A visual canvas for agentic workflows.
Drag, connect, and run multi-step AI pipelines — real model execution, streaming output, durable runs, and full observability. No glue code.
Clients
Teams shipping AI in production
Northwind
dfndr
Leadplus
Basis
illuminate
NIX
linkup
Vantage
Northwind
dfndr
Leadplus
Basis
illuminate
NIX
linkup
Vantage
The problem
Prompts don't ship themselves.
A good prompt is five percent of the work. The other ninety-five is orchestration — and today that means a pile of brittle scripts.
01
Glue code everywhere
Queues, retries, and back-off — hand-rolled for every single step.
02
Black-box failures
A run dies at 2 a.m. and you're left grepping logs for the node that broke.
03
Scattered control flow
Approvals, tool calls, and branching live in one-off scripts nobody trusts.
04
Cost flies blind
Token spend and latency stay invisible until the invoice lands.
The Vesper way
One canvas. Every step. In view.
++++
onboarding.vsp
runningWatch it run
Live node state and streaming output, right on the canvas.
Durable by default
Runs survive crashes, pause on approvals, and replay on demand.
Metered per node
Token spend and latency tracked for every step and model.
The canvas
Design pipelines you can see.
Drag nodes onto an infinite canvas and wire them together. Prompts, images, models, code, conditionals, approvals — every step is a node you can inspect while it runs.
Multimodal
Text, image, and file inputs and outputs.
Any model
Anthropic, OpenAI, Google — bring your keys.
Live runs
Per-node state and streaming, as it happens.
Metered
Tokens, cost, and latency on every step.
Capabilities
Batteries included.
The hard parts of shipping AI — handled. Build, run, watch, govern, and meter every workflow, with no extra services to wire up.
Usage · last 24h
Resource overview
Runs
1,284▲ 12%
Agents
14▲ 2
Tokens
2.4M▲ 8%
Spend
$12.40▼ 6%
gemini-2.5-flashgpt-4oclaude-sonnetpython-workerpeak 14:00 · 61 runs
Blocks32Data8⌘K
Search blocks…
Most used
Model
Pick a model & params
Prompt
Compose the instruction
Conditional
Branch on a condition
HTTP Request
Call an API endpoint
Code
Run a transform step
↑↓ navigatedrag to canvas⏎ add
Live execution
AnthropicOpenAIGooglenode 3/5
1,204 tok118 tok/s$0.0021
Activity
Tariro approved the refund gate · now
Run #1284 recovered after a retry · 3m
Anesu edited the Summarise prompt · 6m
Durable run completed · 9m
Cost per run
$0.0042
▼ 18% vs avginput
$0.0019
output
$0.0015
tools
$0.0008
How it works
Design. Run. Observe.
The whole loop lives in one place. Build the pipeline, fire it off, and watch every step report back — from first node to final output.
1Design
design.vsp
Wire the flow.
Drag nodes onto the canvas and connect them — prompts, models, code, branches, and human gates. No glue code, no YAML.
2Run
run · live
Press play.
Execute in the browser for a fast loop, or hand off to durable server-side runs that survive crashes and resume on approval.
3Observe
metrics
Watch it live.
Streaming output, token cost, latency, and per-node state — all in view as it happens. No black boxes, no log grep.
Use cases
Built for the work you actually do.
One canvas, five very different shapes of work — branching, parallel fan-out, self-correcting loops, and human gates. Pick a lane.
auto-advancing · hover to pause
support-triage.vsp
Branch · human-in-the-looprunning
Classify, draft, and route every ticket — your team only touches the edge cases.
Auto-resolved
82%
Avg handle
11s
To human
18%
Generative workflows
Yes — a workflow made this.
The portrait to the right isn't stock art. It's the output of a five-node Vesper pipeline — a prompt, a style reference, an image model, and a review gate. Point the same canvas at copy, code, audio, or research.
- prompt
- kunoichi mid-motion, flowing silk, chartreuse + coral
- style ref
- editorial portrait · cream studio
- model
- gemini-2.5-flash-image
- pipeline
- 5 nodes · durable run
++++
kunoichi.png
generated1254×1254gemini-2.5-flash-image
Developer experience
Drive Vesper from the terminal.
The canvas is one front-end for the engine — not the only one. Everything runs on a clean REST API, so you can script, schedule, and ship from wherever you already work.
Python CLI
Trigger, inspect, and tail runs.
Full REST API
Every UI action, scriptable.
Terminal UI
A TUI on the same endpoints.
Simulate mode
Zero-setup demos and CI.
vesper — zsh — 80×24
$ python -m vesper_cli trigger call-qa --input ./calls/
✓ run queued id=run_8f2c1a workflow=call-qa
● scoring 14 calls ████████░░ 80%
✓ done — 2 flagged · $0.058 · 11.3s
$
Bring your own keys
Your provider keys, your accounts — no token markup, ever.
AES-256-GCM at rest
Keys are encrypted at rest and decrypted only to call a model.
Never logged
Secrets never hit logs or traces. Redact PII before it leaves the flow.
Pricing
Pay for the canvas, not the tokens.
Bring your own provider keys and pay them directly — Vesper never marks up a single token. Plans cover the orchestration, not the inference.
Hobby
$0
free forever
For trying things out and small personal pipelines.
Start freeIncludes
- 1 workspace
- Browser runs
- Community templates
- Bring your own keys
- 7-day run history
++++Most popular
Pro
$20/ mo
billed monthly · BYOK
For builders shipping real workflows to production.
Start ProIncludes
- Unlimited workflows
- Durable server-side runs
- Full observability & traces
- Approval gates
- 90-day history
- No token markup
Team
Custom
annual · talk to us
For orgs running AI across teams, with governance.
Book a demoEverything in Pro, plus
- SSO & roles
- Audit logs
- Shared multiplayer canvases
- Priority support & SLA
- Dedicated onboarding
All plans are bring-your-own-key · you pay providers directly · cancel anytime
No. Runs execute either in your browser for a fast build loop, or as durable server-side jobs that Vesper manages — queues, retries, and back-off included. There are no extra services to stand up.
OpenAI, Anthropic, and Google out of the box, plus your own Python workers for anything custom. Switch the model on any node without rewiring the graph.
Never. Vesper is bring-your-own-key: you pay providers directly at their rates, and we add zero token markup. Your plan covers the canvas, durable runs, and observability.
Yes. A run survives process crashes and restarts, pauses cleanly on human-approval gates, and can be replayed from any node. You never lose a half-finished pipeline at 2 a.m.
Everything in the UI is backed by a REST API. There's a Python CLI and a terminal UI on the same endpoints, plus a simulate mode for zero-setup demos and CI.
Provider keys are encrypted at rest with AES-256-GCM and are never written to logs or surfaced in traces. You can redact PII in-flow before anything reaches a model.
Yes — the canvas is multiplayer. You'll see teammates' live cursors and edits, and approval gates can route to specific reviewers before a run continues.
Start building
Ship your first agent before lunch.
Design it on the canvas, run it for real, and watch every step report back. No glue code, no infra to stand up.
No credit cardBYOK — your keysFree to start