Agents
Specialists with instructions, a model, tools and a structured output schema.
Features
Everything between a prompt and a product people use.
Vendor.sh is organised by what you are trying to do: build the work, connect it to your product, release it safely, and keep it under control once it is live.
What you can do with Vendor.sh
01 · Build
Shape the work before it touches anything real.
Specialists, shared prompts and a visible graph, so the path a request takes is something the team can read.
Prompts
A versioned prompt library that every agent in the workspace can share.
Workflow canvas
Chain agents, tools and branches on one graph you can inspect.
Tools
Built-in, HTTP and MCP tools, each with deliberately limited access.
Knowledge
Search your own sources from any agent or workflow step.
Templates
Start from a checked agent, workflow or tool instead of a blank page.
02 · Connect
One stable connection for every backend you run.
Your product calls a versioned API. Credentials, provider accounts and private identifiers stay on the platform side.
Platform API
A versioned /api/v1 surface described by an OpenAPI 3.1 contract.
Project service keys
Scoped abilities, a deployment allowlist, an expiry and one-click revocation.
Embeds
Workflow embeds that sit inside your own pages, on a session token.
Connections
Provider and tool accounts live outside workflow content.
Bring your own key
A call on your provider key creates no managed provider debit.
Idempotent runs
Send an Idempotency-Key and a retried request never runs twice.
03 · Release
A version you tested is the version that runs.
Development, Staging and Production are separate places, and a release does not change after it is made.
Environments
Development, Staging and Production, each with its own deployments.
Immutable releases
A released workflow version never changes underneath a caller.
Evaluations
Score a version against real examples before anyone depends on it.
Human approval
A workflow can pause and wait for an eligible reviewer.
Try runs
Run a draft with real inputs before you release it.
Deployments
Point each environment at exactly the release it should serve.
04 · Control
See every step, cost and failure while it happens.
Live work stays readable after launch: who ran what, what it cost, and where it stopped.
Runs
Every run carries a state the whole team can read: queued, running, succeeded, failed.
Traces
Each step with its inputs, tokens and latency.
Cost budgets
Budgets stop spend before the invoice does.
Members and permissions
Server-side permission checks on every action.
Audit trail
Changes to keys, members and providers are written down.
Workspace isolation
Separate workspaces and projects limit cross-workspace access.
Build a useful AI feature, not a collection of settings.
Start from an agent, a workflow, or a ready template.