SERVFLOW

Build agents.
Prove they work.

Servflow is the agent platform for production teams โ€” design agents visually, watch every decision they make live, and gate every deploy on benchmarks.

$brew install Servflow/servflow/servflow-pro
support-agent ยท live runproduction
00:00.04run started ยท trigger POST /support
00:00.31agent thinking ยท intent = refund_request
00:00.58tool_call ยท orders.lookup(#4821) โ†’ found
00:01.02guardrail passed ยท refund โ‰ค $500
00:01.19reply sent ยท 1.2s ยท $0.011 ยท score 94

One loop: build, observe, prove.

Servflow closes the loop between shipping an agent and knowing it works.

01BUILD

Design agents on a canvas, not in a codebase.

Map out how your agent thinks โ€” prompts, tools, guardrails, and decision paths โ€” in a visual builder. Every change syncs to your repo as a declarative file, so agents get reviewed like code.

Visual builder40+ actionsModel-agnosticGit sync
Visual agent builder
02OBSERVE

Watch every decision, live.

The cloud dashboard traces each run step by step โ€” the prompt, the tools called, the data touched, the tokens spent. When an agent does something surprising, you can see exactly why.

Run tracesDecision inspectionCost & token insights
Request Logs
1,247Total Requests
99.2%Success Rate
67msAvg Response
6Endpoints
EndpointMethodStatusDuration
/api/pipelines/executereq_8f2a4b6c
POST
200
127ms
/api/instances/statusreq_3e7f1a9c
GET
200
23ms
/api/webhooks/stripereq_9d4e2f8a
POST
201
89ms
/api/logs/recentreq_5a1b3c7d
GET
200
45ms
03BENCHMARK

Gate deploys on proof, not vibes.

Score every agent version against your own scenarios. If a prompt tweak or model swap makes the agent worse, the regression gate blocks the deploy before your users ever feel it.

Scenario suitesRegression gatesModel comparisons
Benchmark Runssupport-agent ยท v14 vs v13
0.0Overall Score
+0.0vs Last Version
$0.000Avg Cost / Run
0.0sP95 Latency
ScenarioCasesv13v14Gate
refund-requests329196+5Pass
order-lookups249898+0Pass
angry-customer-escalation188492+8Pass
multi-step-cancellations218178-3Fail
out-of-scope-questions338994+5Pass

A runtime built for production

Single binary. Zero dependencies. Runs on your laptop, your VPC, or Servflow Cloud.

agents/support-agent.yamlsynced
agent: support-agent
model: claude-sonnet-5
tools:
  - orders-db
  - refunds
guardrails:
  max_refund: $500
benchmark:
  suite: support-v2
  min_score: 90

Agents as code

Every visual change syncs to your repo as a declarative file โ€” reviewed, diffed, and versioned like the rest of your stack.

AnthropicOpenAIGeminiLlamaLocal

Any model, any step

Mix providers per step and swap them without rebuilding the agent.

refunds โ‰ค $500allow
orders.readallow
db.writeblocked

Guardrails built in

Scope exactly what each agent can touch before it touches production.

12ms

Millisecond overhead

The Go runtime adds almost nothing on top of your models.

Scale out instantly

Run one agent across replicas as demand grows.

Every step traced

Native traces, or export to Jaeger and OpenTelemetry.

Easily Set Up Integrations With Various Services.

Connect your database, LLM, email, vector database.

Ship agents
you can prove.

Join the teams building, observing, and benchmarking their agents on Servflow.

Open source runtimeโ—†Single binaryโ—†Zero vendor lock-in