mOperator is meant to be forked. This is the order to do it in, and roughly how
long each step takes.
The whole agent is the agent/ directory. Nothing about it is hidden behind a
framework abstraction: instructions are markdown, tools are TypeScript files,
schedules are cron plus a prompt. npm run agent:info tells you exactly what the
framework found on disk, which is the fastest way to confirm a change landed.
agent/lib/config.ts is the single customization surface. Read it top to
bottom once — it is short, and every field is something you probably want to
change.
Set at minimum:
BOT_NAME=Opsy
MOPERATOR_ORG_NAME=Acme
MOPERATOR_TIMEZONE=America/New_York
MOPERATOR_FISCAL_YEAR_START_MONTH=2 # if your FY does not start in January
AUTHORIZED_USER_EMAILS=you@acme.com,ops@acme.comThen your conventions, which are the highest-value thing you can teach it:
MOPERATOR_CAMPAIGN_NAME_PATTERN=^(NAM|EMEA|APAC)-FY\d{2}Q[1-4]-[a-z0-9-]+$
MOPERATOR_CAMPAIGN_NAME_EXAMPLE=NAM-FY26Q1-webinar-observability
MOPERATOR_UTM_MEDIUMS=email,paid-search,paid-social,event,webinarIf your naming convention does not fit a regex, replace check_campaign_name
with something that does — it is one file, agent/tools/check_campaign_name.ts.
This is the difference between an agent that guesses at field names and one that
knows your org. When a marketer says "segment", which field do they mean?
Two ways, and you want both eventually:
/audience-vocab — the admin UI. Stored in Redis, live immediately, noAUDIENCE_VOCABULARY in agent/lib/vocabulary.ts — checked in, reviewed,An entry looks like this, and the avoid list is the part people underrate:
{
term: "segment",
aliases: ["customer segment", "tier"],
object: "Account",
field: "Customer_Segment__c",
description: "Primary segmentation tier, set by RevOps monthly.",
commonValues: ["Enterprise", "Mid-Market", "SMB"],
avoid: [
{ field: "Segment__c", reason: "legacy, stopped syncing in 2024" },
],
notes: "For Contact queries traverse Account.Customer_Segment__c.",
}The vocabulary is injected into the system prompt at session start, so an edit
takes effect on the next conversation.
agent/instructions/ is the agent's character. Two files are yours to rewrite:
00-identity.md — how it works and how it talks. The current version10-safety.md — how it treats approvals and limits. Change this if yourTwo files are generated and you should leave alone: 20-capabilities.ts renders
the active integrations, and 30-vocabulary.ts renders the vocabulary above.
Keep the base prompt short. Procedures belong in skills, which load only when
relevant.
agent/skills/ is where the accumulated knowledge of your ops team goes. A skill
is a markdown file with a description in its frontmatter; the agent loads it
when a request matches that description.
The description is a routing hint, not a title. Write it as the situation that
should trigger it:
---
description: Use when someone asks for the weekly exec metrics update, or for numbers that go in the Monday leadership deck.
---
Weekly exec metrics
Pull these five numbers, in this order, from these exact fields...
Good candidates from your own team: the report someone rebuilds every Monday, the
checklist for launching in a new region, the three-step dance your CRM needs for
a lead handoff, the thing that always goes wrong at quarter end.
Start by writing down what you find yourself explaining to new hires.
Delete what you do not use — every tool in the model's context is a chance for it
to reach for the wrong one.
To remove Marketo entirely:
rm agent/tools/marketo.ts
rm -rf agent/lib/marketo
# then drop its entry from INTEGRATIONS in agent/lib/integrations.tsYou do not have to do this for correctness — an unconfigured integration
contributes no tools — but a fork that only mentions the systems you actually run
is easier to read.
Three ship with the repo and all three are inert until you name a channel:
MOPERATOR_CAMPAIGN_DIGEST_CHANNEL=C0123ABC # Mondays
MOPERATOR_AD_SPEND_DIGEST_CHANNEL=C0456DEF # daily
MOPERATOR_TRIAGE_DIGEST_CHANNEL=C0789GHI # FridaysThe prompt in each schedule is the whole job description — edit it directly. To
add your own, drop a file in agent/schedules/; it becomes a Vercel Cron Job on
deploy.
Test one without waiting for its cron:
npm run agent # in another terminal
curl -X POST http://localhost:2000/eve/v1/dev/schedules/campaign-digestPick the behavior that would embarrass you if it broke, and pin it:
// evals/my-team/monday-report.eval.ts
import { defineEval } from "eve/evals"
import { includes } from "eve/evals/expect"
export default defineEval({
description: "The Monday report uses the right source object.",
async test(t) {
await t.send("Give me the Monday leadership numbers")
t.succeeded()
t.calledTool("query_salesforce")
t.check(t.reply, includes("Pipeline"))
},
})
Then npm run eval. This is how you find out a prompt edit broke something,
before your team does.
agent/
├── agent.ts model, reasoning, limits
├── lib/config.ts ← start here
├── instructions/ who it is, how it behaves
├── skills/ ← your playbooks go here
├── tools/ one file per integration + always-on tools
├── lib/<service>/ API clients
├── lib/approval.ts which writes need a human
├── channels/slack.ts who may talk to it, who may approve
├── channels/eve.ts who may reach it from a browser
├── subagents/ specialists with their own tool surface
├── schedules/ recurring jobs
├── hooks/ lifecycle observers (usage analytics)
└── sandbox/ the workspace and its data tooling
evals/ checks you can run against your own fork
src/app/ site, /chat, /console, /analytics, /audience-vocab
npm run agent # terminal chat, hot reload on save
npm run agent:info # what did the framework find on disk?
npm run agent:doctor # do the credentials actually work?
npm test # pure-logic regression net
npm run typecheck
npm run eval # scored checks against a real agentTwo different questions, and you want both:
agent:info reports what the framework discovered and which environmentagent:doctor makes one read-only call per integration and reports whatApprovals are durable pauses, not stored records. A pending approval survives
a redeploy and never expires. Do not build a TTL or a reminder job for it.
Dynamic tool files must keep execute inline. In agent/tools/*.ts files
that use defineDynamic, write execute as an inline function. Assigningexecute: someFunction works on the first step but breaks when the runtime
replays it.
Subagents inherit nothing. A subagent under agent/subagents/ gets
only what is authored in its own directory — that is what makes the read-only
analyst genuinely read-only, and it is also why it declares its own sandbox.
The system prompt configures itself. Do not hand-write a list of integrations
into instructions/; 20-capabilities.ts already renders the live set.