Manual & API
Run from the console — including test runs against sample input — or from your own systems through the API and SDK.
Design the process. Let agents do the work.
AI creates value when it acts inside a real process — not when it sits in another chat window. Gecko workflows connect reasoning, knowledge, systems and people in one versioned, traceable flow.
20step types
4ways to trigger
12routing operators
JSONportable export
↗
Entry · webhookStudent enquiry
AI
EvaluateClassify intent · 0.94
⌕
Lookup knowledgeFind approved policy
◇
RouteConfidence ≥ 0.8?
✉
Send messageRespond with sources
◎
Pass to agentEscalate with context
The builder
A visual builder your IT team can reason about.
Add steps from the palette, wire branches on the canvas and configure each one in the inspector. Data flows between steps with a readable reference syntax — $.input.question, $.previous — so logic stays inspectable.
- AI Workflow Builder — describe the process in plain English and get a step-by-step wireframe, each node marked configured or needing setup
- Accessible outline view — edit the whole workflow without the canvas
- Example gallery — copy proven patterns such as support triage, knowledge answer with review, parallel research and authenticated tool approval
Describe the process
When a new admissions enquiry arrives, draft a personalised reply using our programme pages, have a recruiter approve it, then email the applicant and log it in the CRM.
Proposed wireframe
1Webhook · New enquiryEntryConfigured
2Search website · Programme pagesRetrievalConfigured
3Generate text · Draft replyAIConfigured
4Requires approval · RecruiterHuman-in-the-loopNeeds setup
5Send message · Email applicantActionNeeds setup
20 step types
Every building block a real process needs.
From AI reasoning and retrieval to system actions and human checkpoints — each step records its inputs, outputs and timing on every run.
↗Entry
The start of every run, with an example input to test against.
AIGenerate Text
Draft, summarise, classify or extract — as text, a fixed choice or JSON matching a schema. Can use knowledge, website and MCP tools, and read images.
AIExpand User Query
Rewrite a request into broader search variants for better recall.
AIEvaluate
Score or classify with calibrated probabilities and confidence — choice, scale or yes/no — ready for threshold routing.
⌕Lookup Knowledge
Retrieve from the knowledge folders this step is allowed to see.
⌕Search Website
Search crawled website content with relevance thresholds.
⇅Rerank Results
Re-score results with a dedicated reranking model from Cohere, Amazon Bedrock or a compatible endpoint.
≡Consolidate Context
Merge retrieved context and keep it inside a character budget.
◇Route
Branch on 12 operators — equals, contains, matches, greater than, is true and more — with a fallback path.
⫘Parallel Branch
Fan out into paths that run at the same time.
⋈Merge
Join parallel branches back into a single result.
MCPMCP Tool
Call a named tool on a connected MCP server — the SIS, CRM or your own systems.
◎Pass to Agent
Hand the run to a configured agent, keeping the conversation going.
{}Custom Code
Run Node.js in an isolated sandbox, with encrypted secrets bound at execution time.
⇢Webhook
POST, PUT or PATCH JSON to any approved endpoint, with custom headers.
▤File Action
Ten file operations across S3, Google Drive, OneDrive, SharePoint and Dropbox.
✉Send Message
Notify through Slack, Microsoft Teams, email or SMS.
☎Make AI Call
Place an outbound AI voice call, with voicemail detection and a choice of ten voices.
✓Requires Approval
Pause for a person to approve or reject — from Slack, Teams, email, SMS or the console.
IDAuthenticate
Email a one-time code to verify who the workflow is acting for before continuing.
Triggers
Start work from anywhere.
The same workflow can run on demand, from another system, on a timetable or in response to what happens in a conversation.
Inbound webhook
Every workflow gets its own unguessable URL, so any external system can start a run.
Schedule
Cron schedules with friendly presets — hourly, 9am on weekdays, midnight on the first of the month.
Agent events
When a conversation starts, a message arrives or a conversation closes — blocking the reply or running in the background.
Human-in-the-loop
The approval is a button in Teams — not a login to another system.
When policy says a person decides, the run pauses. Approvers act where they already work, and every decision is recorded against the run.
- Interactive Approve / Reject cards in Slack and Microsoft Teams, or links by email and SMS
- Single-use action tokens, stored only as hashes — approvals can’t be forged or replayed
- Identity checks mid-run — a six-digit emailed code, expiring, with a five-attempt limit
- First-class waiting states — waiting approval, waiting authentication and waiting call — with average review time on the dashboard
G
Gecko AI · Tuition refund workflow
Approval required
Refund £450 deposit to applicant 20231184
Identity verified by email code. Policy 4.2 matched with 0.92 relevance. Drafted by agent “Admissions finance”.
Approve
Reject
ID
Authenticateverified a•••@university.ac.uk
09:41✓
Requires approvalwaiting · sent to 3 approvers
09:41
Decisions you can defend
Route on confidence, not on hope.
The Evaluate step scores or classifies input with calibrated probabilities, so a Route step can send low-confidence cases to people and let high-confidence cases flow. Rerank Results uses dedicated reranking models to sharpen what the AI reads.
- Choice, score or yes/no evaluations with probabilities for every answer
- Reranking models from Cohere, Amazon Bedrock or a Cohere-compatible endpoint
- Structured output — Generate Text can return JSON that matches a schema you define
Question · choice
Which team should handle this enquiry?
Probabilities
Finance
0.94
Admissions
0.04
Student support
0.02
◇Route · confidence ≥ 0.8 → Finance workflow
Custom Code
Extend anything — safely.
When a step needs logic of its own, write Node.js — or let the AI code assistant draft it. Code runs in an isolated sandbox per execution, never on the platform itself.
- Isolated sandbox per execution, with a hard 300-second limit
- Encrypted secrets bound to environment variables only at run time — leaked values are replaced with
[REDACTED]in errors - Test tab runs the code against a sample payload before you save
// Custom Code · runs in an isolated sandbox
const res = await fetch(`https://timetable.example.ac.uk/rooms/${input.roomId}`, {
headers: { authorization: `Bearer ${process.env.TIMETABLE_KEY}` },
});
const room = await res.json();
return {
available: room.bookings.length === 0,
nextFree: room.nextFreeSlot,
};
Runs & versions
Always know which logic produced which outcome.
Drafts are edited freely. Publishing creates an immutable, numbered version, and every run is bound to the version it ran — with a step-by-step record you can open from the canvas.
- Per-step input, output, timing and errors; failed steps light up on the canvas
- AI step detail — model, finish reason, token counts and every tool call
- Cancel in-flight runs, and act on approvals and codes directly from the run view
- Export and import as JSON — copy workflows between teams or take them with you
ID
Authenticateverified · code entered on canvas
41sMCP
MCP Tool · servicedesk.getStatus{"service":"vpn","status":"degraded"}
0.3s◇
Routestatus equals "degraded" → Known incident
0.0sAI
Generate Text · Guidancefinish: stop · 942 in / 188 out tokens
1.9s✉
Send Message · Teamsdelivered to requester
0.4sTwenty step types. Approvals in Slack or Teams. Every run versioned and traceable — and the whole workflow exports as JSON.
Keep exploring
Bring one process
We’ll sketch the workflow with you — approvals included.
Pick a high-volume process with trusted knowledge and a measurable outcome. We’ll show you what it looks like as a governed workflow.
- A 30-minute working session
- Your process mapped to real step types
- The controls and evidence it needs