The workspace
Watch the workspace in motion
Goals in · finished work out
Send one goal.
Get finished work.
Mrrlin plans it, coordinates your agents, sends weak output back, and stops at every step that needs a human.
Your agents, under one harness
One coding agent gives you output. Mrrlin gives you a harness.
Codex, Claude Code, and other agents can write. Mrrlin keeps the brief, routes each task, challenges weak work, and brings you in only for decisions that need you.
Read the workflow transcript
Without a harness, a person carries context between prompts, checks each output, follows up, and decides what happens next. With Mrrlin, one goal becomes a visible plan: work is routed, reviewed, retried when weak, and brought back only for the decisions that need a human.
Inside the workflow
One goal. Every agent visible.
Plan, execution, review, and the approval inbox live in one place — not across five chat tabs.
Accepted — building the plan: 5 goals, 27 tasks, 2 approval checkpoints.
Open12
Needs you2
In review5
Ready8
Active goals
G-01Launch conversion-ready marketing site88%
G-02Founder-led SaaS outreach74%
G-03Turn feedback into a product spec81%
G-04Client campaign pack57%
G-05Technical project audit69%
Approval inbox
Needs you · 10:44
Marketing site is ready to publish
Copy, SEO, schema, and mobile checks passed. Two claims were rewritten after review.
Approve launchSend back
Review · 08:35
6 weak personalization lines rejected
Memory · yesterday
SEO opportunity map saved
To do7
SEO-02
Map support pages from ICP
ANL-07
Set up analytics events
Agent working4
SITE-01
Build landing page and demo section
SEO-05
Draft metadata for 6 pages
Being verified5
COPY-05
Check claims against evidence
LEAD-04 · returned
Reject weak ICP matches
Finished8
POS-01
Positioning → homepage messaging
SCH-03
Schema markup verified
3 agents runningClaude Code 42%Codex 31%other models 27%
Inbox · 4
Check-in · 08:35Outbound sequence is being reviewed
Done · yesterdayClient audit report assembled
Marketing site is ready to publish
The launch package passed copy, SEO, schema, and mobile checks. Two claims were rewritten after review. Nothing is published until you approve.
DeliverableReviewNext
Homepagepassedapprove
SEO metadatapassedpublish
CTA flowsign-offapprove
Launch checklistpassedship
Approve launchAnother review passSend back
Which competitor angle should we avoid?
Two positioning routes are plausible and the evidence is split. Mrrlin needs your judgment before the campaign brief is finalized.
Answer in your own words
Send answerLet the Director decide2 agents waiting
Project memory · 24 documents
DOC-214Marketing site launch package
DOC-218Founder-led SaaS ICP rules
DOC-221SEO opportunity map
DOC-225Client campaign brief
Reviewed launch package
What changed
Homepage messaging now matches the canonical positioning: coordinated agents, verification, retries, and human approval.
homepage.mdseo-checklist.mdlaunch-plan.md
Review notes
Reviewer agents tightened the claims, improved the CTA path, and confirmed that publishing waits for approval.
Next suggested goal
Create the first SEO support page and draft outreach to founder-led SaaS teams.
Model routing
Research & summarizingauto
Code & repo workpinned
Positioning & claimson
Long-form draftingauto
Always ask me before
Sending email to a real contactrequired
Publishing or deployingrequired
Client-facing claimsrequired
Internal drafts and researchnot required
Subscription usage
Codex CLI85% left
Claude Code62% left
1w window · resets in 7d · no metered API calls
Connected agents
Claude Codeonline
Codex CLIonline
Gemini reviewonline
DeepSeekidle
Up next4
1GT-1114Build a weekly acquisition scorecardSpec awaiting approvalgrowthautoFN
2GT-1808Define a 14-day measured pilotstarting soonplanningauto--
Blocked2
GT-1106Recover pillar rankings after checkpoint missWaiting on a handoffSEOauto--
Ask the Director@Mrrlin unblock the handoffs and continue
Explore the workspace
scroll to follow the loop →
Without Mrrlin
You
paste the context again
One model, one chat
one opinion, one pass
“Done.”
you are the only check
The chat ends and the context goes with it. Next time you paste it all again.
With Mrrlin
plan ▸
You
A goal
One sentence, in your own words.
Mrrlin · Director
Spec
Context is pulled from memory, not re-explained by you.
Gate · review panel
Consensus
Independent reviewers, rounds until they converge.
no consensus → rerun
Isolated
Run
Its own git worktree. Every artifact recorded as it goes.
Gate · proof required
Receipt
A live address, a green build, or your own sign-off.
no proof → stays open
Written back
Memory
What was decided, and why.
◂ remember
From goal to finished work
Three steps between
a goal and a
deliverable.
Everything that usually falls on you after the prompt — clarification, routing, review, retries, handoff. Three stages, with two proof gates before work moves on.
01
Shape the work
Plan readyOne plain-language business goal becomes scope, constraints, acceptance checks, and visible tasks before any agent starts.
02
Run and challenge it
Consensus QAEvery task routes to the right agent. Independent review checks claims, evidence, and code quality before delivery, returning weak work for another pass.
03
Approve consequential decisions
Your callRoutine work continues without asking. Only sensitive claims, sends, publishing, and final deliverables wait for your sign-off.
What lands in your inbox
Marketing launch pack
copy acceptedschema verifiedCTA needs sign-off
Founder outbound pack
12 rejected leads3 models reviewedsend list waiting
Repo audit report
risk rankedtests proposedclient report ready
Built for the work that keeps slipping
The work your team keeps postponing.
Founder-led SaaS
Launch a sharper marketing site, SEO base, analytics, and first growth campaign without becoming the AI project manager.
Deliverable
Published site, SEO checklist, launch plan, reviewed copy
8 tasks · founder sign-off
AI-native agency
Turn a client brief into campaign strategy, content, landing variants, and final QA across multiple agents.
Deliverable
Campaign brief, content pack, review notes, client-ready assets
7 tasks · multi-agent QA
Growth lead
Run acquisition and onboarding experiments end-to-end while keeping approvals and measurement visible.
Deliverable
Experiment plan, copy, tracking plan, results memo
6 tasks · metrics attached
Technical agency
Convert a client request or repo audit into scoped work, fixes, tests, and a polished recommendation report.
Deliverable
Delivery plan, implementation tasks, reviewed report
9 tasks · repo context
One harness for the agents you use
One Director routes work to the right agents.
Claude Code, Codex, Sonnet, DeepSeek, or your own local setup. Mrrlin plans, asks clarifying questions, sends important calls to consensus, then routes execution per task.
The state lives in Mrrlin and outlives every session. The model runs on your machine and is yours to swap. Nothing crosses between them without a receipt.
You work fromClaude CodeCodex CLICursorWeb appPhone
Mrrlin — the state that outlives every session
41 tables · the rules are database constraints, not team habits
Tasks and epics
Status, dependencies, and how much autonomy each one is allowed.
Project memory
Specs and evidence, ranked by how current they are — not just by keyword.
Decisions and why
The reasoning is stored with the decision, so the next session inherits it.
Receipts
The proof each finished task produced, attached to the task itself.
MCP · 70 tools · every read compacted, 585,000 → 26,000 characters
Your machine — your models, your keys, your subscription
Mrrlin never holds a model credential and never resells tokens
- 01
Intent
What you actually meant, pinned down before anything gets built.
- 02
Spec
Architect, security and scope reviewers — each on its own model — have to converge.
- 03
Run
An isolated git worktree, never your checkout. Artifacts saved as it goes.
- 04
Review
A separate read — never the executor grading its own work.
- 05
Receipt
A live address, a green build, or your own sign-off.
No receiptThe task stays open. A model saying it finished is not evidence, and the database will not accept it as one.
A resumed run ships a delta, not the whole history — measured on one real task: 641 KB → 25 KB of prompt.
One grant per destination · reversibility recorded · every call logged
The outside — nothing reaches it that you did not grant
Your repository
Branches, pull requests, workflows.
Your site or CMS
Pages that return a real address.
Messaging
Handoffs that actually arrive.
Ad accounts
Read and change, per client.
01
Plan with your context
Claude Code reads the repo and docs. The Director writes the plan, acceptance checks, and open questions before work starts.
02
Consensus finds gaps
Important decisions go through independent model review. Missing evidence and weak assumptions become specs.
03
Route by complexity
Simple work goes to cheaper models. Hard tasks route to Sonnet, DeepSeek, Codex, or specialized agents with visible checkpoints.
You keep the final call
Autonomous doesn't mean invisible.
Routine work continues without another prompt
Weak output is rejected and sent back for another pass
Evidence, reviewer notes, and model consensus stay attached
Publishing, sending, and client-facing claims wait for you
Goal
Launch founder-led SaaS outreach without sending risky claims.
6 weak lead matches rejectedre-run
2 email claims need stronger evidenceblocked
CTA variant B has stronger ICP fitreviewed
Question
Use the competitor comparison angle, or switch to a safer productivity angle?
Approve angleAsk for safer version
FAQ
Before you start.
What makes this different from a chatbot?
A chatbot responds to prompts. Mrrlin turns a business goal into an AI workflow, plans the tasks, runs the work, remembers context, and keeps iterating until the outcome is ready.
Which teams is it for?
Founder-led teams, early-stage startups, small product and growth teams, and agencies or consultants that need repeatable business workflows.
Do we need to change our tools?
No. The goal is to work across the tools you already use, not force a migration to a new system.
How much control do humans keep?
Human review stays in the loop for final delivery and sensitive decisions, while the system handles the repetitive execution work.
Give Mrrlin the first goal.
Start with the site, the SEO base, the outreach, or the client work you keep postponing. Mrrlin will map the execution path.
No credit card · No setup · One goal