Stateful, sandboxed, observable. Start on our cluster — drag-drop to your AWS, GCP, Azure, or on-prem when you're ready.
We didn't invent a new runtime. Agents are just well-behaved workloads on infrastructure you've already hardened. You get production reliability without running production yourself.
Memory, conversation state, and scratch volumes survive pod restarts. No “the agent forgot everything” after a zone drain.
Each agent gets its own egress rules, credentials, and budget caps. A rogue prompt in one sandbox can’t reach another.
We wrap the agent in the resiliency primitives Kubernetes gives you — readiness probes, graceful shutdown, PDBs — with zero YAML to maintain.
Standard wire formats. Ship to Datadog, Honeycomb, Tempo, Mimir — whatever your SRE team already runs.
Every deploy is versioned, every config change auditable, every bad release one click to roll back.
Run on our cluster for day one, promote to your AWS / GCP / Azure / on-prem later with the same manifest. No re-platforming.
Most agent frameworks hand you a demo and a prayer. Codiac Agents hands you four steps and a clear path to production.
Your agent shouldn't be held hostage by the platform that birthed it. Start on our cloud in minutes, then drag-drop the same agent — same manifest, same guardrails — into your cloud or your basement.
Start here. Zero setup, no card — productive in minutes.
AWS, GCP, Azure. Drag-drop the same agent into your account — stateful containers, your keys.
Air-gapped clusters, k8s, or bare metal. Same manifest, same control plane.
Every agent ships logs, metrics, and traces — Prometheus, OpenTelemetry, Grafana. Cost per agent, tokens per model, latency per tool call. No new pane of glass, no proprietary dashboards, nothing your SRE team has to re-learn.
Per-agent counters, gauges, histograms. Rate(), sum(), label-based drilldown — all the PromQL you already write.
Every tool call, every model invocation, every guardrail check — as spans. Ship to Tempo, Jaeger, Honeycomb, Datadog.
Pre-built dashboards for cost, token spend, latency, error rate. Import the JSON, or point at our recording rules.
If you're already running a cluster, we install into it directly — no replatforming, no sidecar lock-in. Agents run next to the rest of your workloads, governed by the same RBAC, NetworkPolicies, and PodSecurity you already configured.
Short answers to the questions we get on every demo call. If something isn't here, ask us — honest answers beat clever marketing.
Codiac Agents is the production runtime for AI agents. Every agent runs in its own sandboxed, stateful container — Kubernetes under the hood, turnkey on the surface. The agent itself is simple OpenClaw, which just works. What Codiac adds is the operational layer around it: sandbox, blast radius, observability, cost caps, and one-click promote from our managed sandbox to your AWS, GCP, Azure, or on-prem.
You pick. While an agent is running it's stateful — backed by Kubernetes StatefulSets with persistent volumes, health checks, and rolling updates, so a pod restart doesn't wipe its memory or lose conversation context. When you want a clean slate instead, the nuke button is one click and really does nuke: memory gone, tokens revoked, no orphan state. Crash-resistant when it matters, ephemeral when you say so.
Yes, and that's one of the things that makes this actually useful. When your agent writes a service, builds a tool, or needs a dev machine, Codiac hosts it right next to the agent — same cloud, same observability. The agent watches its own logs, redeploys, rolls back. The blast-radius walls still apply, so a hallucinated typo can't wipe prod: the agent only touches what you explicitly let it.
No. We don't train models, we don't run notebooks, we don't ship a feature store. This is the production runtime for agents that are already built — whatever framework, whatever model. Think 'Vercel for agents,' not 'SageMaker for agents.' If your team is doing MLOps, Codiac is downstream of that work, not a replacement for it.
Those are agent frameworks or ephemeral-only sandboxes. Codiac Agents is the production runtime that runs them — and lets you pick. Run an agent ephemerally when that's the point (one-shot jobs, demos, anything behind a nuke button) or let it run stateful so a pod crash doesn't wipe its memory mid-task. Build with any framework you like; Codiac adds Kubernetes-grade resilience when it's running, blast-radius controls, audit logs, cost caps, observability, companion-app hosting, and one-click promote from our cloud to yours. Frameworks orchestrate. Ephemeral sandboxes demo. Codiac operates.
No. Codiac is a real cloud operations platform — the same one used to run production clusters, deployments, and infrastructure for regulated customers. Agent hosting inherits every primitive: stateful workloads, multi-cloud portability, identity, secrets, audit, SLOs. It's not a thin marketing wrapper around someone else's compute and it's not a DIY Terraform starter-kit; it's production infra that happens to be turnkey.
Yes — and that's the whole point. Start on our cloud in minutes (no credit card, see below), then drag-drop the same agent into your AWS, GCP, or Azure account when you're ready. One click from the UI or a single CLI command, no YAML to babysit. Same manifest, same guardrails, same observability — just on your compute with your keys. Air-gapped and on-prem installs are supported for regulated workloads, on the same platform your DevOps team already uses for clusters and deployments.
Each agent gets its own isolated cell: separate compute, separate network egress rules, separate credentials, per-agent spend caps. The blast-radius view shows exactly what the agent can reach — tools, data, domains, budget — before it takes an action. One click on the nuke button wipes the agent, its memory, and revokes all its tokens. Nothing leaks past the sandbox wall.
Prometheus metrics, OpenTelemetry traces, Grafana dashboards — out of the box. Cost per agent, tokens per model, latency per tool call, error budgets, SLO-ready metrics. Standard Prom/OTel wire formats so you ship to any backend your SRE team already runs (Datadog, Honeycomb, Tempo, Jaeger, Mimir, etc.).
Claude Opus 4.7 is the default (and the extended-reasoning variant is always available), with first-class support for OpenAI GPT, Google Gemini, Moonshot, Fireworks, and Tencent providers. Bring your own API keys on any plan — the agent uses your keys for inference, so token cost and rate-limit policy stay with your accounts. Swap providers or drop a new model in without restarting the agent — `/models add` is a live hot-reload. Enterprise tier supports self-hosted model endpoints.
Yes. Every agent gets persistent memory backed by LanceDB with cloud storage — conversation history, scratchpad state, embeddings, skill configs all survive pod restarts, rolling updates, and zone drains. Combined with the stateful Kubernetes primitives underneath (StatefulSets + persistent volumes), "the agent forgot" isn't a failure mode you need to design around. Soft-reset the workspace without killing the agent if you want a clean slate mid-session.
Yes — scheduled agents are a first-class primitive. Define cron triggers, preview the resolved delivery targets before the job fires (so you know who's getting that 7am brief), and let the agent run whether anyone's watching or not. Job definitions and runtime state are stored separately so config edits don't clobber in-flight runs.
Yes to the trial, no to the card. Sign up, spin up your claw — no credit card required. You get 24 hours of live use on us; after that you hit a paywall, and one click on any plan picks up exactly where you left off. If you don't upgrade right away your claw drops into zombie mode: instance paused, every byte held in cold storage for 30 days. Come back inside that window and your agent, memory, skills, and config all restore — nothing to rebuild. After 30 dormant days with no upgrade we clean it up.
Flat monthly tiers based on instances, storage, and team size: Starter ($29/mo), Pro ($79/mo), Business ($199/mo). Enterprise is custom for on-prem, SLAs, and regulated deployments. Pricing is driven live from the billing API so it's always current — check the pricing section on the homepage.
HIPAA-ready sandbox is available on request. SOC 2 Type II is in progress. For regulated workloads we support dedicated sandbox clusters, custom model hosting, and fully air-gapped on-prem installs. Talk to us about compliance packets and signed BAAs.
Yes. Every UI action maps to a CLI call; every CLI call maps to an API endpoint. Agents can drive the CLI themselves — that's how we dogfood our own marketing-ops agent. Same mental model for humans and autonomous agents.
OpenClaw is the 348K-star agent harness — the simple runtime your agent actually runs on. No tech mumbo-jumbo, it just works. Codiac Agents is the platform around OpenClaw: stateful Kubernetes primitives so agents survive restarts (persistent memory via LanceDB), companion apps deploy next to the agent, observability ships on day one (OpenTelemetry, Prometheus, Grafana — enabled by default, not a bolt-on), costs cap at the gateway. You get 'it just works' on day one and production primitives on day 1,000 — same platform, no migration.
Yes — that's a first-class feature, not an afterthought. Export the manifest, run it on our cloud, your cloud, or your laptop. Our control plane is optional; the agent itself is portable. No lock-in is a design principle.
Start in our sandbox, promote to your cloud when it earns its keep. No seat fees during pilot.