Sign up

Falcon Studio

Your team, building agents that work together.

Everyone who knows your business shapes the same agents: each person owns their part, drafts are reviewed, and one version goes Live. Then the agents pass work to each other in workflows.

Sign up See how it works

In a private pilot with a small number of companies. Read our Privacy Policy.

Many people, one agent Four people, Dana in support, Luis in sales, Priya in compliance and Omar in operations, each fill the part of one service agent they own: role and purpose, tone and style, knowledge, escalation, tools and actions, and a skill to book a service visit. Each part shows its owner with a lock. The agent then moves from Draft to In review to Live. D Dana Support lead P Priya Compliance Service agent Instructions and skills Draft In review Live Role and purpose D Dana Tone and style L Luis Knowledge D Dana Escalation P Priya Tools and actions O Omar Skill · Book a service visit O Omar L Luis Sales O Omar Operations
Each person fills the part they own, and the agent goes Live only after review.

Inside an agent

What goes into an agent

Think of it as a new team member your whole team trains together.

A service agent sits in the middle, with its versions: Draft, Dev and Live. Only Live answers people. Nine parts slot into it, one at a time:

  1. Instructions, its job description: sections by topic, each written by the person on the team who knows it best.
  2. Skills, training for specific situations: company know-how it uses only when it is needed.
  3. Tools, what it is allowed to do in your apps: look things up, change something, or act outside the company, with a person's approval first where you want it.
  4. Knowledge, the binder of documents it looks things up in: help articles, policies and files it searches for answers.
  5. Memory, what it remembers about each person: preferences, facts, summaries and moments, each kind with its own setting for whose memory it is.
  6. Voice, how it sounds on calls: off, listen only, or listen and talk, with a feminine-sounding and a masculine-sounding voice that each person picks from.
  7. Identities, who it works on behalf of: the people and places it acts for, each with its own email, calling address and, if you want one, a phone number.
  8. Channels, where people can reach it: texts, RCS, email, calls, website chat and apps.
  9. Model and limits, its brain and its spending limit: the AI model it runs on, daily and monthly spend limits, and checks.

When every part is in, Publish runs the checks, they pass, and the agent moves from Draft to Dev to Live.

An agent is its instructions, skills, tools, knowledge and memory, with a voice and identities to act for. Your team shapes each part; only the Live version talks to people.

Built together, working together

People across the company shape the same agent, each in the part they know best. Then one agent can answer as many of your people and places.

  • Each person owns a part: role, tone, knowledge, escalation, tools or a skill.
  • Changes stay drafts until they are reviewed, then one version goes Live.
  • One Live agent can speak as many people and places, each with its own number, email and chat.

One agent, many identities

One agent, many identities, fed by every entry point Entry points sit along the top, each a small node with an icon: POST, an incoming web request, with a Request map node under it that maps the request's fields for the agent; Email; SMS; WhatsApp, marked coming next; website Chat; Call; Connector, a trigger from a connected app such as a new CRM record; and Event, from your systems. Their lines run down into an agent, shown with a spark, which sits above a shelf of identities, shown with person icons: Maya at the Downtown store, the service desk shown with a headset, and the Lakeside location shown as a storefront, each with its own phone number, email address and website chat. A POST request arrives, passes through the request map into the agent, Maya's card attaches and the agent is labelled Acting as Maya, chosen as the customer's usual contact; the reply goes out from Maya's email address. A text arrives on SMS while Maya is still attached, and the service desk attaches as next in turn, so the agent acts as two identities at once; it came in on the service desk's number and the reply goes back to it. Then a chat arrives from website Chat and Lakeside attaches as the store the customer wrote to, and replies in its website chat. After each conversation the identity goes back to the shelf. Incoming Coming next POST Email SMS WhatsApp Chat Request map fields → agent Call Connector Event Agent Service agent Identities People and places your agent speaks as Identity Maya Downtown store (555) 010-2231 maya@shop.example Website chat data-identity="maya" Acting as Maya Rule: their usual contact Identity Sam Service desk (555) 010-4400 sam@help.example Website chat data-identity="desk" Acting as Sam Rule: next in turn Identity Lakeside Lakeside location (555) 010-7788 lake@shop.example Website chat data-identity="lake" Acting as Lakeside Rule: the store they wrote to From maya@shop.example From (555) 010-4400 From Lakeside’s website chat Next conversation Next conversation Incoming POST Email SMS Request map maps fields WhatsApp Coming next Chat Call Connector Event Agent Service agent Identities People and places it speaks as On a call Maya Downtown store Acting as Maya On a call Sam Service desk Acting as Sam On a call Lakeside Location Acting as Lakeside Rule: their usual contact Rule: next in turn Rule: the store they wrote to From maya@shop.example From (555) 010-4400 From Lakeside’s website chat
Identities are the people and places your agent speaks as, and each can have its own phone number, email address and website chat. Each conversation gets the right one.

Every way in knows who it is for

Each identity gets its own addresses, any system can send its requests through a request map, and the apps you connect can start work. Whatever arrives already says who it is for.

Every identity gets its own ways in.

One agent, the Front desk, acts for two identities: Dr. Rivera, a person, and the Downtown store, a place. Each identity has its own set of addresses, drawn as cards: a receiving email such as k7f3q9xm@falcon.com, a SIP address, a phone number marked optional, a public link and a website chat snippet. A new email arrives at Dr. Rivera's address, the message runs up to Dr. Rivera, and the agent shows Acting as Dr. Rivera. Then a call arrives at the Downtown store's SIP address and the agent acts as the Downtown store. Whatever arrives already says who it is for.

Each agent acting for an identity gets its own email, SIP address and link, with a phone number if you want one.

Map any request into your agent.

A POST request arrives with a signature header and a JSON body with four fields: called_number, caller_phone, recording_url and plan. The signature is checked first. Then arrows carry each field into the request map in turn: recording_url becomes the start field recording; called_number looks up the identity the agent acts as, the Downtown store; caller_phone matches the customer; plan becomes a custom field; and the identity is sent on as the Falcon-Identity header. A fixed value is typed into the map: topic equals Call report. The map feeds a workflow whose first stage has two steps side by side, Summarise and Update CRM, both receiving the same start fields. The caller gets back 202 Accepted with a run ID.

Map body fields, query values and headers, or type fixed values, into start fields, identity and customer fields, and headers, with a signature check first.

Connected apps start work too.

Connected apps are drawn as tiles, each with a line into one workflow and its agent. Gmail and Google Calendar, a new email or an event coming up, marked Yours because the account belongs to one person. HubSpot and Salesforce, when a CRM record changes, marked Company. GitHub, on a push, pull request or issue, marked Company. Slack, a message in a channel, marked Company. A schedule, every weekday at 9:00. Another workflow, when it finishes. Your own systems, sending an event such as order.refunded, which waits for approval before it can start work. Outlook mail and calendar and Microsoft Teams are marked coming next. In turn, each live tile lights, a message runs along its line, and the workflow shows what started it. The event from your own systems first changes from Needs approval to Approved.

A connected account can belong to the company or to one person. You choose which of its events start work, and new events from your own systems wait for approval first.

Every channel your customers use.

People reach your agents the way they already talk: email, calls, texts, chat, apps and messaging.

  • Email

    Each agent and identity gets its own address, and people write to it as they would to anyone.

    • Forward from your own mailbox, or use your own domain
    • Sends from your company’s domain
    • Replies stay in the same thread

  • Phone calls

    People call, and the agent answers in a voice of its own.

    • A SIP address, plus a phone number if you want one
    • Listen only, or listen and talk
    • After hours, an AI voicemail takes a message

  • Texts

    Text messages to and from your company’s own numbers.

    • Your own Twilio numbers
    • STOP is honoured on every number
    • Follow-ups only to people who agreed

  • RCS

    Branded messages with your logo and name, from a sender the carriers approve.

    • The company’s shared sender, or one for each identity
    • Preview the registration before you apply

  • Website chat

    A chat window on your site, from one snippet you paste in.

    • Styled to match your brand
    • A snippet for each identity

  • Public link

    A page anyone can open to talk to the agent.

    • Share it in email, posts or print
    • You set who may use it and how often

  • iPhone and Android apps

    The same chat window, inside your own app.

    • Add your app’s bundle ID or package name
    • Shares the rules you set for website chat

  • Slack and Telegram

    Agents send and receive messages there through messaging tools.

    • Slack through your workspace, Telegram through a bot
    • Company rules for each channel

  • WhatsApp Coming next

    Messages on WhatsApp, as with texts.

    • Through your own Twilio account
    • Approved templates outside the 24-hour window

Channels for your systems.

Email, webhooks and EventBridge events all run through the request mapper: body fields, headers and fixed values become start fields, the identity to act as, and the customer.

  • Email from your systems

    Order confirmations, alerts and forms that arrive by email start work.

    • Matched by recipient, sender, subject or keywords
    • Mapped into a workflow or an agent
    • A company catch-all at the end

  • Webhooks

    Any system can send a signed request, and your systems hear back when a run finishes.

    • Signatures checked first (Twilio, Zoom, Stripe and more)
    • Each map has its own address you can rotate
    • Signed webhooks out, with retries and a delivery log

  • Amazon EventBridge Your own AWS

    For companies running Falcon in their own AWS account: events from that account start work, privately. Everyone else sends the same events as signed webhooks.

    • Private sources reachable only from your account
    • Events can wait for approval before they run
    • Same event names and approvals as a webhook source

Agents that work together

A workflow is your team’s playbook. Agents hand work to each other, side chains start on their own, and people step in only where you ask.

  • Each step has an owner on your team, shown in its corner.
  • Side chains run beside the main steps, like booking a call.
  • A step can wait for a person, and nothing runs until they approve.
Agents that work together One run of a service workflow, step by step. A new message to the service desk starts it. An hours check finds the desk open; after hours it would reply with opening times. The front desk agent matches the person to the CRM and routes them to service rather than sales. The service agent handles it, and a side chain starts because the person wants a call: book a call, then send a calendar invite. Another side chain would tell the service manager about a complaint. A refund waits for Priya to approve it, and a tick appears when she does. Taking turns, which is coming next, picks Sam as the next person to follow up. The reply is sent from Sam at the service desk and the run finishes with a signed callback. Each step shows a small avatar of the person on the team who owns it: Omar, Dana, Priya or Luis. Start New message to the service desk O Hours check Open, so carry on O After hours Reply with opening times Agent Front desk D Matched to CRM Sales Service Agent Service agent D Wants a call Book a call Calendar invite sent Complaint Tell the service manager Side chains Waiting for you P Approve refund Priya decides Approve Decline Taking turns L Coming next Maya Sam Jo Next in turn: Sam End Reply sent from Sam · Service desk O Run finished, callback signed Start New message to the service desk O Hours check Open, so carry on O After hours Reply with opening times Agent Front desk D Matched to CRM Service Sales Agent Service agent D Side chains Wants a call Book a call Calendar invite sent Complaint Tell the service manager Waiting for you P Approve refund Priya decides Approve Decline Taking turns Coming next L Maya Sam Jo Next in turn: Sam End O Reply sent from Sam · Service desk Run finished, callback signed
One run: the desk is open, the front desk routes to service, a call is booked on the side, Priya approves a refund, and the reply goes out from the service desk.

Two sides, one clean room Coming next

Sometimes each side of a deal has its own agent, such as a dealership buying leads and a lead source selling them. The agents never talk directly: they negotiate through a Falcon clean room that passes on only what both sides agreed to share.

  • Neither agent sees the other side’s instructions, memory or customer data.
  • Shared: vehicle interest, price range and timing. Held back: name, phone and email.
  • Contact details are released only when both sides say yes.
Two sides, one clean room A buyer's agent for a dealership and a seller's agent for a lead source never talk directly. Each sends messages into a Falcon clean room in the middle. The buyer's agent offers 40 dollars, with a private ceiling of 55; the room passes on only the offer. The seller's agent counters at 48, and the room passes that back. The room shares the customer's vehicle interest, price range and timing, and holds back their name, phone and email behind a lock. They agree at 45, and only then, with both sides agreeing, is the contact released and the lock opens. Neither agent sees the other side's instructions, memory or customer data. Buyer’s agent A dealership Own rules and data Seller’s agent A lead source Own rules and data Falcon clean room Shared Held back Vehicle interest Price range Timing Name Phone Email Offer $40 · max $55 Offer $40 Counter $48 Counter $48 Negotiating Agreed at $45 Both agree → contact released
An offer goes into the room and only the offer comes out, without the buyer’s private ceiling. A counter comes back the same way, and the contact is released once both sides agree.

How steps fit together

  • In order

    Each step starts when the one before it finishes.

  • At the same time Coming next

    A step splits into two that run together, then they join again.

  • Choose a path

    A check sends the work one way or the other, such as open or after hours.

  • Side chain

    A signal, such as “wants a call”, starts extra steps while the main line carries on.

  • Wait

    Pauses for a set time, or for a person to approve, then carries on.

  • Repeat Coming next

    Runs steps again, such as a follow-up every few days, until someone replies or a limit is reached.

  • Take turns Coming next

    The next person or identity in the rotation handles it.

  • Mediate Coming next

    Two parties’ agents negotiate through a clean room; each side sees only what’s agreed, and contact details stay private until both say yes.

Build on what your team already made

Start a new agent from a base your company already trusts. It keeps the base's locked rules, and your team shares skills and access instead of starting over.

  • Build on a base agent: its locked rules carry into every agent built on it.
  • Write a skill or set up a tool once, and any company agent can use it.
  • Invite someone onto one agent, and choose which parts each person can see.
One base, many agents, shared skills and access A company base agent sits at the top with three locked rules: tone, safety and escalation. Three agents are built on it: a service agent, a sales agent and a front desk. The locked rules flow down into each of them, then each adds its own part: returns, pricing and check-in. Below them is a shelf of company skills that all three draw on: Book a service visit, written by Dana and used by two agents, and Look up an order, written by Omar and used by three. Skills from community authors are marked coming next. At the bottom, a panel shows who can reach what. Dana is an administrator for every agent. Luis is a contributor on the sales agent only, and his changes wait for review. Priya is a viewer who sees tone and style only. Sam was invited onto the front desk alone, and it shows for him as Shared with me. Nia, a community author, is marked coming next: she signs herself up, builds her own agent on the base and shares it with Luis. Base agent Company base Rules every agent built on it keeps Tone Safety Escalation Service agent Built on base + Returns Sales agent Built on base + Pricing Front desk Built on base + Check-in Company skills Written once, used by many D Book a service visit Used by 2 agents O Look up an order Used by 3 agents Skills from community authors Coming next Who can reach what D Dana Every agent; can make a version Live Administrator L Luis Sales agent only; changes wait for review Contributor P Priya Sees Tone and style only Viewer S Sam Invited onto the front desk only Shared with me N Nia Coming next Signs herself up as a community author, builds her own agent on the base and shares it with Luis
Every agent keeps the base's locked rules and adds its own. They share the same skills, and each person reaches only what they were given.

For AI consultants and partners Coming next

Agencies that build AI employees for their clients can run every client on Falcon. Publish your base agents, skills and tools once, and each client builds on them under your brand.

  • One partner console for every client you serve.
  • Your base agents set rules a client can't override; their staff edit only their own parts.
  • Each client's usage is billed through with your markup, under your brand.
  • Built on what runs today: base agents, shared skills, access and usage metering.

Become a partner

One partner, many clients Coming next. A partner, Harbor AI, sits at the top with its library of base agents, skills, tools and workflows. Three client companies sit below: Lakeside, Ridge Auto and Maple Homes. The same base flows into each, with rules the client can't override, shown as locks. Each client adds its own part: their FAQ, trade-ins and listings. Each client shows Harbor AI's brand at the top, and a billing line at the bottom: usage of 120 dollars, billed at 144 dollars with the partner's markup. Partner Coming next H Harbor AI Partner library, one console for every client Base agents Skills Tools Workflows Harbor AI Lakeside Same base + Their FAQ Usage $120 Billed $144 Harbor AI Ridge Auto Same base + Trade-ins Usage $310 Billed $372 Harbor AI Maple Homes Same base + Listings Usage $85 Billed $102
Each client gets the same base under your brand, adds its own parts, and is billed its usage plus your markup.

From the first message to your systems

Anything can start a workflow. The agents call your systems, and each agent remembers only what you allow.

Work comes in, work goes out

Work comes in, work goes out Five things can start work: an incoming request from another system such as Order paid, with its signature checked; a new email; a text or chat; a finished call recording; and a schedule. Each passes through a request map that picks out the order id, customer and topic, then into a workflow where a front desk step hands to the service agent, with a side chain that books a call when the person wants one. The agent's tool calls go out: create a ticket in your system, update the CRM through an app connection, and email the reply from Maya's address, which waits for a person to approve it. When the run finishes, Falcon sends a signed callback to the system that started it. Starts work Incoming request Order paid Email New message Text or chat New message Call recording Finished Schedule Every Monday Request map order id · customer · topic Signature checked Workflow Front desk Agent: Service Wants a call Book a call Side chain Goes out Your system Create ticket App connection Update CRM Send from Maya Email the reply Callback Run finished Falcon signs its callbacks Waiting for you Starts work Incoming request Order paid Email New message Text or chat New message Call recording Finished Schedule Every Monday Request map order id · customer · topic Signature checked Workflow Front desk Agent: Service Wants a call Book a call Side chain Goes out Your system Create ticket App connection Update CRM Send from Maya Email the reply Waiting for you Callback Run finished Signed callback to the system that started it
Anything can start a workflow, and agents can call your systems back, signed both ways, with a person approving what matters.

What the agent remembers

What the agent remembers A customer, Jordan Lee, writes: mornings work best, please text rather than call, the van needs brakes, call back Friday. The agent notices those phrases. Its memory strategy, Service customer memory, keeps preferences, commitments and facts about the vehicle, never keeps card numbers or health details, keeps memories for 12 months, and asks rather than guesses. The phrases become saved memories on Jordan Lee's record: best time mornings and channel text, shared with the service team; call back Friday and van brakes due, kept for this agent. Then a signed memory.updated event with only the changed fields, preferred channel text and best time morning, goes to your CRM and your data warehouse. 1 · Conversation Mornings work best, and please text rather than call. The van needs brakes. Call back Friday? Will do. I’ll text you Friday morning. 2 · Memory strategy Service customer memory Kept for 12 months Keeps Preferences mornings · text Commitments Call back Friday Vehicle facts Van: brakes due Never keeps Card numbers Health details Asked, not guessed: the agent checks before it saves. mornings · text Call back Friday Van: brakes due 3 · Saved memory J Jordan Lee Customer since 2024 Just saved Best time: mornings Shared with the service team Channel: text Shared with the service team Call back Friday This agent Van: brakes due This agent 4 · Sent to your systems memory.updated preferred_channel: text best_time: morning Only what changed Webhook Your CRM Webhook Your data warehouse 1 · Conversation Mornings work best, and please text rather than call. The van needs brakes. Call back Friday? Will do. I’ll text you Friday morning. 2 · Memory strategy Service customer memory Kept for 12 months Keeps Preferences mornings · text Commitments Call back Friday Vehicle facts Van: brakes due Never keeps Card numbers Health details Asked, not guessed: the agent checks before it saves. mornings · text Call back Friday Van: brakes due 3 · Saved memory J Jordan Lee Customer since 2024 Just saved Best time: mornings Shared: service team Channel: text Shared: service team Call back Friday This agent Van: brakes due This agent 4 · Sent to your systems memory.updated preferred_channel: text best_time: morning Only what changed Webhook Your CRM Webhook Data warehouse
Each agent follows a memory strategy you set: what to keep, what never to keep, and for how long, and every change can be sent to your own systems.

Fast checks on everything

Every message, file and script gets quick checks from jev by TypeSafe, a small, fast model that answers yes/no, pick-one or score questions about a piece of text with a confidence, in a fraction of a second and at a tiny cost. Code makes the final decision, and anything uncertain goes to a person.

Every input, checked Coming next

Every input, checked A text from Sam asks: can you lock in 2% for me? Call me on 555-0142. Layer 0 is code, always on, with no model: it masks the phone number and applies quiet hours and consent. Layer 1 is jev, always on: instructions to the agent 0.03, asked to stop 0.01, someone in danger 0.02, and a triage label, rates or advice, at 0.88. Triage escalates to Layer 2, the specialist checks: clean room, company rules and abuse. Company rules apply, so the message goes to Review, waiting for you. A message that just says ok with a thumbs-up takes the fast path: it skips layers 1 and 2 and passes. On the way out, the agent's reply, yes, I can promise you 2%, is checked against the company rule never promise a rate. It scores 0.96, so it is blocked and not sent. Comes in ok 👍 Text from Sam Can you lock in 2% for me? Call me on 555-0142 Layer 0: code Always on · no model 555-0142 [phone] Quiet hours Consent Layer 1: jev Always on Instructions to the agent? 0.03 Asked to stop? 0.01 Someone in danger? 0.02 Triage Rates or advice 0.88 Escalates Layer 2: specialist checks Only when triage escalates Clean room Company rules Company rules Abuse Fast path skips layers 1–2 Pass Goes on Pass Goes on Review Waiting for you Review Waiting for you Block Not sent Block Not sent Goes out Reply from the agent Yes, I can promise you 2%. Company rule: never promise a rate Promises a rate? 0.96 Every input, checked A text from Sam asks: can you lock in 2% for me? Call me on 555-0142. Layer 0 is code, always on, with no model: it masks the phone number and applies quiet hours and consent. Layer 1 is jev, always on: instructions to the agent 0.03, asked to stop 0.01, someone in danger 0.02, and a triage label, rates or advice, at 0.88. Triage escalates to Layer 2, the specialist checks: clean room, company rules and abuse. Company rules apply, so the message goes to Review, waiting for you. A message that just says ok with a thumbs-up takes the fast path: it skips layers 1 and 2 and passes. On the way out, the agent's reply, yes, I can promise you 2%, is checked against the company rule never promise a rate. It scores 0.96, so it is blocked and not sent. Comes in ok 👍 Text from Sam Can you lock in 2% for me? Call me on 555-0142 Layer 0: code Always on · no model 555-0142 [phone] Quiet hours Consent Layer 1: jev Always on Instructions to the agent? 0.03 Asked to stop? 0.01 Someone in danger? 0.02 Triage Rates or advice 0.88 Escalates Layer 2: specialist checks Only when triage escalates Clean room Company rules Company rules Abuse Fast path skips layers 1–2 Pass Goes on Pass Goes on Review Waiting for you Review Waiting for you Block Not sent Block Not sent Goes out Reply from the agent Yes, I can promise you 2%. Company rule: never promise a rate Promises a rate? 0.96
Each bar is jev’s confidence. A question about rates goes to a person, a reply that promises a rate is held, and a plain “ok” goes straight on.

  • Identity fields checked on save

    A person’s or place’s details are checked for instructions to the agent, sensitive data and abuse when saved or imported.

    Live

  • Safer scripts and tools

    Every script and tool call is checked for risky or irreversible steps, such as a force push.

    Coming next

  • Side chains and signals

    An independent check, not the agent grading itself, decides when a side chain starts.

    Coming next

  • Smart file scan

    Files and GitHub pull requests are checked for secrets, personal data, hidden instructions and duplicates.

    Coming next

  • Coding agents find the right files

    They ask one question across a repository and read only the files that matter.

    Coming next

  • Ask about documents

    Agents ask a question about documents without loading them into the conversation.

    Coming next

  • Tidy long conversations

    A check spots when a long conversation can be summarised, and never in the middle of a task.

    Coming next

jev never decides alone: code has the last word, a confident answer never grants permission, sensitive checks hold when jev is unavailable, and every decision is recorded.

How Falcon bridges the gap

AI teammate products are finished and easy, but built for a team’s own work. AgentCore is the secure, scalable platform, but you have to build the agents. Companies that want customer-facing agents have had to pick one and live without the other. Falcon sits in the middle.

Finished product

AI teammates

  • Named coworkers you message
  • Do your team’s own work in your apps
  • Work through a shared cloud computer
  • Each person sets up their own bots

Falcon bridges the gap

Falcon

  • Named agents a team shapes without code
  • Face your customers as your people and places
  • Every channel, approvals, consent and reports
  • Runs on AgentCore, in Falcon’s AWS or yours

Platform to build on

Amazon Bedrock AgentCore

  • Secure runtime, memory, identity, tools
  • Any model or framework
  • You write and run the agents
  • Built for developers at scale

  1. From AI teammate products: the finished experience

    Named agents you shape in plain words, starters for common jobs, Ask Studio to help, and nothing to code or host.

  2. From AgentCore: the enterprise platform

    Every Falcon agent runs on AgentCore: isolated sessions, managed memory, identity and tool gateway, observability and scale, in Falcon’s AWS account or the company’s own.

  3. What Falcon adds in the middle: the company layer

    Many people configuring the same agents with roles, drafts and a Live version; identities with consent and their own numbers and addresses; every channel with replies routed to the right agent; approvals, spend limits and reports; and all of it on MCP.

  4. The result

    A company gets customer-facing agents working in days, without building a platform and without giving up control of its data, its spend or what gets sent.

What makes Falcon different

A team shapes each agent

Several people edit one agent within their roles, as drafts; nothing reaches customers until a version is made Live.

Agents act as your people and places

One agent can speak as Dana, the Downtown store or Riverside, each with its own email, number, hours, voice and consent. Identities never sign in to Studio.

Every channel, one conversation

Texts, RCS, email, calls and chat. A reply always finds the right agent and identity, even across channels, and follow-ups only go where the person agreed.

People stay in charge

Drafts wait for approval, agents ask an admin when unsure, and risky actions need confirmation. Nothing is sent that shouldn’t be.

Decisions you can check

Small yes/no and routing decisions use a model built for calibrated answers, with the probability recorded, so they are cheap, fast and auditable.

Everything is on MCP

Anything you can do in Studio, an AI assistant can do through MCP within your access, by design and checked on every build.

Your data, your account

Agents run in AWS, and a company can choose its own AWS account. Reports show counts and cost, never what people said.

Spend you control

Limits, reservations and kill switches on every agent and workflow, so a runaway loop can’t run up a bill.

What teams build

Some of these run today. The ones marked Coming next are designed and on the way, but not built yet.

Dating concierge

Love’s Confidant plans dates with members: a researcher, a planner and a booker work at once, a check makes sure the plan is complete, and the plan is texted back.

  • Workflows
  • Side chains

A store that answers for itself

The Downtown store has its own number, hours and voice. Calls outside hours go to an AI voicemail that takes a message and starts a follow-up.

  • Identities
  • Numbers
  • Voicemail

Front desk for sales and service

One agent greets, looks the person up in past conversations and the CRM, then hands to the sales or service agent.

  • Front desk
  • Routing
  • Coming next

Email replies with approval

A new email to a mailbox gets a drafted reply that waits in Waiting for you until someone sends, edits or drops it.

  • Gmail
  • Approvals

Calls turned into reports

A finished Twilio, Zoom or RingCentral recording becomes a transcript, a summary with follow-ups, and a note on the caller’s history.

  • Recordings
  • Reports
  • Coming next

Sales reps taking turns

New leads rotate between reps who are available, and each customer keeps their rep for later conversations.

  • Taking turns
  • Hours
  • Coming next

A coding agent that commits as your team

An agent makes changes and commits to GitHub as the identity it acts for, with the requester credited.

  • GitHub
  • Identities
  • Coming next

Help answers from your documents

Agents answer from help articles and policies you upload or connect, and say when they don’t know.

  • Knowledge

Outreach that respects STOP

Messages carry topics, replies route by keyword or content, and opting out of marketing never stops order updates.

  • Topics
  • Consent
  • Coming next

A look inside

Screens from Falcon Studio as it is built today, running on example data in one pilot company’s colours. See every feature, and which are live.

Workflow steps with retries if a step fails, and two side chains: one for when the person is interested, one for when they reply STOP.
Side chains and retries. Decide what happens when someone says yes, says STOP, or a step fails.
An identity's contact details: its own receiving email, SIP address, phone number, link and website widget code.
Identities with their own ways in. Each person or place an agent acts for gets its own email, number, link and chat widget.

Put your agents to work for your customers.

Falcon is in a private pilot. Tell us about your team and we’ll be in touch.

Sign up