One MCP server.
Every tool your agent wishes it had.
Social data, lead enrichment, web scraping, a real Postgres database, memory, schedules, queues and an inbox. One URL, plugged into the agent you already run. An agent with no hands is just a chat window.
Copy the setup prompt.
Your agent does the rest.
Fetch and execute the appropriate instructions to set me up for superagnt_ from https://mcp.superagnt.com/agent-setup/prompt.mdA real Postgres, born with the workspace.
A folder of markdown gets an agent surprisingly far, until the third time it re-reads everything to answer one question. Everything yours collects lands in an actual database it can create tables in, migrate, and query with SQL. Not a key-value toy: the memory layer real work sits on.
- »tables, SQL, migrations and CSV loads: tools, not a dashboard
- »every scrape, enrichment and job result lands somewhere queryable
- »10 GB to start, isolated per workspace
data_instagram_*posts · profilesdata_agnt_people_*enrich · finddata_web_*scrape · crawlagnt_db_insert- @lena.buildsinstagram001
- sarah chenpeople002
- northloop.comweb003
- @marco.runsinstagram004
- iris waltonpeople005
select count(*) from leads;5The work, on a wall. Not in a CSV.
One tool call pins a query as a live widget, and the widgets add up to an operation. Below: a sales agent's own board, the funnel it is working, the pipeline it moved overnight, and the two decisions it parked for its founder. Shared with a link; nobody asks for an export again.
Pipeline
sales-agnt · week of aug 31Sales Funnel
this week's moversOutreach Conversion
last 30 daysThis Week's Motion
replies + demos · dailyPipeline Growth
open pipeline $k · weeklyWaiting on You
founder gateAnnual at $18,000/yr, two months free vs monthly. Contract drafted; goes out the moment you approve.
Signs this week at $12,000 with 12% off list. Margin still clears the floor.
Overnight Run
02:00–08:00 · autonomousRun it from your phone. Or the channel your team lives in.
Bind your agent to a phone number and to a Slack channel. After that you work with it the way you work with a person: text it from the airport, read the answer in the same thread, and when it needs a yes before something goes out, that yes is one reply. An agent that only talks in a terminal is an agent you have to go sit in front of.
- »bind a phone number or a Slack channel, nothing to install
- »answers land in the same thread, on your phone or in the channel
- »approvals arrive as a message. one reply releases the send
A 200-email day. Two of them reach you.
Connect Gmail or Outlook once and your agent gets the mailbox as tools: search, read, draft, reply, send, label, file. Triage and escalation come built in, so you are not writing rules for what a receipt looks like. Most inbox agents get expensive because they pull every message into context. Yours opens the ones that matter and leaves the rest alone.
- »triage and escalation built in: what needs you gets escalated, the rest gets handled
- »it reads what matters and skips the rest, so a 200-email day stops costing tokens
- »Gmail and Outlook connect once: search, read, draft, reply, send, label, file
The data your agent cannot get from the open web, one call away.
Profiles, posts, followers, ad libraries and keyword search across the big social platforms, plus any page on the web as clean text. Nine sources on one key, structured JSON whichever one you hit. No developer accounts to apply for, no scrapers to keep alive every time a layout changes.
- »search by role, company and city
- »find a work email from a name and a company
- »verify it is deliverable before you send
- »find a mobile number when the email gets ignored
- »turn a profile url into a full record
- »one person, or a whole list in one pass
- »search by industry, headcount and geography
- »find the companies that look like your best ones
- »turn a domain into size, industry and location
- »pull every public address on that domain
- »watch hiring, funding and tech signals
- »one domain, or a whole list in one pass
Priced per call, in credits, on the same key as everything else.
» the upstream bills are ours
What one agent learns, every agent knows.
Your agents share one memory. What the research agent learns on Monday, the outreach agent uses on Tuesday, because both write to the same graph of people, companies and projects. The usual alternative is one memory silo per platform: facts you cannot read, cannot move, and have to teach again from scratch the day you switch tools.
- »any harness, any agent, one graph and one set of facts
- »people, companies and your business in one place you own and can export
- »ingest a document once; every agent retrieves it by meaning
Knowledge graph
shared by every agent you run
Hand it forty thousand rows. It finishes without you.
A data job runs outside the chat. Your agent writes the recipe once (reformat this row, verify this email, enrich this signup) and the pipeline runs it as code over the whole list. Nothing to babysit: no context window to blow, no session to hold open, and items that fail get retried. Push those same forty thousand rows through chat instead and you pay a model to think forty thousand times.
- »it runs on the pipeline, not in chat: no context window, no open session
- »an item that fails is retried, and the job resumes where it stopped
- »2,000 email checks cost 2,000 lookups, not 2,000 model turns
agnt_data_job_submitcallback received- ✓08,412reformatted
- ↻19,077retried, ok
- ✓33,190reformatted
It works while you sleep. The output is waiting at 9am.
Three standing primitives give your agent a pulse when nobody is typing. Schedules fire on a cron line. Queues drain themselves overnight. Webhooks wake a session the moment something happens in the outside world. Without them your agent only moves while you are sitting there typing at it.
» every firing is a fresh session, and every session leaves a receipt you can read
daily-digest0 7 * * * · schedule→ session started · 07:00lead-queuequeue · 214 items→ 12 items processedstripe-eventswebhook · inbound→ payload delivered to session
Every agent gets its own custom toolkit.
Mint one MCP server per agent and hand it only the tools its job needs. The researcher gets the data sources. The outreach agent gets the inbox and the database. Nobody gets what it does not need. The alternative is one over-powered agent with every tool switched on, and no sentence in a system prompt stops that agent from using them.
digs through X, the web and your knowledge base
finds people, sends the email, logs the touch
reads the numbers, renders the recap
- »the researcher gets data and knowledge; it cannot send an email
- »the outreach agent gets the inbox and the db; it cannot scrape the web
- »revoke one agent's server at 2pm and the other two keep working
When one agent isn't enough, deploy a fleet.
The toolkit's ceiling: the agent you are talking to can build agents of its own. Deployed agents run on our managed runtime with their own prompts, tools, schedules and channels, and everything they do is observable, down to the individual tool call.
- »your harness creates the agent; ours keeps it alive
- »schedules, channels and a knowledge base wire in at deploy time
- »sessions, tool calls and spend are all queryable after the fact
- livepipeline// pipelineenriching lead→ company + 6 signals
- liveinbox// triagedrafting reply→ inbound from acme.co
- liveresearch// signalsgathering intel→ 12 sources
- liverevops// revopsflagging for approval→ refund $480
- livescheduler// schedulerscheduling meeting→ tue 14:30
Build your toolkit. Pay for the hands you use.
A $19 base covers the connection, every data source and unlimited vendor connections. Click the modules your agent actually needs and the price follows you down the page. Calls meter in cents against a balance your agent can read mid-run, and a call that fails is free.
The builder base. Always on.
» modules · click to add · tiered ones step price for volume
Every module. Top tier of each.
Stop choosing. All six modules at their highest tier, $25/mo of credits, 600 rpm.
free / pay-as-you-go$0 · 60 rpm · 3 connections · $0.50 starting grant. The first call always costs nothing.
A credit is a cent.
One balance per organization, split into a data wallet for API calls and an AI wallet for the model tokens a deployed agent burns. The split is there so a runaway agent cannot drain the wallet your lookups run on.
1 credit = $0.01You pay per call, and only when it works.
Every billable response carries what it cost and both remaining balances in its meta block, so the agent can read its own spend mid-run. A call that fails is not charged.
meta.costCentsCommit to credits, discount every call.
Credit packages ride the same subscription as a commitment dial: the more you commit to monthly, the deeper the per-call discount, from 10% at $29 up to 25% at $999. Modules never change the discount, and credits never switch on a module.
10–25% offPer-source pricing is published.
Each source prices its own calls: a web fetch is a fraction of a cent, an enrichment that walks several providers costs more. The number for a given source lives on that source's page.
/tools/<source>One URL, every workspace
The same address for everyone. Your key decides which workspace it resolves to and what it can reach.
https://mcp.superagnt.com/mcpMeet superagnt_
Same toolkit, with an operator already sitting behind it. You describe the job in plain language. It builds the agent, wires the tools, runs on a schedule and reports back.
Answered before you paste a key.
01How does auth work?
Two ways in. Claude Desktop speaks OAuth natively: you add the endpoint as a custom connector, approve it once, and that client gets its own credential, so no token is ever pasted or revealed. Every other client sends your workspace token as a bearer header. The block at the top of this page prints the exact snippet for whichever one you run.
02What counts as a credit?
A credit is a cent. Data credits pay for data API calls, connection proxy calls, database usage and webhook events. AI credits pay for the model tokens and runtime a deployed agent burns. They are separate wallets on purpose, so an agent that runs long on tokens cannot empty the balance your lookups draw on. Credits come off on success, and every billable response reports what the call took and what is left.
03Are there rate limits?
Yes, per workspace, on a rolling 60 second window. Every key in a workspace shares one bucket, and going over returns a 429 carrying a retryAfter that tells you how long to wait. The ceiling scales with your plan, so a heavier plan buys throughput as well as credits.
04Is my data isolated?
Yes. Keys, database, files and memory are all workspace scoped. The Postgres database is provisioned per workspace with no connection string handed out, and memory never crosses a workspace boundary. Your data and your agent's memory are yours, and they are not used to train foundation models.
05Can my agent enable more tools itself?
Yes, and that is the intended path. A fresh connection starts on a small default surface so the tool list stays readable. From there agnt_tools_search finds what else exists and agnt_tools_enable switches a family on, so the agent widens its own toolset mid task instead of waiting for you to tick a box. agnt_tools_disable takes it back off.
06Do I need to be a developer?
No. For most clients the wiring is one line you paste into a terminal or a config file, and for the agent clients there is a prompt you can hand to the agent so it does the wiring for you. After that you talk to your agent, not to us. If you would rather not run MCP at all, every tool is also a REST endpoint you can call with the same key.