AGENTKIT V2 IS LIVE

Ship AI agents that actually do the work.

The framework for building, deploying, and observing production AI agents. From prototype to scale in days, not months.

book a demo

start building

~/agent · run #4827
$agent run "audit Q3 invoices for anomalies"
→ initializing agent · gpt-5 · 12 tools available
step 1planbreak task into 4 subtasks
toolstripe.invoices.listperiod=2025-Q3
retrieved 1,284 invoices · $2.4M total
step 2analyzescanning for outliers, duplicates, off-pattern amounts
toolpython.executestatistical_outliers.py
flagged 7 invoices > 3σ from mean
toolpostgres.querycross-reference vendor_history
step 3verify3 of 7 match known vendor patterns · 4 require review
step 4reportdrafting summary + slack notification
toolslack.postchannel=#finance-alerts
✓ done · 4 anomalies flagged · 6.0s · 12 tool calls · 0.34¢
$
agent run trace· 767540.0s

Trusted by 500+teams

MeridianVanguard°/EquinoxOAKLINELumina.

Everything you need to ship production agents.

A complete framework for building, deploying, and observing AI agents — from first prototype to millions of runs in production.

RUNWAY

Agent Execution

Run agents in production with auto-retry, parallel execution, and structured output. Spin up workers without managing infra.

RUNWAY

Tool Library

Connect agents to your stack with pre-built tools for Postgres, Stripe, Slack, GitHub, and more. Or define your own in TypeScript.

RUNWAY

Context & State

Persistent memory across runs, sessions, and users. Vector search and structured state, handled without a separate database.

RUNWAY

Multi-Model Routing

Route between GPT-5, Claude, Gemini, and open models per task. Failover, fallback, and cost optimization built in.

RUNWAY

Observability

Every run, tool call, and token logged. Replay any execution, debug failures, and ship fixes without guesswork.

RUNWAY

Security & Compliance

SOC 2 Type II, GDPR, and on-prem deployment. Granular role-based access, full audit logs, and data residency you control.

From chaos to clarity in ship production agents.

A comprehensive suite designed specifically for the complex workflows of modern legal and compliance professionals.

Define your agent

agent.yaml
agent:
name: invoice-auditor
model: gpt-5
tools:
- stripe
- postgres
- slack
guardrails:
max_cost: $1.00
timeout: 30s

Deploy with one command

deploy.log
→ building agent.yaml
→ resolving 12 tools
✓ build complete · 1.2s
→ pushing to us-east
→ allocating workers
✓ deployed · v1.4.2
live · 4 regions · 0 errors

Observe and iterate

trace · run #4827
step 1plan4 subtasks
toolstripe.invoices
→ 1,284 retrieved
toolpython.execute
→ 7 outliers
step 2verify
toolpostgres.query
done✓ 4 anomalies · 6.0s · 0.34¢

Define your agent

agent.yaml
agent:
name: invoice-auditor
model: gpt-5
tools:
- stripe
- postgres
- slack
guardrails:
max_cost: $1.00
timeout: 30s

Deploy with one command

deploy.log
→ building agent.yaml
→ resolving 12 tools
✓ build complete · 1.2s
→ pushing to us-east
→ allocating workers
✓ deployed · v1.4.2
live · 4 regions · 0 errors

Observe and iterate

trace · run #4827
step 1plan4 subtasks
toolstripe.invoices
→ 1,284 retrieved
toolpython.execute
→ 7 outliers
step 2verify
toolpostgres.query
done✓ 4 anomalies · 6.0s · 0.34¢

Built for every team shipping agents.

Whether you're a solo founder shipping your first AI feature or a platform team running thousands of agents in production, Agentkit gives you the same primitives — and grows with you instead of forcing a rewrite.

agentkit · productionlive · run #4827
agent.ts
1import { Agent } from "agentkit"
2
3// audit Q3 invoices for anomalies
4const agent = new Agent({
5model: "gpt-5",
6tools: ["stripe", "postgres", "slack"],
7memory: true,
8trace: true,
9})
10
11await agent.run(prompt)
execution trace
step 1plandecompose into 4 subtasks0.2s
toolstripe.invoices.listperiod=Q30.8s
→ 1,284 invoices · $2.4M
toolpython.executeoutliers.py1.3s
→ 7 invoices > 3σ
step 2verifycross-reference vendors2.1s
toolpostgres.queryvendor_history2.6s
→ 4 require review
done✓ done4 anomalies · 12 tool calls · 0.34¢3.2s
runs / 30d1.2M+12%
p50 latency1.4s−8%
cost / run$0.34avg/run
error rate0.02%−40%

book a demo

start building

Usage-based pricing that scales with you.

Start free, pay for what runs. No seats, no platform fees, no surprise overages — just runs, tokens, and the tools you actually use.

HOBBY

Free

Includes 1,000 agent runs/mo

Models: GPT-4o, Claude Haiku

Tools: 5 built-in

Support: Community

Try it for FREE

HOBBY

Free

Includes 1,000 agent runs/mo

Models: GPT-4o, Claude Haiku

Tools: 5 built-in

Support: Community

Try it for FREE

PRO

$20/mo + usage

Includes 25,000 runs included

After: $0.002/run

Tools: All tools + custom

Support: Email + Slack

start building

Enterprise

Custom

Includes Unlimited + SLA

After: Custom rate

Models: + private models

Tools: + on-prem connectors

Support: Dedicated CSM, SOC 2

book a demo

Used by teams who ship

From prototype to production. No seat costs. No guessing. Just runs that work.

We collapsed three tools into Agentkit. Our engineers stopped arguing about runtimes and started shipping faster. It's the first agent platform that actually felt native to both prototype and production.

Sana Lindqvist

VP of Engineering at Meridian

Replay traces alone paid for the migration. We caught a tool-call regression in minutes that would have shipped to production on our old stack.

Marcus Okafor

Staff Engineer at Vanguard°

Our agents went from demo-grade to dependable the week we switched. Multi-model routing meant we stopped writing failover code by hand.

Priya Anand

Founder at Equinox

We were spending $40k a month on GPT-5. Switched to Agentkit's router and dropped to $11k by sending 70% to smaller models. No quality drop.

Aïda Renault

Head of AI at Lumina

Our support team now resolves refund requests end-to-end. All the tool connectors worked out of the box. What took a quarter on our old stack took half a day.

Marcus Chen

Director of Operations at OAKLINE

SOC 2 and on-prem deployment were non-negotiable for us. Agentkit was the only platform that didn't ask us to compromise on compliance.

James Porter

VP of Security at Meridian

Questions? We have the answers.

How runs work, what you'll pay, and what happens when you scale. / No surprises. No footnotes. Just the answers that matter.

What counts as a run?

A run is one complete agent execution — from the initial prompt to the final response, including any tool calls, retries, and model invocations along the way.

What happens if I exceed my plan limit?

Nothing breaks. You'll be charged a per-run overage rate ($0.002/run on Pro). We'll notify you at 80% and 100% of your limit so there are no surprises.

Do you charge for failed runs?

No. If a run fails due to an Agentkit error, it's free. You only pay for successful agent executions that return a result.

Can I bring my own model API keys?

Yes — on Pro and Custom plans you can connect your own OpenAI, Anthropic, Gemini, or self-hosted endpoints. You'll only be charged for Agentkit infrastructure, not model tokens.

Is the free tier really free forever?

Yes. 1,000 runs per month, every month, no credit card required. It's a permanent free tier, not a trial.

Do you offer discounts for startups or open-source projects?

Early-stage startups (under $5M raised) get 50% off Pro for 12 months. Open-source maintainers get Pro free. Email startups@agentkit.dev to apply.

Can I self-host Agentkit?

On-prem and private VPC deployment are available on the Custom plan. The core runtime is also open-source and can run on your own infrastructure.

What's included in the Custom plan?

SSO/SAML, dedicated infrastructure, custom SLAs, audit logs, a named support engineer, on-prem deployment, and volume pricing on runs.

Can I cancel anytime?

Yes. Cancel from your dashboard in one click. Monthly plans stop at the end of the billing period. Annual plans are refunded pro-rata for unused months.

What counts as a run?

A run is one complete agent execution — from the initial prompt to the final response, including any tool calls, retries, and model invocations along the way.

What happens if I exceed my plan limit?

Nothing breaks. You'll be charged a per-run overage rate ($0.002/run on Pro). We'll notify you at 80% and 100% of your limit so there are no surprises.

Do you charge for failed runs?

No. If a run fails due to an Agentkit error, it's free. You only pay for successful agent executions that return a result.

Can I bring my own model API keys?

Yes — on Pro and Custom plans you can connect your own OpenAI, Anthropic, Gemini, or self-hosted endpoints. You'll only be charged for Agentkit infrastructure, not model tokens.

Is the free tier really free forever?

Yes. 1,000 runs per month, every month, no credit card required. It's a permanent free tier, not a trial.

Do you offer discounts for startups or open-source projects?

Early-stage startups (under $5M raised) get 50% off Pro for 12 months. Open-source maintainers get Pro free. Email startups@agentkit.dev to apply.

Can I self-host Agentkit?

On-prem and private VPC deployment are available on the Custom plan. The core runtime is also open-source and can run on your own infrastructure.

What's included in the Custom plan?

SSO/SAML, dedicated infrastructure, custom SLAs, audit logs, a named support engineer, on-prem deployment, and volume pricing on runs.

Can I cancel anytime?

Yes. Cancel from your dashboard in one click. Monthly plans stop at the end of the billing period. Annual plans are refunded pro-rata for unused months.

Usage-based pricing that scales with you.

Start free, pay for what runs. No seats, no platform fees, no surprise overages — just runs, tokens, and the tools you actually use.

book a demo

Start free, pay for what runs. No seats, no platform fees, no surprise overages — just runs, tokens, and the tools you actually use.

book a demo