Two firsts in one tool. An agent lifecycle that lets you build, eval, replay, and ship agents in one place — and a workspace where engineers and business users align at the development stage, not three sprints later when the customer complains.
Powered by Smart-Code: no API keys, no model-picking — we route to the right model per task and you start coding instantly. 500 credits free for 30 days, then $9.99/mo, or free forever on local Ollama.
Agent Foundry: Run, Trace, Evals, Tools, and Replay any production run in one surface. Cursor, VS Code & Windsurf optimize for writing code; we own the loop after you write the agent.
First IDE for agents 2Engineering ships the agent, the domain expert skills it, the compliance owner guard-rails it, operators monitor it — same project, same source of truth, while it's being built. Stop discovering misalignment at customer feedback.
#1 unique 3The frictionless on-ramp for developers and business users alike: intelligent routing across labs, automatic failover, a live token meter. The easy way in — especially if you don't want to manage API keys.
The on-rampTwo structural firsts — built for agents and devs + business co-building — with Smart-Code as the zero-friction way in.
Smart-Code is the easy way in — for developers and business users alike. It picks the right model for the question you just asked. Claude when you want fast. Codex when you need deep. Automatic cross-lab failover, a live token meter, always the right one. Never your problem.
No API keys · 500 credits free for 30 days · then $9.99/mo · or stay free forever on Local · Ollama (offline)
On every other team, the business writes a spec, engineering builds for three sprints, and the gap only surfaces when the customer complains. CodeNow is the first IDE where the domain expert and the engineer work the same agent, in the same project, while it's being built — skilling it, guard-railing it, and watching runs together. Alignment moves to the development stage, where it's cheap to fix.
Your agent worked in production last week. You changed the code today. Did you just break it? Every CodeNow run — local and production — is captured step by step: prompt, tool calls, model response, cost, tokens, latency. Click Fork to current code, and CodeNow re-runs that exact production input against the code on your machine right now, side-by-side with the original. No Langfuse tab. No Helicone account. No midnight pytest harness. No duct tape — it's in the editor where you already work.
On first launch, CodeNow scans your machine for credentials and configs from the tools you already use. One click — your AI keys, GitHub auth, git identity, and project remotes are all wired up.
The full comparison, row by row → Everything Cursor, Windsurf, Codex & Zed do, plus the agent + oversight layer they don't. As of v0.1.54 CodeNow ships SMART-Code with Token Economics — a multi-model coding agent that auto-routes between Anthropic Haiku 4.5 (speed) and OpenAI GPT-5 Codex (depth) per task, now with open-weight Kimi K2 + Qwen coders (Fireworks) selectable across SMART-Code, Cmd+K, autocomplete & the multi-model Studio, drag-portable Agent Foundry agents, screenshot/vision input, and cross-agent context continuity (all v0.1.54), with automatic cross-lab failover (if one lab is down or over quota, your turn lands on the other — no dead turns), cache-stable context blocks + per-turn cache-hit % surfaced in the routing drawer. The Token Economy Index measures $ per resolved task — the metric nobody else publishes. No model-picking. Metered in credits (1 credit = 1,000 tokens) — 500 free for 30 days, then from $9.99/mo; credits only meter CodeNow-served inference, and BYOK is always free. CodeNow also ships TrustModel AGP as built-in agent-governance — every agent in your Fleet gets a TrustScore from minute one (no separate signup, no separate bill). Plus full parity on the IDE essentials — Cmd+P, Cmd+Shift+F, command palette, @ mentions, Apply, Tab autocomplete, codebase Q&A, source-control GUI, Vim mode, real-time CRDT pair-programming (v0.1.28), Codex-style Task Hub (v0.1.32), CodeNow Sidecar (Mode C) pushing skills + guardrails into agents deployed in your own AWS/GCP/Azure cloud (v0.1.28), Agent Fleet auto-detection across every project (Claude Agent SDK · OpenAI Agents · Google ADK · Microsoft Agent Framework · LangGraph · CrewAI · MCP — v0.1.44), built-in agents (SMART-Code · Composer · Cmd+K — v0.1.44), import any MCP server or CLI agent into Business mode (v0.1.49), browser-as-agent-tool (navigate/eval-js/get-dom/console — v0.1.42), cross-agent context via memory.jsonl ledger (v0.1.42), open-tab + chat-history persistence across restarts (v0.1.46), Live Agent Replay, anonymous read-only share links, channel-separated prompt-injection defense, SHA-256 Merkle audit trail, Cisco DefenseClaw runtime (partnership →), mobile companion app at codenow.pro/app that can drive your desktop's live SMART-Code loop from your phone (Desktop Relay — v0.1.50), one-click deploy to Vercel / K8s / AWS / Azure / GCP, and the only Human Oversight surface for non-engineers — all in one window.
| Capability | CodeNow | Cursor | Windsurf | Codex CLI | Zed | VS Code |
|---|---|---|---|---|---|---|
| Core IDE features (parity as of v0.1.7) | ||||||
| Cmd+P fuzzy file finder | ✓ | ✓ | ✓ | CLI — n/a | ✓ | ✓ |
| Cmd+Shift+F project-wide search | ✓ | ✓ | ✓ | grep/rg | ✓ | ✓ |
| Cmd+Shift+P command palette | ✓ | ✓ | ✓ | CLI — n/a | ✓ | ✓ |
| Inline edit (Cmd+K) with streaming diff | ✓ | ✓ | ✓ | CLI prompt | Agent panel | via Copilot |
| Multi-file agentic edit (Cmd+I / Composer) | ✓ | ✓ | Cascade | ✓ | Agent panel | No |
| @ mentions for file / URL context in chat | ✓ | ✓ | ✓ | via stdin | ✓ | No |
| "Apply" button on AI-suggested code blocks | ✓ | ✓ | ✓ | CLI — n/a | ✓ | No |
| Tab autocomplete (ghost text) | ✓ | ✓ | ✓ | CLI — n/a | Zeta (open-weight) | via Copilot |
| Codebase Q&A panel ("Ask the repo") | ✓ | ✓ | Cascade | via prompts | Agent thread | No |
| Source-control GUI (stage / diff / commit) | ✓ | ✓ | ✓ | CLI — n/a | ✓ | ✓ |
| Real terminal (no sandbox) | ✓ | ✓ | ✓ | ✓ (it IS the terminal) | ✓ | ✓ |
| Where CodeNow goes further | ||||||
| CodeNow Smart Coding — dual-lab coding agent that auto-routes between Anthropic Haiku 4.5 (speed) and OpenAI GPT-5 Codex (depth) per task. No model-picking. 30 days free with 50K tokens/day across both, then $9.99/mo. Read the pitch → | Yes (v0.1.35) | Single-lab (Anthropic OR OpenAI) | Single-lab | Single-lab (OpenAI) | Single-lab (hosted) | No |
| Built-in agent — no API key required — bundled in $9.99 sub via cloud proxy (Anthropic + OpenAI keys provided by us) OR fully offline via local Ollama. Both included. | Yes (v0.1.34) | Bundled cloud only | Bundled cloud only | OpenAI account required | Bundled cloud only | BYOK only |
| CodeNow Sidecar (Mode C) — push skills + guardrails into agents deployed in your own AWS/GCP/Azure cloud; outbound long-poll, no inbound firewall holes (Docker image, k8s manifest, hardened systemd all shipping) | Yes (v0.1.28) | No | No | No | No | No |
| Pair-programming with live CRDT cursors — Yjs + Supabase Realtime, both devs editing the same file with presence; no screen-share required | Yes (v0.1.28) | Workspaces (no live cursors) | Limited | No (CLI) | Yes | Live Share |
| Audit-log forwarding from customer cloud — sidecar accepts agent runs on loopback, forwards to CodeNow tagged with sidecar_id; closes the runs loop for Mode C agents | Yes (v0.1.28) | No | No | No | No | No |
| 7-day trial counter + Stripe checkout in app — "Day N of 7 · Upgrade" pill in topbar opens $9.99 Payment Link; webhook flips plan automatically | Yes (v0.1.29) | No | No | No | No | No |
| Model Arena — every flagship model answers the same prompt side-by-side; "Pick winner" persists as adoption signal | Yes (v0.1.20) | No | No | No | No | No |
| Provider transparency drawer — every API call logged with cost, tokens, latency, cache hits. No silent rerouting | Yes (v0.1.20) | No | No | No | No | No |
| Capability tuning per-provider — surface each lab's premium knobs (extended thinking, reasoning effort, web grounding) per-project | Yes (v0.1.20) | No | No | No | No | No |
| Anonymous read-only share link — mint URL, recipient clones & opens project in viewer mode (Monaco read-only + revoke) | Yes (v0.1.21) | No | No | No | No | Live Share (requires install + sign-in) |
| Multi-project tabs in one window | Yes | No | No | No | No | No |
| Switch AI models in one click — 5+ models side-by-side | 5+ models | Cursor only | Cascade only | No | parallel agents (single thread) | via extensions |
| Cross-LLM knowledge portability — one source emits CLAUDE.md, AGENTS.md, GEMINI.md, .aider.conf.yml | Yes | No | No | No | No | No |
| Agent Foundry — Run/Trace/Evals/Tools/Prompts panel for agent-building | Yes | No | No | No | No | No |
| Live Agent Replay — every run captured, replayable, forkable to current code | Yes | No | No | No | No | No |
| Multi-harness scaffolds (Claude · Cursor · Microsoft · Google ADK) | 4 SDKs | Cursor SDK | No | No | ACP — bring-your-own | No |
| Per-project MCP connections (49 vetted, hosted & local) | Yes | Global only | Global only | No | Global only | Global only |
| API keys encrypted in the OS keychain | ✓ | user config | user config | No | user config | user config |
| Detects existing CLI auth (codex / gemini / claude) automatically | ✓ | No | No | No | No | No |
Prompt caching (cache_control: ephemeral) across providers | Yes | No | No | No | No | No |
Channel-separated prompt-injection defense (<untrusted-data> wrap) | Yes | No | No | No | No | No |
| SHA-256 Merkle hash chain over JSONL audit trail | Yes | No | No | No | No | No |
| Cisco DefenseClaw runtime (Skills/MCP scanner + Guardrail Proxy) | Yes | No | No | No | No | No |
| Embedded browser control — AI navigates, clicks, types, scrolls, screenshots the built-in browser panel | Yes (v0.1.35) | No | Yes | No | No | No |
| Mobile companion app — review runs, approve trust gates, edit skills from your phone | Yes (v0.1.35 — codenow.pro/app) | No | Yes | No | No | No |
| JIRA → multi-model context — click a ticket, all agents (Claude/Codex/Gemini) receive full ticket context automatically | Yes (v0.1.35) | No | No | No | No | via extensions |
| One-click deploy to Vercel / AWS / Azure / GCP / K8s — signed TrustScore artifact + GitHub Action wrapper | Yes (v0.1.26–27) | No | No | No | No | No |
| Human Oversight for Agents (PMs, Ops, HR, Finance, Sales, Marketing, Legal, CX, Executives) | ||||||
| Train — domain expert teaches the agent (Skillify, no code) | Yes | No | No | No | No | No |
| Guardrail — visual rules enforced as system prompt on every run | ✓ | No | No | No | Tool permissions | No |
| Monitor — every run captured with input/output/TrustScore | ✓ | No | No | No | No | No |
| Override — pause, replay, or yank a deployed agent in one click | ✓ | No | No | No | No | No |
| Dev + business user side-by-side on the same agent | Yes | No | No | No | No | No |
| Workforce view — humans + agents in one panel | ✓ | No | No | No | No | No |
| TrustScores on every agent output | ✓ | No | No | No | No | No |
| One-click publish to Agent21 marketplace | Yes | No | No | No | No | No |
| Where Zed goes further today (CodeNow roadmap) | ||||||
| Real-time CRDT collab (live cursors, simultaneous edits) | Yes (v0.1.8 — Yjs) | No | No | No | Yes | Live Share |
| Voice / video pair-programming | Via Slack / Discord / Zoom integration | No | No | No | Built-in | Live Share |
| First-class Vim mode | Yes (v0.1.8 — monaco-vim) | via extension | via extension | No | Yes (native) | via extension |
| Slack-style team channels in the editor | Yes (v0.1.8) | No | No | No | Yes | No |
| GPU-rendered native (Rust) — 120fps editing | Electron + Monaco | Electron | Electron | No | Yes (Rust) | Electron |
| Editor + agent open source | Wiki public; IDE source closed | No | No | No | Editor + Zeta + ACP | Yes |
| Tech ↔ Non-Tech Collaboration | ||||||
| One tool for engineers and any business user | Yes | No | No | No | No | No |
| Engineering ↔ business handoff in same workspace | ✓ | No | No | No | No | No |
| Audit trail of who trained / guard-railed / overrode each agent | ✓ | No | No | No | No | No |
| Pricing & token economics | ||||||
| Bundled coding agent — included in subscription, no API keys | Yes (Claude + Codex auto-routed) | Yes (single lab) | Yes (single lab) | OpenAI account required | Hosted (single lab) | No |
| Free trial — try before paying | 30 days · 50K tokens/day · both labs | Free tier: very limited | Free tier: very limited | No standalone trial | Free editor (paid models) | Free editor |
| Token spend transparency — see your daily usage live | Live meter · per-day reset · which lab answered each turn | Opaque credit pool drains | Opaque quota burn | Your OpenAI dash | Opaque hosted billing | via your own extension |
| Multi-lab side-by-side — Claude AND Codex in one agent, auto-picked | Yes — Smart routing per turn | Single-lab (pick at sub time) | Single-lab | OpenAI only | Single-lab | No |
| BYOK still available — bring your own keys for premium models (Opus / GPT-5 / Gemini 2.5) | Yes — per-provider tabs (Claude Code · Codex · Gemini · Aider · Copilot) | Yes, but Tab autocomplete still needs Pro sub | Individual users only · certain models · not Teams | Yes (OpenAI native) | Hosted-only by default | Yes |
| Local mode — fully offline · zero connectivity · unlimited | Yes (Ollama · always free) | No | No | No | No | via Continue extension |
| Subscription cost | 500 credits free for 30 days · then $9.99/mo (Developer Pro) · $29.99 (Business) · $39/seat (Team) | $20/mo Pro · $60/mo Pro+ · $40/seat Business | $20/mo Pro · $200/mo Max · $40/seat Teams | ChatGPT Plus / API only | Free editor · paid hosted models | Free |
Try with no credit card, no API keys. CodeNow's hosted proxy puts both Anthropic Haiku 4.5 and OpenAI GPT-5 Codex at your fingertips, with daily reset. After 30 days, $9.99/mo to keep both labs — or stay free forever on Local (Ollama, offline).
Cursor & Windsurf make you pick a model per turn. We pick for you in milliseconds — fast Claude for chat-like questions, deep Codex for refactors and reasoning. See how →
The chat panel shows your daily usage live — and tags every assistant message with the lab + model that answered. No "Pro+ credit pool drained" surprises. No black box.
Want Claude Opus or GPT-5 frontier or Gemini 2.5? Use the dedicated Claude Code, Codex, Gemini, Aider, Copilot bottom tabs — each runs the official CLI with your own keys. CodeNow Code Cloud is bundled; BYOK premium is parallel, not replacement.
Plus everything else $9.99 buys you: multi-project tabs · all 4 SDK scaffolds · Agent Foundry · Live Replay · Sidecar · Human Oversight surface · Task Hub · JIRA-in-IDE · pair-programming with CRDT cursors. CodeNow Smart Coding → · All release notes.
Beta users came from every IDE in the category. Here's what made them stay. Each one is a real complaint we've heard from devs running Cursor / Codex / Windsurf / Copilot daily — and how CodeNow handles it.
Cursor & Windsurf show a dropdown. CodeNow's Smart routing picks Claude for fast questions, Codex for hard ones — in milliseconds, automatically. You ask. We pick. You ship.
Every other IDE locks you to one lab. CodeNow bundles Anthropic Haiku 4.5 AND OpenAI GPT-5 Codex. One $9.99/mo. Auto-failover if either lab has an outage. Vendor lock-in is the old way.
No credit card. No API keys. No "your free credits expired after 200 messages." 50K tokens/day across BOTH labs for a full month. Then $9.99 — half what Cursor or Windsurf charges.
You work on 3 repos. Cursor wants 3 windows. Windsurf wants 3 windows. CodeNow gives you 3 tabs — each with its own AI model, terminal, MCP connections, agent state. Stop alt-tabbing.
Sixth sidebar tab. Every agent task = own git branch (codenow/tsk-xxx), status pill (running / needs-review / merged), conversation, files-changed. Cross-project. Ask Claude to "fix X" and it lives as a Task you can review, merge, or discard. See Task Hub →
Install Ollama, click Local mode. Every keystroke, prompt, response stays on your laptop. No proxy, no telemetry, no daily cap, no monthly fee. Privacy-first devs love this. Cursor & Windsurf don't ship anything close.
Every agent run captured as JSONL. Click Fork to current code → to re-run any past prompt against your current state, side-by-side diff. Cursor doesn't have a replay system. Neither does anyone else.
Push skills + guardrails into agents deployed in your own AWS / GCP / Azure VPC. Outbound HTTPS only, no inbound firewall holes. Docker image + k8s manifest + hardened systemd unit all ship. No other IDE has a customer-cloud agent push channel.
Build with Claude Agent SDK (TS or Python), Cursor SDK, Microsoft Agent Framework (Azure), or Google ADK (Vertex). One IDE, every harness. Pick the right tool for your enterprise stack, not whoever your IDE vendor partnered with.
CodeNow's terminal is a real node-pty shell. npm test runs against your real Node. gcloud deploy talks to your real GCP. Cursor's terminal is wrapped; Codex CLI IS the terminal but lacks the IDE. CodeNow is both.
GitHub · JIRA · Slack · HubSpot · Salesforce · Postgres · Stripe · 42 more. Scoped per project — your work repo doesn't share connectors with your personal sandbox. Cursor & Windsurf only support global MCP.
Yjs + Supabase Realtime — both devs editing the same buffer with presence indicators, no screen-share required. Cursor's Workspaces feature has no live cursors. Windsurf has limited collab. CodeNow shipped this in v0.1.28.
Click a JIRA ticket in the rail, all selected agents (Claude / Codex / Gemini) get the full ticket context — summary, priority, labels, description — at once. Stop pasting JIRA into chat. Branch auto-pin keeps your active ticket visible as you code.
Switch to Business mode — same window, no IDE chrome. Your PM trains agents via Skillify (no code), legal sets guardrails as toggles, ops monitors a Workforce panel, finance sees TrustScores on every run. Cursor / Windsurf / Codex are dev-only tools.
Bundle prompts + evals + tools + guardrails + skills into a deployable package, list on the Agent21 marketplace, start earning. Built-in monetization. No other IDE has a path from "I built an agent on Friday" to "it's making money on Monday."
Tried Cursor or Windsurf? You'll feel the difference in 30 seconds.
Whether your developers build it from scratch or your business user imports one from the Agent21 marketplace, the next steps are the same: skill it, guardrail it, monitor it, then list it back on Agent21 to monetize.
Devs write the agent in Agent Foundry. Run / Trace / Evals / Tools / Prompts — any stack (Python, TS, Rust, Go), any LLM. The agent runs on your existing infra.
Click Share → invite your domain expert as a collaborator. They open the same project, train it (Skillify), set guardrails, and monitor every run. Files in git, activity in Supabase — both sides see it live.
One click in Agent Foundry bundles the manifest (prompts, evals, tool sigs, guardrail policy) and POSTs to the Agent21 marketplace. Your agent gets a public listing with provenance — built-by, runs-at, data access, price.
Browse the Agent21 catalog from inside CodeNow, paste an Agent21 URL, or wire up an HTTP endpoint. CodeNow registers the agent with full provenance — who built it, where it runs, what data it touches.
Your domain expert layers company-specific knowledge on top, sets guardrails, monitors every invocation. The base agent runs wherever it ran before — your overlay stays here.
Publish your skilled, guard-railed version of the agent back to Agent21 — under a new slug, under your team's name, with your overlay baked in. Your domain expertise becomes a sellable layer on top of someone else's agent.
Every business deploying AI agents has the same fear: drift. The agent gets clever in the wrong way. Hallucinates a number. Acts on stale data. Says something off-brand. CodeNow's Business mode is built so the human responsible for the work stays responsible — and stays in charge.
Skillify lets a domain expert teach the agent the way the work actually gets done in your company. Every skill is versioned, attributable, and reversible.
Visual rules for data access, allowed actions, tone, escalation. Saved guardrails are sent to the agent as a system prompt on every invocation — so the human owner's policy is enforced, not assumed.
Every agent run is captured with input, output, tools used, and a TrustScore. Review activity at human pace, not log-tail pace.
One click to pause an agent, replay a prompt with a fix, or yank a deployed agent off its function. The human is always one click away from "stop."
Built for the engineers running 4–6 codebases at once. Vibe-coders, consultants, founders, agency teams.
All your projects as tabs in one window. No more 6 VS Code instances. Switch context instantly.
Claude Code, OpenAI Codex, Gemini, Aider, Copilot. Switch between AI models with one click — per project.
Your project context, rules, memory, and skills live in one canonical bundle. CodeNow generates the right format for every AI CLI — CLAUDE.md, AGENTS.md, GEMINI.md, .aider.conf.yml. Switch from Claude to GPT-5 to Gemini without re-onboarding. The only IDE that survives an AI vendor change.
Full system access. Push to GitHub, create Jira tickets, deploy to Vercel, send Slack messages. Everything your terminal can do.
API keys and access tokens encrypted at rest via your OS keychain (macOS Keychain, Windows DPAPI, Linux libsecret). Never written in plaintext, never echoed to your shell history.
Built-in Jira, Slack, Gmail, GitHub, Stripe, Figma. Per-project connections — each project shows only what it uses.
Not 47 menus. Not 200 settings. Just files, editor, terminal, and AI. Built for vibe-coders, not configuration warriors.
Same IDE for everyone. Shared project configs. Engineers on Mac, Windows, and Linux all use the same tool.
The first IDE built for agents. SDLC tools optimize for writing code. Agent Foundry models the loop after the code is written — Run, Trace, Evals, Tools, and Replay — so your agent isn't a script you hope works, it's a service you can prove works.
Cursor and VS Code optimize for writing code. Most builders today spend 80% of their time on the loop after they've written the agent — iterating prompts, eval-ing across models, debugging traces, testing tools. Agent Foundry is the panel none of them ship.
Paste a prompt, fan it out to Claude Opus, GPT-5, and Gemini 2.5 in parallel. Side-by-side cards with response, latency, tokens. Stop guessing which model is best for the job.
Every Run is auto-saved. Click any trace to inspect aggregate cost, fastest model, full system + user prompt, per-model output. Replay with one button. No browser-tab Helicone trip required.
JSON suites with six assertion types. Run-all against every model. Pass/fail diff vs your last run, so you catch regressions before the customer does. Snapshot a Run as an eval in one click.
Auto-discovers Python @tool and TS/JS tool defs across your project. Click Test, paste JSON args, see the real return. No more "let me write a 200-line test harness for one function."
Markdown files with frontmatter {{variables}}. Fill-in-modal substitution before Run. Every save is logged to history — diff today's prompt against last week's without ceremony.
Build with Claude Agent SDK, Cursor SDK (Composer 2 + multi-model), LangGraph, OpenAI Agents SDK, CrewAI, MCP, or raw fetch. Studio's eval/trace/replay/publish loop wraps your code — doesn't replace it. Cursor commoditized models. We commoditize harnesses.
Every agent invocation — local dev and Agent21 production — is captured as a step-by-step timeline. Tool calls, model responses, cost, tokens, latency. Click Fork to current code → to re-run the same input through your local agent and diff side-by-side. The screenshot Cursor can't ship: prod replay, not just dev sessions. Watch the 60-second demo →
What you actually need when you're shipping an agent, not a CRUD app.
| Capability | CodeNow | Cursor | VS Code | Windsurf |
|---|---|---|---|---|
| Parallel multi-model invocation in one click | ✓ | No | No | No |
| Model Arena — same prompt to every flagship model with adoption-signal voting | Yes (v0.1.20) | No | No | No |
| Trace explorer with cost / tokens / latency per call | ✓ | No | Extension | No |
| First-class eval suite + pass/fail diff vs last run | ✓ | No | No | No |
| Tool isolation runner (test a single tool with JSON args) | ✓ | No | No | No |
| Prompt versioning with variable substitution | ✓ | No | No | No |
| Snapshot a successful run → eval test case (1 click) | ✓ | No | No | No |
| Stack-agnostic (Python / TS / Rust / Go / any SDK) | ✓ | TS-leaning | ✓ | TS-leaning |
| Replay any production run + Fork to current code | Yes | No | No | No |
| Multi-harness scaffolds (Claude · Cursor · Microsoft · Google · …) | 4 shipped | Cursor SDK only | No | No |
| Open auditable code (no black-box agent runtime) | ✓ | No | ✓ | No |
Pick the runtime your enterprise already standardized on. CodeNow ships first-class scaffolds for the harnesses Microsoft shops, Google Cloud shops, and AI-native teams actually use in production — with the same lifecycle (Test → Replay → Evals → Publish) wrapping all of them.
The official Anthropic harness. Tool-use, MCP servers, subagents. Calls the API directly — no Claude Code CLI required.
npm i @anthropic-ai/claude-agent-sdk
Cursor's full agent harness — Composer 2 + any frontier model, sandboxed cloud VMs, subagents, hooks. Public beta. Drop into CodeNow's lifecycle wrapper.
npm i @cursor/sdk
Microsoft's official agent SDK (Semantic Kernel + AutoGen merged, v1.0 GA). Azure OpenAI native with AzureCliCredential or env-var auth. The default for the vast majority of enterprises living on Microsoft stack — Windows, Office, Azure, .NET for 50 years and counting.
pip install agent-framework
Google's Agent Development Kit — workflows, agent teams, native Vertex AI deploy. Optimized for Gemini 2.0, model-agnostic. One env-var flip from AI Studio (local) to Vertex AI (production).
pip install google-adk
Graph-based orchestration with cyclic flows, persistent state, checkpointing. The agent framework Python shops know.
Handoffs, tracing, guardrails — OpenAI's official agent framework with first-class GPT support.
Multi-agent crews with role-based collaboration. The framework when your "agent" is actually a team.
Expose your agent as a Model Context Protocol server. Plug into Claude, Cursor, CodeNow — anywhere MCP runs.
No SDK, no opinions. Just an Agent21 manifest you wire to your own invocation URL. For the harness CodeNow doesn't ship yet.
One IDE, every harness. Don't bet your company on a single vendor's runtime. Switch harnesses by picking a different kit — same lifecycle, same eval suite, same publish flow, same prod replay.
Every agent built in CodeNow comes pre-wired to TrustModel.ai — replay, cost tracking, guardrails, audit logs, eval harnesses, HITL approvals. No SDK install. No separate setup. Sign in with CodeNow → land in TrustModel Console for fleet-wide monitoring.
See what's built in →Each project remembers which model it uses. Use Claude for the React app, Codex for the Python service, Gemini for the docs.
Get CodeNow for your OS. Drag to Applications and launch.
CodeNow scans for keys and configs from VS Code, Cursor, Claude Code, gh CLI, and more. One click imports the lot.
Claude, GPT-5, Gemini, Aider — switch in one click. Your project knowledge follows you.
One-click connect to Jira, GitHub, Slack, and more via MCP.
The editor is free forever — full IDE, your own keys, local runs. Smart-Code is the easy on-ramp: start with no API keys (we pick the model and pay for the inference), 500 credits free for 30 days, then $9.99/mo — or stay free forever on local Ollama. Bring your own keys anytime, always $0 to us.
SSO (SAML/OIDC), SOC 2 Type II attestation, on-prem / VPC deployment, contractual SLA, indemnification, dedicated CSM, training pack. Volume discounts above 50 seats.
1 credit = 1,000 tokens. Credits only meter SMART-Code — the inference CodeNow serves and pays for. The Token Economy panel shows exactly what every task costs as it runs, so there are no surprises. Run out and you're never locked out: switch to your own API key (BYOK) and keep going free, or top up. Using your own keys is always $0 to us.
BYOK inference under your own provider keys is always free — we never mark it up. SMART-Code credits are billed monthly per the tier above. Cancel anytime; usage in the current billing period is honored. Terms · Privacy.