Agent InfrastructureBeta

Deploy your agent wherever you want.

Flowyn keeps it running, observable, controlled, and connected. It records each run, its traces, model and tool steps, token usage, and the decisions people make, through one REST API.

Flowyn is the infrastructure around your agent, not the intelligence inside it. Keep your framework, your model provider, and your hosting.

Start building Read the agent quickstart

Run recorded

illustrative

refund-agent · ticket-1042

run reported by your runtime

running

steps: 3
model: gpt-4.1-mini
tokens: 876
cost: estimated

The boundary

The infrastructure around your agent, not the intelligence inside it.

Your agent decides what to do. Flowyn keeps the record of what it did, what it cost, and what a person decided, and gives you the primitives around it.

You own

The agent's code and framework
Model choice, prompts, memory, and tools
Where and how the agent runs
Calling your model provider
Waiting for a decision and acting on it

Flowyn provides

A durable identity for each agent
A record of every run and its outcome
Traces, model steps, and tool steps
Token usage, latency, and cost
Human approval decisions
Encrypted provider credentials

What each word means

Running, observable, controlled, connected.

Each word maps to something Flowyn does today, and to a line it does not cross.

Running

Your agent runs on your infrastructure. If it exposes an HTTP endpoint, a Flowyn schedule or durable HTTP job can call it on time with retries, and a heartbeat or HTTP monitor opens an incident when it stops answering.

Observable

Every run, trace, model step, and tool step, with status, timing, errors, token usage, and cost, readable through the API and the dashboard.

Controlled

A person approves or rejects sensitive actions, flags change behavior without a redeploy, and provider keys stay encrypted on the server. Flowyn controls what it owns; it never claims to pause, resume, or cancel a process it does not host.

Connected

Signed, retried webhooks carry results to the rest of your system, and document parsing turns the PDFs and DOCX files your agent reads into Markdown or JSON.

Capabilities

Six capabilities for agents you run yourself.

Each one is a REST resource with its own documentation. Traces, model usage, approvals, and provider connections also have pages in the Flowyn dashboard.

POST /v1/agents

Agent registry

Register each agent you deploy and give it a durable identity: a name, a description, and your own external_id. Archive it later without losing its history.

Up to 100 active agents per environment
Archived agents keep their id for attribution
Framework-neutral: identity and advisory thresholds, no code or model policy

Agent registry docs

POST /v1/agent-runs

Agent runs

Record each execution your runtime performs. A run starts running and you report one outcome: succeeded, failed, or cancelled. Flowyn records the run; it does not host or execute it.

running → succeeded, failed, or cancelled, exactly once
Your own external_run_id for correlation
Failures keep a short error code and message

Agent runs docs

POST /v1/agent-telemetry

Traces and steps

Send spans in batches from wherever your agent runs. Model calls, tool calls, and other operations become the steps of a run, with hierarchy, status, timing, and errors.

Batches of up to 50 spans
Model, tool, and generic operation steps
Metadata only: no prompts, completions, tool arguments, or results

Traces and steps docs

GET /v1/agent-telemetry/usage

Model usage and cost

Input, output, and cached tokens, latency, and cost per model call and per trace, totalled by model, provider, service, or step name over the last hour to 30 days.

Reported cost wins; otherwise an estimate from Flowyn's price catalog
Unknown cost is shown as unavailable with a reason, never as zero
Estimates, not an invoice

Model usage and cost docs

POST /v1/agent-approvals

Human approvals

Your agent asks for a decision about one action. A signed-in person approves or rejects it in the dashboard. Your runtime reads the decision and decides how to continue.

Only a signed-in person can decide; API keys cannot
The first decision is final and records who made it
Flowyn records the decision; it does not pause or resume your agent

Human approvals docs

POST /v1/provider-connections

Provider connections

Store an OpenAI API key encrypted on Flowyn's servers, per environment, and test whether OpenAI accepts it. Optional: your agent can keep calling its provider with its own keys.

OpenAI today
Write-only: no response ever returns the key
Explicit connection test against the provider

Provider connections docs

One run, end to end

From registration to result, in nine calls.

Your runtime makes these calls from wherever it runs. Pick a step to see the request.

curl -X POST https://api.flowyn.dev/v1/agents \
  -H "Authorization: Bearer $FLOWYN_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: register-refund-agent" \
  -d '{
    "name": "refund-agent",
    "description": "Handles refund requests from support tickets.",
    "external_id": "refund-agent"
  }'

Framework-neutral

Keep the framework you chose.

There is no plugin to install and no SDK required. Any agent that can make an HTTPS request can use these endpoints.

OpenAI Agents SDK
LangGraph
CrewAI
Mastra
PydanticAI
Custom Python
Custom Node.js

Trace and span ids use the W3C Trace Context format: 32 and 16 lowercase hex characters. Ids your tracing already produces can be sent as they are.

What Flowyn does not do

No hidden runtime, no pretend control.

Flowyn cannot deterministically control an agent process it does not host, so it does not claim to. A cancelled run is an outcome your runtime reports, and an approval is a recorded decision your runtime acts on.

Host, deploy, or execute your agent's code
Run model inference or proxy your model calls
Store prompts, completions, tool arguments, or tool results
Provide agent memory or a vector database
Orchestrate agents or replace your framework
Pause, resume, signal, or cancel a running agent
Enforce an approval inside your runtime

Built on the five primitives

One account, one API key, the same primitives.

Agent infrastructure is not a separate product. It composes Process, Execute, Connect, Observe, and Configure, so your agent can use all of them with the same key.

Parse the PDFs and DOCX files your agent reads into Markdown or JSON.

Documents & files

Call your agent's HTTP endpoint on a schedule, or run its outbound HTTP calls as durable jobs with retries.

Durable HTTP & schedules

Send results to other systems as signed webhooks that are stored, retried, and replayable.

Reliable events & webhooks

Traces, steps, and model usage, plus monitors and heartbeats on the endpoints your agent depends on.

Monitors & issues

Agent registry, provider connections, human approvals, and flags your agent reads at run time.

Flags & remote configuration

Frequently asked questions

Is Flowyn an agent framework?

No. Flowyn is the infrastructure around your agent, not the intelligence inside it. Keep the framework, model provider, and hosting you chose. Flowyn records identity, runs, traces, model usage, and human decisions through a REST API.

Which frameworks does it work with?

Any agent that can make an HTTPS request: OpenAI Agents SDK, LangGraph, CrewAI, Mastra, PydanticAI, or your own Python or Node.js code. There is no framework plugin to install; you call the REST API.

Does Flowyn run or host my agent?

No. Your agent runs wherever you deploy it. An agent run in Flowyn is a record your runtime reports: it starts running and ends succeeded, failed, or cancelled. If your agent exposes an HTTP endpoint, a Flowyn schedule or durable HTTP job can call it, like any other endpoint.

Can Flowyn pause, resume, or cancel my agent?

No. Flowyn cannot reliably control a process it does not host, so it does not claim to. A cancelled run is an outcome your runtime reports. Human approvals record a decision; your runtime reads it and decides what to do.

Does Flowyn see my prompts or tool results?

No. Telemetry is metadata: names, kinds, timing, status, errors, provider, model, and token counts. Prompts, completions, tool arguments, tool results, headers, and URLs are not accepted.

Do I have to store my model provider key in Flowyn?

No. Provider connections are optional. Your agent can keep calling OpenAI or any other provider with its own credentials. Flowyn records usage you report either way.

How is cost calculated?

If you report a cost for a model call, Flowyn uses it. Otherwise Flowyn estimates it from the reported tokens and its price catalog of selected OpenAI, Anthropic, and Google models. A model outside the catalog shows cost as unavailable with a reason. These are estimates, not an invoice.

Give your agent infrastructure.

Create a workspace, register an agent in the test environment, and send its first trace.

Start building Read the docs