Skip to main content

Claude Agent SDK·n8n·RAG·MCP·Voice·Evals

AI agents and workflow automation, built as code in a repo you own.

You describe the manual job in a few sentences. A senior engineer ships the agent that does it: support chat answering from your own docs, lead enrichment, internal ops agents, voice, content pipelines. Every one tested with evals before it touches live traffic, every one running on your infrastructure with your API keys.

Cancel us and the agent keeps running, because it was never ours.

Your repo, your infra, your keys · Evals before live traffic · Month to month or one-off

Map my workflow, freediagram in days

Two fields here. The brief below arrives pre-filled, three fields to go.

One run, traced

Watch a single run, including the part that goes wrong.

Every agent we ship traces its runs like this. Timestamps, what each step read, a retry when a system times out, and an eval gate that stops low-confidence output before a person ever has to catch it. This is what “you can trust the automation” looks like when it’s written down.

A sample agent run traced step by step: trigger, classification, a CRM fetch that retries after a timeout, a policy lookup, a drafted reply, an eval gate, human approval, and the sent-and-logged result.
run · support-inbox-triage
  1. 09:14:02

    trigger · inbound email

    New message in the support inbox. The agent wakes up; nobody had to notice.

  2. 09:14:03

    classify intent

    Read the email, matched it to “billing question” with the categories you approved.

  3. 09:14:03

    fetch CRM record

    First attempt timed out at 3s. Logged, waited, retried.

  4. 09:14:07

    fetch CRM record · retry 1

    Customer record and plan pulled. The retry is in the trace, not swept under it.

  5. 09:14:08

    check policy · RAG

    Searched your own policy docs and cited the paragraph it will answer from.

  6. 09:14:10

    draft reply

    Wrote the answer in your support voice, with the citation attached.

  7. 09:14:11

    eval gate · score 0.92

    Scored against the eval suite. Threshold is 0.85: below it, the run stops here and pages a person instead.

  8. 09:14:11

    held for human approval

    The draft waits in your queue. On day one everything holds; you loosen it per intent, category by category, when the trace has earned it.

  9. 09:31:47

    approved · sent · logged

    Reply sent, CRM updated, the whole trace stored. The person spent thirty seconds approving instead of fifteen minutes writing.

9 events · 1 retry · eval 0.92 ≥ 0.85 · human approved · full trace stored

Below the eval threshold, nothing sends without a person. Above it, you decide per category how much runs on its own, and the trace is always there to check.

Map my workflow, free

The agent is a file

“Built as code” means you can read it.

This is the run above, as the file that produces it, simplified for the page. Every behaviour is a line: the trigger, the steps, the eval gate, the approval hold. Your team can read it, diff it, and revert it, because it lives in your repository next to everything else you own.

agents/support-triage.tssimplified · the shape is the point
  1. export const agent = defineAgent("support-triage", {
  2. trigger: inbox("support@you.com"),← what wakes it
  3. steps: [
  4. classifyIntent(categories),← your categories
  5. fetchRecord(crm),← retries built in
  6. checkPolicy(rag(docs)),← cites your docs
  7. draftReply(voice),← your support voice
  8. ],
  9. evalThreshold: 0.85,← below this, a person
  10. approval: "human",← loosened per category
  11. trace: true,← the run log above
  12. });

$ git push · deploys from your repo, on your infrastructure · evals run before traffic does

No platform login we hold over you, no workflow trapped in a vendor’s canvas. If we disappear tomorrow, this file still runs, and any engineer can pick it up.

Map my workflow, free

agents.registry · what we build

The AI agents and workflow automation we build. Each shipped to a repo you own.

06 / 06 · code in your repo

agents · sample registry · illustrative
06 entries · 5 prod · 1 staging
  1. agent.ops@ v3.4.1

    Internal agents

    Syncprod

    Ops bots, call summarizers, lead enrichment, admin tools. The busywork your team does by hand, now a job that runs on its own.

    Signature

    ops: (task: Brief) → Result<Action[]>

    stack · Claude Agent SDK · n8n

    • Model

      Claude Opus

    • Latency · p95

      2.1s

    • Cost / call

      $0.018

    • Eval pass

      96.2%

  2. agent.rag@ v2.7.0

    RAG support chat

    Streamingprod

    Chat that answers from your knowledge base, not thin air. Every answer is grounded and cited. When it is not sure, it hands off to a person.

    Signature

    rag: (msg: Msg, ctx: Conv) → Stream<Reply>

    stack · pgvector · LangGraph

    • Model

      Claude Sonnet · pgvector

    • Latency · p95

      1.4s

    • Cost / call

      $0.014

    • Eval pass

      94.1%

  3. agent.voice@ v1.9.0

    Voice agents

    Realtimeprod

    Inbound qualifying, outbound calls, booking. The agent talks, transcribes, and writes the result back to your CRM.

    Signature

    voice: (audio: Stream<PCM>) → Stream<Turn>

    stack · Vapi · Retell · WebRTC

    • Model

      Haiku · Vapi · Retell

    • Latency · p95

      560ms

    • Cost / call

      $0.21 / min

    • Eval pass

      91.8%

  4. flow.automate@ v4.1.2

    Workflow automations

    Asyncprod

    Form-to-action chains, enrichment, lead scoring, reporting. Built on n8n, self-hosted. The manual handoffs just stop.

    Signature

    automate: (trigger: Event) → Run<Job>

    stack · n8n · Make · Zapier

    • Model

      Routed · Sonnet / Haiku

    • Latency · p95

      320ms

    • Cost / call

      $0.004

    • Eval pass

      98.4%

  5. agent.content@ v2.3.0

    Content pipelines

    Multi-agentprod

    One pillar becomes a week of posts, locked to your tone. The agent drafts, a senior editor reviews, you ship.

    Signature

    content: (pillar: Brief) → Atom[12]

    stack · Claude · n8n

    • Model

      Claude Opus · Sonnet 4.6

    • Latency · p95

      4.8s

    • Cost / call

      $0.42 / piece

    • Eval pass

      89.6%

  6. agent.admin@ v1.5.4

    Internal LLM tools

    Syncstaging

    Small web apps your ops team uses: tag records, summarize queues, batch-extract data. Grind work goes from hours to a click.

    Signature

    admin: (rows: Record[]) → Annotated[]

    stack · Next.js · Claude

    • Model

      Haiku

    • Latency · p95

      180ms

    • Cost / call

      $0.0006 / row

    • Eval pass

      95.3%

What we build · all of it traced and yours to inspect

The stack is yours to inspect, swap, or fork. Every choice is one a senior with about a decade can defend, with the evals to back it. One of 8 practices. Hire it alone or add more, never a forced bundle.

Map my workflow, free

Agent diagram · sanitized snapshot

The pipeline we actually ship.

AI automation services, in plain terms: a senior engineer turns a manual process into AI agents and workflow automation that run on their own. Five nodes, each with its own latency and cost. You get this back in your audit, drawn for your workflow. Evals on every node, tracing before live traffic.

Diagram v 04 · sample

Pipeline · support.deflection · prod
Sample · illustrative
  1. 01

    Trigger

    Form / Webhook

    p95

    80ms

    cost

    $0.001

  2. 02

    Retrieve

    pgvector / Pinecone

    p95

    120ms

    cost

    $0.0008

  3. 03

    LLM

    Claude Opus

    p95

    1.4s

    cost

    $0.012

  4. 04

    Eval

    Langfuse / Braintrust

    p95

    60ms

    cost

    $0.0003

  5. 05

    Action

    CRM / Slack / DB

    p95

    90ms

    cost

    $0.0005

Total p95

1.75s

Cost / call

$0.014

Eval pass · illustrative

94.2%

Sample agents · illustrative · code in your repo
Code in your repo
  • Support deflection bot

    Grounded answers, low-confidence to a humanillustrative
  • Lead enricher

    Form to enriched CRM recordillustrative
  • Call summarizer

    Every call summarized and filedillustrative
  • Content pipeline

    One pillar into a week of postsillustrative

This is what an audit reply looks like. Real diagram, real costs, real evals. No deck.

Map my workflow, free

Workflow audit · sample finding card

What we check before an agent ships.

Every engagement opens with an audit. FAIL items become sprint one. You see all of it, including what we would not touch yet.

Audit card · sample · finding rowsExported with the diagram
  • FAIL
    Hallucination management
    We check grounding, confidence thresholds, and the human fallback, because those three keep wrong answers away from customers.
  • FAIL
    Observability and tracing
    No tracing wired. You cannot see the prompt, context, or tool calls. Debugging is guesswork.
  • WARN
    Eval suite
    Zero evals. No way to catch quality slipping when a prompt changes.
  • WARN
    Vendor lock-in
    Hardcoded to one provider. No abstraction layer. Switching means a rewrite.
  • PASS
    Data residency and security
    Vendor confirmed not to train on your inputs. PII scrubbing in place.
  • PASS
    API key custody
    Keys stay in your vault, rotated on a schedule. No shared keys in the code.

Findings like these land in your inbox in days, with the diagram. Every build then clears two QA passes before live traffic.

Map my workflow, free
Operating principles · in writing
Principles I to IV

Four principles.
Wired into every agent.

Most AI projects fail for the same few reasons. These four put them on the table, in writing, before an agent goes live.

  1. PI

    You own the prompts, the code, and the infra

    Prompts and code live in your repo, version-controlled. Infra runs on your accounts. Token costs pass through at provider cost. We never hold your API keys.

  2. PII

    Evals and tracing on day one

    Tracing is wired before the first live call. Every prompt change runs the evals before it ships. You see what the agent did, why, and what it cost. No black box.

  3. PIII

    Low-confidence answers go to a human, not a guess

    A confidence threshold sits on every output. Below it, the answer routes to a person, not the customer. Every claimed fact shows its source.

  4. PIV

    Mutual NDA, and no training on your data

    Signed before we look at your workflows. Vendor contracts confirmed not to train on your inputs. PII scrubbing in place.

Operator's mark

You can cancel any month, month to month. Most clients stay: most clients stay well past the first project.

Eval ledger · sample build · v0.1 → v1.0

Every version runs the evals. None ships without them.

Production gate · ≥ 90% pass

evals.run · sample build
iterations · test cases
  1. v0.1
    Day 03

    Baseline. Zero-shot Sonnet, no retrieval, no HITL.

    84 tests · Senior engineer

    46.2%
    p951.9sCost$0.018
  2. v0.2
    Day 07

    pgvector retrieval over your KB. Confidence threshold at 0.72.

    124 tests · Senior engineer

    62.8%▲ 16.6
    p951.7sCost$0.014
  3. v0.3
    Day 12

    System prompt rewrite. Voice locked to your brand doc.

    188 tests · Senior engineer

    71.4%▲ 8.6
    p951.6sCost$0.014
  4. v0.4
    Day 18

    Human fallback below 0.65 confidence. Citations surfaced.

    246 tests · Senior engineer

    82.0%▲ 10.6
    p951.5sCost$0.013
  5. v0.5
    Day 22

    Router added. Haiku for tier-2, Opus for tier-1. Cost halved.

    312 tests · Senior engineer

    87.6%▲ 5.6
    p951.4sCost$0.009
  6. v1.0shipped
    Day 28

    Evals past the production gate. Rolled out to 100% traffic.

    312 tests · Account lead

    94.2%▲ 6.6
    p951.4sCost$0.009
Production gate · cleared (sample)
v1.0 · pass 94.2% · ≥ 90% required · cost ↓ 50%

The first agent reaches production in weeks. The diagram and audit land in days.

Map my workflow, free
AI questions

What people ask before they let an agent into the stack.

Plain answers on who owns the prompts, how we keep the agent honest, and what it costs.

  • You do, from the first commit. The agent is code in your repository, the prompts live in that code, and it runs on your infrastructure with your API keys. Cancel us and everything keeps running, because none of it was ever ours.

  • There is no rate card. You describe the workflow, we map it for free, and the map comes back with a fixed quote: a one-off build price or a month-to-month retainer, your call. Model and infrastructure costs pass straight through to you at cost, on your own accounts, with no markup.

  • Every run is scored against an eval suite before anything sends. Below the threshold, the run stops and a person gets it instead. On day one every action holds for human approval, and you loosen that per category only after the trace has earned it. When something does slip through, the full trace shows exactly which step went wrong, so it gets fixed in code rather than argued about.

  • Almost certainly. Agents connect to your CRM, helpdesk, database, inbox, and Slack through their normal APIs. Week one produces a written map of every system involved, its rate limits, and its owner, so surprises surface before the build starts rather than after.

  • No. It runs on its own, and the repo ships with docs written for an engineer who has never met us. Because the agent is plain, readable code rather than a locked vendor canvas, any developer you hire later can maintain or extend it.

  • The workflow map comes back in days and tells you exactly what the build involves. Simple workflows ship fast; nothing goes live before the evals pass on real examples from your own data, and the first stretch always runs with a human approving every output.

Don't see your question?Send a quick message →
Reply · within 1 business day
Node · AI-PRAC-2026/Closing brief
Closing brief

One step

Get your first workflow mapped, free.

Describe the manual work in a few sentences. In days you get a diagram of the automation, an eval plan, a cost estimate, the named engineer, and a price. Retainer or one-off, your call.

  • 01Free diagram and audit in days
  • 02A senior engineer named on day one
  • 03Code in your repo, evals before live traffic
  • 04Token cost passes straight to you
Node · audit.intake
v 1.0
Free workflow auditThe only form

Describe the workflow. The map comes back in days.

A few fields. One business day to a senior reply. No follow-up sequence.

Capacity is limited on purpose: seniors do the work, so we take a few new engagements at a time. The audit is how we choose.

Mutual NDA · Code in your repo · Token cost passthrough

Sealed · Read by a senior engineer

Workflow map · free

Production AI agents, not slides

Map workflow