Skip to content

Agents that do the work, not just the talking.

Bunshin gives an agent your knowledge, your tools and a workflow to follow — then puts it on the channels your customers already use. Every step it takes is on the record, and a human can take the thread at any point.

Access is by application. Tell us what you’d put an agent on.

app.bunshin — Run trace
Support agentrun 8f2a·416 steps · 3 credits
  1. Received

    Telegram · Priya M.

    “Where is order 4471? It was meant to arrive Tuesday.”

  2. Retrieved

    Shipping policy · 2 passages

    Delayed-delivery terms, replacement window

    180ms
  3. Tool

    orders.lookup

    order_id 4471 → in_transit, ETA 2 days

    340ms
  4. Branch

    Workflow · Late delivery

    Delay under 5 days → apologise, offer tracking

    20ms
  5. Replied

    Sent to customer

    Tracking link, revised ETA, replacement offer if it slips

    1.2s
  6. Handed off

    Refund requested → Dara

    Above the agent's approval limit. Thread passed intact.

Grounded.

It answers from your material

Attach your documents and the agent answers from them. You can test what it retrieves for a given question before a customer ever sees it.

Capable.

It resolves, not just responds

It calls the tools you give it and follows the workflows you draw, so a request can end in an action rather than a description of one.

Accountable.

Every run is on the record

Each run reads back as a trace: what it retrieved, what it called, what it said, what it cost, and where it stopped and asked for a person.

It answers from your documents, and you can check that before shipping it.

Attach files, pasted text or a URL. Ask the retrieval tester the question a customer would ask, and read back the exact passages the agent would answer from.

  • Files, text or a link

    Upload a PDF or spreadsheet, paste text straight in, or point it at a page you already publish.

  • Retrieval is inspectable

    See which passages matched a question and how strongly, before a customer ever asks it.

  • Update once, answers follow

    Change the document and every future run reads the new version. No retraining, no waiting.

app.bunshin — Knowledge · Test retrieval
My order is a week late, can I get a refund?
Returns policy 2024.pdf0.91

Unopened items may be returned within 30 days of delivery for a full refund.

Shipping terms.docx0.84

Where a delivery is more than 5 days late the customer may request a replacement at no cost.

Support macros.md0.61

Always confirm the order number before offering a refund or replacement.

A request can end in an action, not a description of one.

Attach the tools the agent is allowed to use and it will call them mid-conversation — looking up an order, tracking a shipment, booking a slot — with the arguments and results kept in the run.

  • Scoped per agent

    An agent only holds the tools you attach to it. Nothing else is reachable.

  • Approval where it matters

    Mark a tool as needing approval and the agent stops at it rather than deciding for you.

  • Every call is recorded

    Arguments in, result out, how long it took. Nothing happens off the record.

app.bunshin — Agent · Tools
orders.lookupRead
refunds.issueNeeds approval
shipping.trackRead
calendar.bookWrite
orders.lookup · 340ms

→ { "order_id": "4471" }

← { "status": "in_transit", "eta_days": 2,
"carrier": "DHL" }

For the requests that already have a procedure, draw the procedure.

Branching workflows sit on the canvas next to the agent. Where a decision has a rule, the rule decides; where it needs judgment, the branch ends at a person.

  • Drawn, not written

    Conditions and branches on a canvas. No expression language to learn.

  • Test before it goes live

    Run sample requests through the workflow and watch which branch each one takes.

  • A branch can end at a human

    Handoff is a destination you can design toward, not an error the agent falls into.

app.bunshin — Workflows · Refund triage
Refund askedWithin policy?order age · conditionIssue refund, confirmHand to a personyesno
Tested against 14 sample requests12 resolved · 2 handed off

The moment it needs judgment, you step in.

Take any conversation over and reply as yourself. The agent stops immediately, and picks up again with everything you said already in its context.

  • One tap to step in

    Open any conversation and start typing. The agent goes quiet the moment you do.

  • The customer sees one thread

    No transfer notice, no restart, no repeating themselves. It stays the same conversation.

  • Hand it back when you're done

    The agent resumes with your replies as part of the record.

Telegram — handed to Dara
Customer

It's been eight days. I'd rather just have my money back.

Dara took over this conversation
You

Hi, it's Dara from support. That's on us — I've refunded it in full just now, and you'll see it in 3 working days.

Autonomy you can revoke mid-sentence.

Handoff is a designed state, not a failure path. Read any run while it happens, take the thread as yourself, and hand it back with everything you said already in the agent’s context.

Three layers, and no deployment pipeline.

There is no build step, no hosting decision and no promotion from staging. Connecting a channel is the deployment: the agent is live on it the moment the token lands.

01

The channel

Telegram takes one bot token. Email gives you an inbound address. The web widget is a single line of script.

02

The context

Documents, tools and workflows attach to the agent. Change any of them and the next run uses the new version.

03

The oversight

Runs are traced, conversations can be taken over, and every credit is accounted for against the run that spent it.

You pay for the work the agent does. Credits are metered per run and drawn from a prepaid balance — no seats, no subscription, nothing that renews while you're not looking.

  • Metered per run, not per seat

    A run is one request the agent works through. Invite your whole team without the bill moving.

  • The balance is always visible

    It sits on the screen you use most, and every run reports the credits it consumed.

  • Nothing renews on its own

    When the balance reaches zero the agent pauses until you top it up. It cannot quietly bill you.

  • Priced before you are charged

    Per-credit pricing is confirmed with you in writing before the first credit is drawn.

We take on a set number of teams at a time.

Tell us what you want an agent to handle and which tools it would need to touch. If it’s a fit, we set the workspace up with you.

Apply for access

Straight answers. Including the ones we can't answer yet.

Both. You attach tools to an agent and it calls them mid-conversation — looking up an order, tracking a shipment, booking a slot — then answers using what came back. Any tool you mark as needing approval stops and waits for a person instead of acting.

Put an agent on the work that repeats.

The requests that arrive a hundred times a week already have a procedure. Give that procedure to an agent and keep the judgment calls for yourself.