Skip to content

Clawsseum is in early access. Request access

Now in early access

Build AI agents that work in production.

Clawsseum gives your team one SDK to build agents, 100+ adapters to connect them to your models, tools and channels, and evals that catch regressions before your customers do.

  • Free to start
  • Bring your own model keys
  • TypeScript and Python

support-agent / evals

Live

Eval run 2481

v13 (new) vs v12 (production)

dataset support/refunds, 412 cases from production

Win rate

61.3%

176 wins67 ties44 losses287 / 412 played
ConversationVerdict
  • Asks for a manager after denialv12 better
  • Refund over $500, needs approvalv13 better
  • Partial refund for unused seatsv13 better
  • Chargeback threat, hostile toneTie
  • Refund requested on day 34 of 30v13 better
  • Duplicate charge after plan upgradev13 better

Example eval run: a new version is tested on real production conversations before it is deployed.

102 adapters for the stack you already run.

  • Gemini
  • Mistral
  • Llama
  • Zendesk
  • Stripe
  • HubSpot
  • GitHub
  • Linear
  • Jira
  • Notion
  • Postgres
  • Snowflake
  • WhatsApp
  • Gmail
  • Shopify
  • MCP
  • OpenAPI

The problem

Agent demos are easy.
Production is hard.

Any team can get an agent to work once. Keeping it working across new models, more systems and real customers is where agent projects stall.

01

Models change constantly.

Every new model is better on average and worse somewhere specific. Teams either stay on old versions or upgrade without knowing what broke.

02

Integrations are glue code.

Every CRM, ticketing tool and chat channel is wired by hand, with its own auth, retries and failure modes. Most agent code ends up being plumbing.

03

Regressions go unnoticed.

A prompt change fixes one flow and breaks three others. Without tests on real conversations, customers find the problem first.

Platform

Everything you need to
ship agents.

Build agents with the SDK, connect them with adapters, test every release with evals, and run them on a runtime built for production. Use all four together or adopt one at a time.

01Agent SDKBuild

Define agents in code.

Write agents in TypeScript or Python with typed tools, durable memory and plans you can review in a pull request. Run them locally against real adapters or recorded fixtures.

  • Typed tools with zod or pydantic
  • Durable, resumable runs
  • Local runtime with hot reload
agents/support-agent.ts
import { agent, tool, z } from "@clawsseum/sdk";
import {
  openai, google, zendesk, stripe, slack,
} from "@clawsseum/sdk/adapters";

const refund = tool({
  name: "issue_refund",
  input: z.object({
    chargeId: z.string(),
    amount: z.number().positive(),
  }),
  sideEffect: true, // pauses for approval
  run: (args, { adapters }) =>
    adapters.stripe.refunds.create(args),
});

export default agent({
  name: "support-agent",
  model: openai("gpt-5"),
  fallback: [google("gemini-2.5-pro")],
  adapters: [zendesk(), stripe(), slack()],
  tools: [refund],
});
02AdaptersConnect

Connect to any model, app or channel.

Models, channels, apps and databases all share the same typed interface. Swap one model for another, or Slack for Teams, in config. Every adapter gets least-privilege scopes by default.

  • Scoped credentials per adapter
  • Retries, rate limits and fallbacks built in
  • Any MCP server or OpenAPI spec
support-agent / adapters6 healthy
  • O

    OpenAIModel

    gpt-5

    primary
  • Google GeminiModel

    gemini-2.5-pro

    fallback
  • S

    SlackChannel

    #support-escalations

    threads
  • ZendeskApp

    tickets:read, macros:read

    scoped
  • StripeApp

    charges:read, refunds:write

    approval
  • PostgreSQLData

    replica, billing schema

    read-only
$ clawsseum adapters swap model google:gemini-2.5-pro --eval
03EvalsTest

Test every release on real traffic.

Turn production conversations into datasets, run each new version against the one in production, and score both with model graders and deterministic checks. Ship on data, not on a gut feeling.

  • Datasets from production traffic
  • Model graders plus deterministic checks
  • Eval checks in CI
Quality score, support/refunds6 versions
v8v9v10v11v12v1388.7%

Win rate

63.8%

Ties

22.1%

Regressions

0 critical

04RuntimeDeploy

Run agents safely in production.

A durable runtime that pauses risky actions for approval, enforces budgets, redacts PII before it reaches a model and traces every step.

  • Approval policies for risky actions
  • Budgets and kill switches per agent
  • OpenTelemetry traces
run 8f12 / support-agent@v13Paused
  • zendesk.tickets.get
    118ms
  • model.plan
    842ms
  • stripe.charges.list
    204ms
  • model.decide
    611ms
  • stripe.refunds.create
    awaiting approval

Refund of $740.00 exceeds the $500 policy limit.

ApproveDenybudget $0.011 of $0.05

Workflow

From first commit to production in four commands.

The CLI, the SDK and the dashboard share one definition of your agent. What you run locally is exactly what runs in evals and in production.

  1. 01

    Create an agent

    A typed project with an agent, a tools folder, a starter eval dataset and a local runtime.

  2. 02

    Connect your tools

    Connect systems with OAuth or API keys. Scopes stay least-privilege until you widen them.

  3. 03

    Run evals

    Test the new version on recorded production traffic and compare it with the version in production.

  4. 04

    Deploy

    Ship with approvals, budgets and tracing on. Roll back to the previous version in one command.

~/acme/agents

$ clawsseum init support-agent --template support

created agents/support-agent.ts

created evals/support/refunds.dataset.ts (24 seed cases)

ok local runtime ready on http://localhost:7420

$ clawsseum adapters add zendesk stripe slack

zendesk connected tickets:read macros:read

stripe connected charges:read refunds:write (approval)

slack connected #support-escalations

$ clawsseum eval run --dataset support/refunds --baseline v12

412 cases 263 better 91 tie 58 worse

win rate 63.8% critical regressions 0 cost -22%

ok release checks passed: v13 is ready to deploy

$ clawsseum deploy --env production

ok support-agent@v13 live in production (us-east)

approval policy refunds>500 -> #support-leads

budget $0.05 per run, kill switch armed

$

Evals

Catch regressions before your customers do.

Evals that live in a notebook never block a release. Clawsseum runs them in your CI and before every deploy, so a worse version cannot reach customers by accident.

01

Record

Capture real conversations and tool calls from production, with PII removed before storage, and save the useful ones to a dataset.

02

Replay

Run every case against the new version and the version in production, side by side. Adapters that take actions run in a sandbox or from recordings.

03

Score

Model graders score outcomes, deterministic checks enforce your policies, and people review the close calls. Release checks decide what ships.

Version comparison

dataset support/refunds, 412 cases

#VersionModelScoreStatus
1support-agent@v13gpt-588.7%+3.7candidate
2support-agent@v12gpt-585.0%production
3support-agent@v13-geminigemini-2.5-pro84.6%-0.4experiment
4support-agent@v11gpt-582.6%retired
5support-agent@v13-mistralmistral-large81.2%-3.8experiment
pull request 1184checks
  • build

    Successful in 41s

  • unit tests

    312 passed

  • clawsseum / evals

    v13 better than v12 on 63.8% of cases, 0 critical regressions

  • clawsseum / policy

    refund limits, tone, PII: all clear

Ready to merge

Adapters

102 adapters.
No glue code.

Every adapter shares the same typed interface: actions with schemas, scoped credentials, retries, rate limits and redaction. Agents never see a raw API key, and every call is recorded so evals can replay it.

Bring your own systems.

Point Clawsseum at an MCP server or an OpenAPI spec and get a typed adapter with an allowlist, scoped auth and recording included. Internal services get the same guardrails as Stripe or Salesforce.

adapters/internal.ts
import { adapter } from "@clawsseum/sdk/adapters";

// Any MCP server becomes a scoped adapter.
export const github = adapter.fromMCP(
  "npx -y @modelcontextprotocol/server-github",
  { scope: ["issues:write", "pulls:read"] },
);

// So does any internal API with an OpenAPI spec.
export const billing = adapter.fromOpenAPI(
  "https://billing.internal/openapi.json",
  {
    auth: adapter.auth.oauth2({ secret: "BILLING" }),
    allow: ["GET /invoices/*", "POST /credits"],
  },
);

Runtime and security

Built for agents that take real actions.

Refunds, account changes, outbound email. When an agent can take actions, you need controls on what it is allowed to do.

Security overview

Bring your own keys

Use your provider contracts and rate limits. We never mark up tokens.

Scoped credentials

Each adapter gets least-privilege scopes from a vault. Agents never hold raw secrets.

Approvals

Risky actions pause for a human in Slack or the dashboard, based on your policies.

PII redaction

Sensitive fields are masked before any model sees them.

Budgets and kill switches

Cap cost per run, per agent and per day. Stop an agent everywhere in one click.

Traces

Every step, prompt and tool call, searchable, with OpenTelemetry export.

Data residency

Run the control plane and traces in the US or the EU.

Your cloud

Self-host the runtime in your own VPC when data cannot leave it.

Use cases

What teams build with Clawsseum.

The pattern is the same everywhere: an agent that reads from a few systems, takes action in one, and has to pass its evals before every release.

01
S

Support resolution

Resolve billing and account tickets end to end, with refunds above a threshold sent to a human.

release check resolution rate at least 85%, 0 policy violations

02

Engineering triage

Turn alerts and error spikes into deduplicated, assigned issues with a first diagnosis attached.

release check correct owner on 90% of cases, no duplicate issues

03
S

Revenue operations

Keep the CRM honest: enrich accounts, log calls, draft follow-ups and flag stalled deals.

release check zero writes outside allowed fields, human send on all email

04
N

Finance operations

Match invoices to payments, chase missing receipts and prepare month-end reconciliations.

release check 100% match accuracy on the golden set before any write

Early access

Ship your first agent this week.

We are onboarding a small number of teams each week. Leave your work email and we will get you set up.

Weekly cohorts. No spam, one onboarding email.

Need a security review, a custom adapter or your own cloud? Contact sales