Fully Managed Agentgateway
as a Service

Deploy Agentgateway as a fully managed service starting at €9/mo. Get automated backups, SSL, updates, support and monitoring included.

Agentgateway is an open-source proxy for agent traffic. It fronts Model Context Protocol tool servers, agent-to-agent calls and model providers behind a single address, combining the convenience of a hosted AI-gateway subscription with the control of software you run yourself. It carries ordinary HTTP and gRPC as well, so one proxy covers both kinds of traffic.

✓Free 7-day trial  ✓99.9% Uptime SLA  ✓No credit card  ✓Cancel anytime

✓Free 7-day trial  ✓99.9% Uptime SLA
✓No credit card  ✓Cancel anytime

Agentgateway

Agentgateway

STARTING AT

€9/month
Automated Backups
Monitoring
Automated Updates
Auto SSL

USAGE

Unlimited
Human Support
Custom Domains
Terminal Access
File Manager Access
Deploy in your region — 21 locations worldwide
GermanyFinlandNetherlandsUKSwedenUnited StatesCanadaSingaporeJapanAustraliaBrazilSouth Africa+9 more →
Agentgateway Preview Image

ABOUT THE SOFTWARE

What is Agentgateway

Agentgateway is an open-source data plane for agent traffic. It sits between your agents and everything they call, adding identity checks, per-tool access rules, retries, rate limits and recorded telemetry to calls that would otherwise leave your network unchecked.

The project is written in Rust and ships as a single static binary alongside a container image and a small command-line tool, so a working deployment is one process and one configuration file rather than a stack. It was originally built by Solo.io and contributed to the Linux Foundation, where it is developed in the open under the Apache 2.0 licence with contributors from several large infrastructure vendors.

Its scope is wider than the name suggests. As well as Model Context Protocol tool servers and agent-to-agent traffic, it terminates ordinary HTTP and gRPC, and it presents model providers behind one OpenAI-shaped endpoint with failover between them. Configuration is a flat file the process watches: change a route and the running proxy picks it up, with the previous version kept in place if the new one fails validation.

FEATURES

What Agentgateway does

Agentgateway is a proxy first and an AI product second. The list below is what the software does; the platform work around it is the same for every application we run. These are the capabilities most teams put it in front of production for.

One endpoint for many model providers

Presents several model providers behind one OpenAI-shaped API with failover, so a provider outage or a model rename is a routing change.

Model Context Protocol federation

Aggregates several MCP tool servers into one address and one tool list, so an agent connects once instead of once per server.

Local tools made remotely reachable

Launches a process-based MCP server and exposes it over streamable HTTP, turning a laptop-only tool into one a hosted agent can call.

Per-tool access rules

Evaluates expression-based rules against verified token claims, so a read-only tool stays open while a writing tool is limited to one group.

Identity checks at the edge

Validates signed tokens, supports browser login through an OpenID Connect provider, and handles keys for callers that can do neither.

Upstream credential handling

Holds provider keys as environment references and supports token exchange, so a short-lived token can replace a long-lived static key.

Telemetry built for model traffic

Exports traces over OpenTelemetry and metrics in Prometheus format, including token counts per model call and per session.

Ordinary traffic management

Retries, timeouts, rate limits, load balancing, TLS termination and header rewriting for HTTP and gRPC alongside agent traffic.

WHAT'S ALWAYS INCLUDED

Every app. Fully managed.
Nothing extra to pay for.

Every app on DANIAN ships with the same operational baseline. You do not choose a backup add-on or a monitoring tier. Hosting, patching, backups, SSL, and support are part of the €9, and they apply to Agentgateway the same way they apply to everything else in the catalog.

Automatic updates and patches

Apps run the latest stable version. Security patches applied silently, with rollback if needed.

Daily off-site backups

Multiple daily backups in redundant off-site locations. One-click restore if anything goes wrong.

24/7 uptime monitoring

Continuous monitoring with instant alerting. We respond before you notice.

SSL, firewall, DDoS protection

Auto-renewing SSL, hardened firewall rules, DDoS mitigation on every deployment.

Performance and scaling

We monitor resource usage continuously. When your app needs more headroom, we flag it and upgrade with your explicit approval.

Dedicated engineering support

Real engineers by email and ticket. DNS, SMTP & migration help. All included in €9.

WHY MANAGED

Why teams pick managed Agentgateway

The protocol underneath these agents changed twice in a year, first to require proper authorization and then to move sessions toward a stateless model. Teams that wired their own proxy in a hurry are now the ones maintaining it, which is the usual reason a managed instance starts to look sensible.

Running the proxy is the easy part. What follows it is upgrade reading: recent releases changed how MCP authentication is attached to a route, changed the shape of one authentication field, changed what a rejected guardrail request returns to the client, and stopped publishing one of the two container image variants. None of that is unreasonable for a project moving at the speed of the specification it implements. All of it lands on whoever owns the upgrade.

The failure modes are specific and they are not obvious the first time. A browser-based client fails silently if the session and protocol-version headers are missing from the cross-origin header sets. A second replica breaks live sessions unless session lifetimes are bounded. The built-in editing interface rewrites the file it manages and drops your comments. Most of the configuration hot-reloads from the watched file, but the top-level block holding logging and the administrative address does not, so a change there quietly does nothing until the process restarts.

The instance can be configured so those traps are closed at setup rather than left as an exercise. That means the glibc image pinned, both stream transports enabled on one route, the two Model Context Protocol headers already present in the cross-origin sets, session lifetimes bounded, a ceiling on connection duration for long-lived streams, the metrics endpoint and the editing interface kept off the public listener, and upstream provider keys referenced from the environment instead of written into the file you edit. Body and prompt recording can be left off until you ask for it.

REVIEWS

Hear from customers ​like you​​​​​​​

Successful businesses and professionals around the world rely on DANIAN every day

USE CASES

Three teams who run Agentgateway on DANIAN

Three teams, drawn from the kinds of deployment this software attracts. Each one has a different reason for putting a proxy in front of agents, and each one made a different configuration decision because of it.

NINE-PERSON AI CONSULTANCY

One gateway per client, tools scoped by group

Runs a separate instance per client in the region that client's staff sit in. Each instance federates three or four Model Context Protocol servers into one tool list, with expression rules keyed to group claims so a client's own staff see their tools and the consultancy's engineers see all of them. Provider keys stay per client.

FOURTEEN-PERSON PRODUCT TEAM

Two model providers, one endpoint, per-team ceilings

Moved every model call in the product behind one address so a provider incident is a routing change. Two providers are configured with failover, token counts are exported to their existing Prometheus stack, and per-team rate limits stop a runaway evaluation job from consuming a month of quota in an afternoon.

SIX-PERSON INTERNAL TOOLS GROUP

Process-based tools reached by a hosted agent

Wraps three process-based MCP servers that read from internal systems and exposes them over streamable HTTP so a hosted assistant can call them. Signed tokens from the company identity provider are required on every route, connection duration is capped, and body recording is left off because the tools return staff records.

COMPARISON

Four ways to run Agentgateway

The honest comparison is not a per-seat blowout. Hosted gateway subscriptions in this category tend to charge a flat fee with unmetered users, then bill by request volume. What changes between the four paths is who does the operational work and where the records live.

Path1 user5 users10 usersWhat it costs you beyond money
Hosted gateway subscription (Portkey Production tier)$49/mo$49/mo$49/moUsers are not metered, so the seat count is not the axis. Cost rises with request volume: a set number of logs is included and additional blocks are billed on top. Records sit with the vendor.
Self-host on a VPS~$24/mo~$24/mo~$24/moServer cost is flat, your time is not. You own kernel patching, TLS renewal, upgrade reading, backups and the 3am page. Budget several hours a month once agents depend on it.
Home server~EUR 180 one-time+ electricity+ your bandwidthWorkable for experiments. A proxy that agents reach from outside needs a static address, inbound ports, a renewing certificate and an uplink you trust. Residential connections supply none of those reliably.
DANIAN managed AgentgatewayEUR 9/moEUR 9/moEUR 9/moFlat, whatever the seat count. Updates, backups, monitoring, certificates and support are included, the region is yours to pick, and terminal and file access stay open so you can read the configuration yourself.

Figures are the published list prices at the time of writing, in the vendor's own currency. Check them before you rely on them.

BY INDUSTRY

Agentgateway for specific industries

A proxy in front of agents is where two obligations become practical rather than aspirational: knowing who called which tool, and being able to show it later. These four sectors have a named rule that says so, and each one changes how we configure the instance.

Broker-dealers work under SEC Rule 17a-4 and FINRA Rule 4511, which require business records to be preserved for three years with the first two readily accessible, and six years where the rule names no period. Electronic records must sit on write-once storage or in a system that keeps a verifiable record of every change. On this app we enable request and response recording on the routes carrying model and tool calls and ship the output to storage you control, with the retention window set by you. A research assistant calling an internal filings tool then leaves one entry per call, naming the caller, the tool, the arguments and the time.

The AI Act places recruitment and candidate-selection systems in its high-risk category, and Article 26(6) requires deployers to keep the logs those systems generate for at least six months. A screening flow usually spans several calls: a parsing tool, a ranking model, sometimes a retrieval step, each one a hop nobody records by default. Routing them through one gateway makes the whole path a single recorded sequence per candidate. We set the retention window on the instance to match the longest period you are subject to and keep the record in the region you picked at signup, so the six-month floor is a setting rather than a project.

Professional conduct rules require a lawyer to protect client confidences and to supervise the assistance used on a matter, which now includes the tools an assistant calls on the lawyer's behalf; the model rules also require certain client records to be kept for five years after the representation ends. The practical control is scoping. We key the expression rules to matter groups, so a paralegal's assistant reaches the document search tool for their own matters and nothing else, and the tool list it receives is filtered to match. A drafting assistant calling a clause library produces a record of matter, clause set and model, while body recording stays off so privileged text is not copied anywhere.

Systems holding health records fall under 45 CFR 164.312, which requires unique identification for every user, mechanisms that record and examine activity in those systems, and a check that the person or service asking is who it claims to be; the documentation rule at 45 CFR 164.316 sets a six-year retention period. Shared API keys fail the first of those outright, which is the usual state of a hand-wired agent setup. We require a signed token carrying a caller identity on every route and refuse anonymous access, so each tool call resolves to one named service or person, with entries retained for the six-year window in the region you chose.

FAQ

Frequently asked questions

Three groups of questions. The first is about running this specific proxy, answered by the people who run it. The second is about moving an existing instance across. The third is about the platform underneath, and applies to every application we host.

01

Technical and configuration

Both. It can launch a process-based Model Context Protocol server as a child process and expose it over streamable HTTP, so a tool that only ran on a laptop becomes reachable by a hosted agent. Remote targets are proxied directly on the same listener.
Yes, both on the same route, which matters because clients differ in what they negotiate. Both can be kept enabled by default, so an older client and a newer one can share a single endpoint instead of needing two deployments.
Almost always the cross-origin header sets. The session identifier and protocol version headers have to appear in both the allowed and the exposed sets, and a browser client fails quietly rather than usefully when they are missing. Both sets can be pre-populated at setup.
Access rules are written per tool and evaluated against the verified claims in the caller's token. A read-only search tool can stay open while a writing tool is limited to one group. The tool list is filtered per session, so a client never sees what it cannot call.
The new configuration is validated first, then swapped in as a whole. Requests already in flight finish against the previous version and new requests use the new one. If validation fails, the previous configuration keeps serving and the error is logged.
No. A file that fails validation is rejected and the last working version stays active. The subtler trap is the built-in editing interface, which rewrites the file it manages and drops hand-written comments. That interface can be kept off the public listener.
Listeners, routes, backends and policies reload from the watched file without dropping traffic. The top-level block holding logging and the administrative address does not, so an edit there does nothing until the process restarts. We schedule those with you.
As environment references rather than literal values inside the configuration file, so the file can be edited, copied and version-controlled without carrying the secret. Where an upstream accepts a short-lived token, we prefer token exchange over a long-lived static key.
Increasingly, yes. Session identifiers carry a lifetime and the protocol has moved toward stateless operation, so a second instance can serve a request the first one started. Clients that still pin a session keep working and reconnect if their instance goes away.
Traces go out over OpenTelemetry and metrics are exposed in Prometheus format, including token counts per model call and per session. Request and response bodies, prompts included, can be recorded when you want that. It can be left off, because prompts usually carry customer data.

02

Migration and onboarding

21 datacenter locations across six continents. You choose the region at provisioning. Application data sits in the region you choose; pick whichever is closest to your users or matches your data-residency preference.
Yes. Request a region migration from the dashboard and we run the move in the background. The system emails you when the migration completes; total transfer time depends on data volume but typical instances finish in a few hours. There is no extra charge for a region change.
Yes. Full data export is available at any time, in a portable format you can bring to any infrastructure.
Send the configuration file and the environment references it expects. We start it in the region you pick, put a staging hostname in front of it, and let you exercise your agents against that before anything moves. The file format is the one you already use.
The parts we control are quick: a file, a set of credentials and a hostname. The pace is set by your side, namely re-issuing provider keys and confirming each tool target still answers. We do not quote an activation window for a step we do not schedule.
Yes. Custom domains are included and certificates are issued and renewed for you. Agents that address the gateway by hostname keep working once that name resolves to the new address, and we usually stage the change behind a second hostname first.
Bring the issuer and audience settings across and existing tokens keep validating. If you are changing identity provider at the same time, both issuers can be accepted on the same route while clients rotate, and the old one is removed afterwards.
Yes. Browser login through an OpenID Connect provider is supported by the software itself, and we will wire the issuer, client and redirect settings with you. Machine callers usually keep using signed tokens or keys, and both can run on one instance.
Yes. What you have is a plain configuration file, your recorded output and a public container image; no part of it is a format only we can read. Ask and we hand back the file and a copy of the data.

03

Billing, support, and platform

€9 covers everything we do for that app: hardware in the region you choose, daily off-site backups with one-click restore, automatic security patches and version upgrades, 24/7 monitoring, SSL and firewall, and engineering support on email and the ticket section in the client portal. There are no setup fees or hidden line items. For more info see our Pricing page.
If you decide to continue, we charge €9/app/month from day 8. If you don't, the trial ends and you can export your data. No card is required for the trial, and we never auto-charge you without explicit consent.
No. The €9/month is flat regardless of how many users log into your app. Add 5 users or 50; the price doesn't change.
24/7 email and ticket support, both staffed by engineers who run the systems. We handle DNS configuration, SMTP setup, app integrations, performance tuning, troubleshooting, and migration help. Response time is typically under an hour. There is no tier system — every customer gets the same support.
Yes. Cancel from the dashboard. We don't charge a cancellation fee, we don't lock data, and we will export your data to you on request before deletion. data to you on request before deletion.
Every customer instance is backed up daily to a separate region from the primary. We test restores. You can request a restore at any backup point within the retention window — usually 7 days for daily backups.
Your application data sits in the region you choose at provisioning — 21 datacenter locations across six continents. Account-level data (billing, account email, support ticket history) is processed centrally. Application data region is picked by you, per app.
99.9% uptime SLA on every app, every tenant. Service credits are documented at danian.co/service-level-agreement. The status page is located at status.danian.co.
When your tenant approaches the resource ceiling — the base tier holds 1 vCPU/RAM, 30 GB storage — we notify you. Resource upgrades happen with your explicit consent; we will not upgrade your tenant or charge you without it.
We wait. We don't suspend the app or delete your data on the first failed charge. We email you, you fix the card on file, and we continue.
Invoices can be downloaded from the billing dashboard in PDF the day each charge succeeds. EU VAT is added where applicable and the VAT-reverse-charge regime applies for VAT-registered businesses with a valid number.
150+ open-source apps across automation, team chat, file sync, analytics, AI, password management, email marketing, dev tools, project management, smart home, CMS, and federated social. See the full catalog →
Yes. Every instance comes with a web-based terminal and a file manager in your DANIAN management dashboard. Useful for managing your data and customizations.
Resources scale with your usage. If your app needs more vCPU, RAM, or storage, we add it — and we ask first before any change to your plan. €9 is the floor; resource-heavy workloads may price higher, but you'll always know in advance.
Yes. We have both a Partner program and an Affiliate program available. Anybody can sign up.
No contract. No minimum commitment. Cancel anytime from the dashboard with one click. The 7-day free trial requires no credit card. After the trial converts to paid, you can still cancel at any month without notice or penalty.

DEPLOY IN YOUR REGION

21 datacenter locations on six continents

Pick the region closest to your users.

United States, Germany, Finland, Singapore, Australia, Brazil, Canada, Netherlands, UK, Spain, Italy, France, Sweden, Malaysia, India, Japan, Mexico, Poland, South Korea, Chile, South Africa and more coming soon

Global Reach Map

Try managed Agentgateway for 7 days

No card. Cancel from the dashboard.