Runtime Review (2026): Pricing, Features & Honest Verdict

Runtime is a control plane for AI coding agents: sandboxed cloud machines where Claude Code, Codex and Gemini CLI run under your team’s guardrails, triggered from Slack, Linear or GitHub. The product is sharp and the paperwork is a year out of date. Best for: engineering teams running agents past one laptop. Price: Free, then from $99/seat/mo. Rating: 7/10.

What is Runtime?

Runtime (runtm.com) is a Y Combinator company that gives coding agents somewhere safe to work. You point it at a repo, it builds a snapshot of your environment, and every agent session gets its own isolated machine with the tools, secrets and permissions you decided in advance. Claude Code, Codex, OpenCode, Gemini CLI and GitHub Copilot all run inside it. The pitch is that agents are already in your company and nobody can see what they are doing.

What we found is a company mid-pivot. The GitHub organisation was created in December 2025 and the first commit landed on 2 January 2026. The original product was a cloud IDE for coding agents. The homepage today sells something wider: agents that read a Zendesk ticket, query Postgres, trace a Stripe webhook and draft eleven refunds for a human to approve, with the four named use cases being technical support, engineering, implementation management and incident response. Forbes covered the company in June 2026 as part of a piece on YC’s latest batch. The repositioning is recent enough that large parts of the site have not caught up, which turns out to be the most interesting thing about reviewing it.

What are Runtime’s key features?

One sandboxed machine per agent session

Every session boots its own cloud environment with a full filesystem, a terminal and a live preview URL. Nothing is shared between sessions. The free tier gives you a 4 vCPU / 4 GB box and one session at a time; Teams doubles that to 8 vCPU / 8 GB with unlimited parallel sessions; Enterprise goes to custom sandboxes up to 120 GB of RAM or lets you plug in your own execution layer if you already run Firecracker or gVisor. This is the part of the product that works exactly as advertised.

Environment templates that skip the Docker file

You import a repo once, Runtime installs the toolchain, and it snapshots the result. Later sessions boot from that snapshot in seconds instead of reinstalling everything. The docs claim this covers monorepos and microservices without you writing Docker or Terraform. For a team where four people each keep a slightly different local setup, the snapshot is the actual product: it makes “works on my machine” a property of the org rather than the laptop.

Guardrails with approval gates

Each agent only sees the tools you grant it. Databases can be mounted read only, individual actions can be marked as needing human sign-off, and Enterprise adds a network allowlist per agent. Runtime’s example is a refund agent that can read Stripe and the charges table freely but must stop and ask a human before refunding more than $500. Credentials are injected at run time and the security page says they never touch disk.

A stored record of every run

Prompt, steps, tool calls, files touched, duration, cost and result are all kept and searchable. There is an org-level dashboard breaking spend down by agent, user and team. Agents also read their own previous runs before starting a new one, which is the closest thing here to a moat: the value compounds inside your account rather than in the model.

How much does Runtime cost?

Runtime starts free and the first paid tier is listed as “starting from $99 per seat per month”, with Enterprise priced by conversation. The free plan includes 500 credits, one user, one parallel session and access to three agents. Teams unlocks unlimited users and sessions, all agents, real-time collaboration, shared templates and secrets, role-based access control, spend controls and priority support. Enterprise adds SSO, SCIM, audit logs, self-hosting, SLAs and a named support engineer.

Against the category that is expensive at the low end and reasonable at the high end. Ona (formerly Gitpod) starts at $20 a month with add-on credits from $10 per 40 OCUs. E2B gives a free Hobby tier with a one-time $100 of usage credit and charges $150 a month for Pro on top of per-second compute. Runtime’s $99 per seat has no usage-based component published at all, which is unusual in a category that otherwise meters everything, and for a five-person team it is $495 a month before a single model token.

Two things about that price deserve attention. The word “from” is doing real work: there is no published ceiling, no seat minimum and no annual discount anywhere on the page, so $99 is a floor rather than a rate. And there is no published startup discount either, only an FAQ line saying “if you’re early stage, reach out and we’ll work something out.”

What does Runtime not tell you about credits and self-hosting?

Two gaps are large enough to change how you buy this.

The first is the credit. Runtime’s free plan advertises “500 credits included” and that phrase is the only appearance of the word anywhere we could find on the site. It is not defined on the pricing page, not in the FAQ underneath it, and not once in the documentation at docs.runtm.com. There is no rate card, no conversion to tokens, minutes or sessions, and no statement of what happens when the 500 run out. You cannot work out from published material whether 500 credits is an afternoon or a quarter. Meter it on the free tier before you commit anyone to a seat.

The second gap runs the other way, and it is in your favour. The pricing table lists “Self-host Runtime” as an Enterprise-only line with a dash against Free and Teams, and the deployment row reads “SaaS / SaaS / SaaS or self-hosted”. The documentation contradicts that on its own front page, which offers “Open Source (Self-Hosted): the open-source core you can run on your own infrastructure, deploy with Docker Compose and bring your own cloud provider, full control over compute, networking and data residency.” The repository backs the docs up: runtm-ai/runtm licenses the control plane API, the build worker and the infrastructure configs under AGPLv3, and the CLI under Apache-2.0. GitHub reports the licence as NOASSERTION because of that split, not because it is closed. So the headline Enterprise feature is on GitHub under a copyleft licence, and the sales page does not mention it.

One more thing to weigh before signing. Bring-your-own-key is a dash on the free plan and a check on Teams. On the free tier you must spend the undefined credits; the only way to pay Anthropic or OpenAI directly is to buy a $99 seat first. Paying a seat fee for the privilege of also paying your own model bill is a defensible product decision, but it makes the free tier a demo rather than a plan.

How stale is the rest of the site?

We check legal pages against marketing pages on every review, and here the two do not describe the same company. Runtime’s Terms of Service were last updated on 21 January 2026 and define the service as “a cloud development platform where you build software with AI coding agents in isolated sandbox environments, including workspace management, live preview, terminal access, GitHub integration, and deployment.” The privacy policy repeats it. Every page under /compare is stamped “Updated Feb 2026” and describes Runtime as a platform with “live preview, multiplayer, background agents, git workflows, and observability.” None of that language matches the homepage that now sells Slack-triggered support and incident-response agents, and the pricing page still says Teams is “for teams collaborating on production software.”

The terms also never name a legal entity. They are governed by Delaware law and operated by “Runtime”, with no Inc., no LLC and no registered address on the site. Total liability is capped at twelve months of fees or $100, whichever is greater. Set against that, the security page claims SOC 2 compliance, says the audit ran against the AICPA Trust Services Criteria and offers the report on request through a trust centre at trust.runtm.com. An unnamed operator with a real SOC 2 report is a strange combination, and legal review at any company large enough to care about SOC 2 will ask who the counterparty is.

The open-source side is small but real. The repo has 294 stars, 33 forks, 23 open issues and 16 tagged releases, with the last push on 21 August 2026. Three contributors have committed, and the top one has 76 of the roughly 100 commits. That is a young project with concentrated key-person risk, which is normal at this stage and should still be priced in.

Who is Runtime best for?

Use Runtime if your team already runs coding agents and the problem has stopped being capability and started being control. The combination that justifies the seat price is guardrails plus a stored record of every run: knowing which agent touched production, what it cost and who approved it. Teams with non-engineers who want to trigger agents from Slack or Linear get more from it than teams of five backend developers who are happy in a terminal. It is also a reasonable fit if you want the option of moving the control plane into your own VPC later, because the AGPL core means that door is open without an Enterprise contract.

Skip Runtime if you are one person. The free tier gives you one session, one 4 vCPU machine, no API key of your own and an unquantified credit budget, and $99 a month for a solo seat buys you far more compute elsewhere. Skip it if your legal team needs a named counterparty before a security review starts. And skip the cloud entirely if data residency is the whole reason you are shopping: self-host the AGPL build and pay nothing.

What are the best Runtime alternatives?

box by ASCII is the cheapest way to get the same shape of thing. Persistent Ubuntu VMs with SSH, snapshots and forking, from $20 a month for about 555 hours of a 4 vCPU / 8 GB machine, billed by the second. You get no control plane, no approval gates and no audit trail, so you are buying the machines and building the governance yourself. Its servers are in the EU only.

Ona (formerly Gitpod) starts at $20 a month on its Core plan with add-on credits from $10 per 40 OCUs, and scales to machines with 32 vCPUs and 128 GB of RAM. It comes from the dev-environment world rather than the agent world, so it is stronger on standardised workspaces and weaker on agent-specific guardrails. It states plainly that it never uses your code to train models, which Runtime’s site does not say anywhere.

E2B is the infrastructure layer most agent products are built on rather than a finished platform. The Hobby tier is free with a one-time $100 of usage credit and a one-hour session cap; Pro is $150 a month, lifts sessions to 24 hours and lets you set CPU and RAM, on top of per-second compute at $0.000014 per vCPU per second and $0.0000045 per GiB per second. Pick it if you are building a product for your own users, not tooling for your own team.

Final verdict: is Runtime worth it?

The product is better than its website. Sandboxed sessions, environment snapshots, per-agent permissions, approval gates and a searchable record of every run are the right set of things to build, and the free tier is enough to see whether agents survive contact with your repo. Teams at the stage where three people are running Claude Code against production and nobody can answer “what did it change” will get value from the $99 seat quickly.

What holds the score at 7.0 is everything around the product. A credit is the unit you are spending on the free plan and it is defined nowhere. Your own API key costs $99 a month to plug in. The terms, the privacy policy and every comparison page describe a February product that no longer exists, and no legal entity is named beside a SOC 2 claim. Self-hosting is sold as an Enterprise feature and published under AGPLv3 on GitHub, which is good news you have to find yourself. Buy it for the guardrails, meter the credits before you commit seats, and read the repository before you read the pricing table.