Cevran — Features

01 · PLAIN LANGUAGE

Type the sentence.
I'll do the parsing.

No prompt engineering. No slash commands to memorize. Ask the way you'd ask a teammate who happens to know your stack — and I'll figure out which tools to touch.

·Multi-step intents in one sentence ("pull these, draft those, ping #revenue").

·I ask before guessing. If two interpretations are plausible, I'll surface both.

·Reference past runs by ID, name, or "the one from Tuesday."

"Pull the open deals from Salesforce idle 7+ days, draft a check-in email to each owner, and post the list in #revenue."

QUERY salesforce.opportunities · idle ≥ 7d · stage != closed →

DRAFT gmail.compose · per_owner · check-in · review_first →

POST slack.message · #revenue · summary_format →

GitHub CONNECTED · 1-CLICK

Slack CONNECTED · 1-CLICK

Salesforce CONNECTED · 1-CLICK

Notion READY · 1-CLICK

Linear READY · 1-CLICK

Stripe API KEY · FALLBACK

OAUTH · DEFAULT· API KEY · FALLBACK· WORKSPACE-SCOPED

02 · CONNECTIONS

One click whenever
I can. Keys only when
I can't.

I default to OAuth so setup is fast and you stay in control. The connector layer handles tokens, scopes, and refresh — you never paste a key into chat.

·3,200+ apps. Browse the directory.

·Manual keys live in a vault, scoped per-workspace, rotated quarterly.

·Custom connector requests turn around in ~2 business days.

03 · EXECUTION

Multi-step work,
end to end.

Cevran chains tool calls, branches on results, retries on transient failures, and writes everything down as it goes. You can step away — the log is the same whether you watched or didn't.

·Branching: if the API returns empty, switch tools or ask.

·Retries with exponential backoff. Transient ≠ failure.

·Stream-friendly: read the log as it writes, or jump to the end.

14:08:12 ● plan drafted · 7 steps 0.4s

14:08:13 ● salesforce.query · 12 rows 0.8s

14:08:14 ● gmail.compose · drafted 12 5.1s

14:08:19 ◐ slack.post · preparing list writing…

— ◇ self-review queued

— ◇ open drafts for approval queued

— ◇ report.send queued

I drafted 12 Gmail follow-ups for owners of idle deals. Same skeleton, customized per deal stage. Nothing's sent yet.

DRAFT · 1 OF 12

To: jamal.k@acme.com
Subject: Quick check-in on the Q3 renewal

Hey Jamal — circling back on the renewal. Last we spoke you were looping in procurement; happy to send a fresh quote or jump on 15 minutes if helpful…

THRESHOLD: REVIEW-FIRST FOR ANY EMAIL · TUNED PER TOOL

04 · REVIEW-FIRST

I pause before
I push send.

Irreversible actions — send, pay, delete, deploy, escalate — sit in an approval queue until you tap through. You choose the threshold per tool: review every time, or just for amounts over $X.

·Per-tool thresholds. Stripe refunds > $100, all Slack DMs, every Gmail send.

·Bulk approve, review-each, or step back into the chat to retune.

·Every approval is logged with reviewer, timestamp, and the diff that was approved.

05 · SELF-VERIFICATION

It checks its own
work — first.

Every plan ships with checks. Cevran runs them against the result before handing back. When something feels off — even when nothing technically broke — it flags the moment instead of shipping it.

·Static checks: counts match, schemas align, links resolve.

·Semantic checks: tone, fit, contradictions with prior runs.

·Failed checks aren't silent — they appear in the report.

all 12 owners reachable OK

no duplicate threads in last 14d OK

drafts conform to brand voice OK

2 deals had calls already booked FLAGGED

Acme and Globex have demos booked this week — sending a check-in would read cold. Pulling them from the batch unless you'd like otherwise.

Ten outreach emails are out, two are paused for context, and #revenue has the list.

Pulled 12 idle opportunities (≥ 7 days no touch). Drafted personalized check-ins for each owner; sent 10 after your approval; held 2 because the deals have demos this week. Posted the full list in #revenue with links back to each draft.

STEPS

7

TOOLS

3

SENT

10

FLAGGED

2

06 · REPORT & AUDIT

A clean write-up.
Every time.

Every run closes with a plain-language summary plus the receipts: which tools were called, with what arguments, what came back, what was skipped and why. Export the audit log to your warehouse on the Team plan.

·Plain summary first. Receipts below. Same in chat, email, or PDF.

·Audit export: JSONL or Parquet, dropped to S3 or BigQuery.

·Replay any run with the same plan, different inputs — or fork it.

07 · NATIVE WORKSPACE

Where the work lands.

Most agents hand back a wall of text and leave you to file it. Cevran ships with a calendar, notes, and tasks of its own — surfaces it reads and writes directly, with no OAuth dialog and no export step. The output of a run has a home, and the agent keeps it current.

·Calendar — unified view of local events, Outlook sync, and your scheduled runs. The agent checks free/busy and books.

·Notes — markdown, searchable, labelled. The agent captures and appends so nothing is lost between runs.

·Tasks — Kanban + list, drag-to-reorder, cross-linked to runs and projects. The agent creates, moves, and closes.

TO DO · 2

Refund duplicate charge

FROM RUN 0x9B2A

Reply to vendor thread

IN PROGRESS · 1

Draft Q2 changelog

CEVRAN · WRITING

DONE · 2

Post deal list to #revenue

Book follow-up demos

Cevran moved “Draft Q2 changelog” to In Progress and linked the run.

EVERYTHING IN ONE TABLE

The whole picture.

What ships today, what's in beta, what's on the near roadmap. No dark patterns — if it's not here, I don't do it yet.

Capability	Drafts	Operator	Team
Plain-language task understanding	●	●	●
1-click OAuth integrations	3	UNLIMITED	UNLIMITED
Manual API key fallback	—	●	●
Multi-step execution	●	●	●
Self-verification on every run	●	●	●
Review-first approval queue	●	●	●
Per-tool review thresholds	—	●	●
Run log + report	●	●	●
Built-in Calendar (local + Outlook sync)	●	●	●
Built-in Notes (markdown, searchable)	●	●	●
Built-in Tasks (Kanban + list)	●	●	●
Agent reads & writes your workspace	●	●	●
Audit export (JSONL / Parquet)	—	—	●
Shared workspaces, roles	—	—	●
SSO · SCIM	—	—	●
BYOK (bring-your-own-key)	—	—	●
Dedicated execution capacity	—	—	●
Named onboarding engineer	—	—	●

Less a chatbot,
more a coworker.

Type the sentence.
I'll do the parsing.

One click whenever
I can. Keys only when
I can't.

Multi-step work,
end to end.

I pause before
I push send.

It checks its own
work — first.

A clean write-up.
Every time.

Where the work lands.

The whole picture.

Hand it off. See what happens.

Less a chatbot,more a coworker.

Type the sentence.I'll do the parsing.

One click wheneverI can. Keys only whenI can't.

Multi-step work,end to end.

I pause beforeI push send.

It checks its ownwork — first.

A clean write-up.Every time.

Where the work lands.

The whole picture.

Hand it off. See what happens.

Less a chatbot,
more a coworker.

Type the sentence.
I'll do the parsing.

One click whenever
I can. Keys only when
I can't.

Multi-step work,
end to end.

I pause before
I push send.

It checks its own
work — first.

A clean write-up.
Every time.