. It matches each one to their profile first, and those lookups are free; it charges 1 credit per person it adds to the list, and nothing for people already in it. Its answer says what it charged in creditsCharged.
Five research tools that call outside providers charge per call: 10 credits for search_or_scrape_web_with_firecrawl_api_tool and google_maps_search_tool, 20 credits for google_jobs_search_tool, and 100 credits for lead_researcher_tool and ask_fuse_brain. The charge is taken before the provider is called, so a search that finds nothing still costs its price. They cost the same when the assistant in the Fuse app uses them.
Every tool call needs a balance of at least 100 credits, because several tools charge after their work finishes. Below that, calls return “Insufficient FuseAI credits. Add credits to keep using these tools.” The check itself costs nothing, and deep_research_status is exempt so that a run that spends your balance down can still return its report.
Behavior and safety
- Every tool runs with your permissions. It sees and changes only what your Fuse user can.
- Changes happen straight away. There is no staging and no undo. A tool that sends an email or a LinkedIn message, approves a scheduled send, executes a Smart Action, pushes records to your CRM or posts to Slack does so as soon as it runs.
- Merging duplicate contacts can’t be undone. Contacts are shared, so a merge changes them for everyone who has them. The merge tool merges only the groups the assistant names, so ask it to show you each group, and the contact it keeps, before you agree.
- Creating a campaign never launches it. A new campaign is a draft that sends nothing. Launching is a separate tool call, and the launch tools tell the assistant to launch only when you ask and to confirm with you first. Say “don’t launch it” when you only want the draft.
- Tools carry hints for your client. Each tool has the standard MCP annotations:
readOnlyHintwhen it only reads, anddestructiveHintwhen it deletes, removes, stops, disconnects or overwrites something. Sending messages, launching a campaign and pushing to your CRM change things outside Fuse without deleting anything, so they are not marked destructive. Keep your client’s approval prompt on for every tool that isn’t read-only. - Text written by others is marked. Tools that return emails and replies, LinkedIn messages and invitations, call transcripts, Smart Action details, web pages or Fuse Brain answers put that text between markers that tell your assistant to treat it as data and not to follow instructions inside it. This helps against prompt injection, where an email or a web page tries to give your assistant orders; keep approval on for write tools all the same. The Tool reference marks these tools third-party text.
- Long jobs run in the background. Imports, enrichment, smart columns and Deep Research answer straight away and finish in the background. The assistant can check on them, and the results also appear in the Fuse app.
Results and errors
Tools answer with text. Almost always it is JSON, such as{"balance":8250} from get_credits_balance or a data array with pagination from a list read; download_list_csv_tool returns JSON whose data field holds the CSV text. For a list over 20,000 rows it starts a background export and returns its jobId instead; download that file from the Fuse app.
Tools marked third-party text in the Tool reference wrap a successful result in markers, with a random 16-character id that changes on every call:
isError is never wrapped.
When Fuse rejects a call, for example because an argument is missing, an id is unknown or a campaign isn’t ready to launch, the result is flagged as an error (isError: true) and its text is JSON with an errors array. Each entry has a code, and a message when there is more to say, so the assistant can correct the call and try again. Eight tools that run in the MCP server itself, such as ask_fuse_brain, lead_researcher_tool and the Deep Research pair, report problems in their normal text instead, without isError:
Tools served by Fuse’s backends have 60 seconds to finish.
search_people and prospect_search_tool stop sooner: a search that has not finished 28 seconds after the call answers with the code SEARCH_TIMED_OUT, and nothing is charged. When the search’s own lookups of company names, email addresses or phone numbers ran out of time instead, the code is SEARCH_LOOKUPS_TIMED_OUT, nothing is charged, and searching again gets further because the lookups that finished are kept. Trying again after SEARCH_TIMED_OUT can work too, because the company names it looked up are kept; if it times out again, narrow the search with a job title or a location. The research tools that run in the MCP server itself have no server-side limit, so your client’s own timeout applies. The work that takes longer, such as imports, enrichment, smart columns and Deep Research, is designed to answer at once and continue in the background, so the limit matters mostly for very large synchronous reads and exports.
Rate limits
MCP usage is limited per user, independently of your REST API rate limits:
Every request to the server counts, including the connection handshake and the tool listing; a batch of messages sent in one request counts once. A task that chains several tools costs one request per call, so a long task can take a dozen or more. Beyond a limit the server returns
429 with the code RATE_LIMIT_EXCEEDED, and your assistant should back off and retry. Some tools also keep the per-user limits they have in the app. search_people and prospect_search_tool, for example, share the people search limit of the app’s search page, 30 searches a minute and 2,000 a day per user; past it they answer with the code PROSPECT_SEARCH_RATE_LIMITED.
Client limits
With close to 300 tools, the tool list is large: around half a megabyte of descriptions and input schemas. Clients handle a list like this differently:- Cursor has, in some versions, passed only the first 40 tools across all connected servers to its agent. If Fuse tools are missing, turn off servers or tools you don’t need.
- Any client that loads the whole list into its model’s context spends a noticeable share of it on Fuse’s tool descriptions. Clients that let you switch individual tools off help here. Claude Code searches the list on demand instead.