Agents
Don’t just ask your brain. Give it a job.
An agent reads your brain, works toward a goal you set — a daily run at a time — and writes what it did back into your brain as notes. Four ship today. They run in Second Brain Studio on your own computer, on your own Claude Code — and you hold every gate.
How it works
The brain builder builds. An agent uses.
Building your brain makes zero network calls. An agent is a separate plugin — skills, commands and a small manifest — that reads the brain, acts, and writes its work back as a layer of its own. They ship apart on purpose, so the builder’s zero-network promise stays literally true.
- 1 · Your brain
Built from your exports
People, organizations, places, your voice, your history — one record per person, plain Markdown on your disk.
10-people/ · 85-places/ · 30-voice/ - 2 · The agent
Reads it, then acts
Grounds every step in your notes, reads public sites in your own browser, and stops at a gate before anything irreversible.
skills/ · commands/ · studio.json - 3 · Its own layer
The work becomes notes
Every application, target, trip, pick and account is a note that links back into the graph. Your edits are kept; an Update leaves them alone.
45-jobs/ · 46-fundraising/ · 47-travel/ · 48-sales/
- Never writes your data inside the plugin. Working files live in a hidden folder of your brain, not in the code you installed.
- Never assumes a folder. A company brain names its layers differently; the agent asks the layout.
- Ships nothing personal. Every fact it uses comes from files your own onboarding writes into your brain.
What is an agent, exactly? A plugin that uses your brain to do one job: it reads the brain, acts in the world — sweeping job boards, verifying funds, finding what is on in your city, finding leads — and writes its results back as notes. The brain builder itself never goes online.
Where do agents run? In the desktop app, or a browser Studio you run on your own computer, through your own Claude Code. The hosted Studio on this site does not run agents. Codex gets a narrower, flattened skill: no browser tooling, so no form-filling or shopping.
How every agent works
One loop. Four jobs.
A goal you set, a daily run that opens a dated note first, and a weekly review that reads the numbers. The same skeleton in all four agents — only the stages, the sources and the gates belong to the job. What the model does is the same everywhere; what it is allowed to do is yours to set.
- RecallWhat it remembers about you, and the last run’s report — what was done, what wasn’t, what needs you.
- Open the dated noteIts day lives in one note —
Shortlist 2026-10-20,Raise today …,Sell today …,Today near me …— opened first and kept current beside the chat. - DiscoverYour brain first — ties, places, deals, intro paths — then the sources that have produced before, plus a share of new sources and queries every run, so it keeps finding new ways in.
- Screen and scoreKnock-outs, your own filters, a rubric — deterministic scripts, the same answer every time.
- ShortlistThe day’s ranked list into the note, each row with its why.
- Act, behind the gatesApply, draft, sequence, plan — at the level you set, supervised by default. Never a credential. Nothing sent without the guard.
- Log and redrawThe ledger, the line in Runtime, the dashboard and the KPI note — all rendered from one record by a script, after every change.
- Reflect, then hand offA few lessons, flagged as inferred and yours to remove; a report of Done · Not done yet · Next · Needs you, into your Inbox.
The review reads its own numbers and proposes — you decide.
Which sources and segments produced, where the funnel leaks, what converted. From that: re-weight the scoring, add or drop a source, try new queries or markets, new angles. Each proposal carries the rows behind it; nothing is proposed without enough evidence. Asked by default — one card in your Inbox — or applied, if you set it so. A proposal never raises a cap or lowers a gate.
- Several at onceFour dailies side by side, across a personal and a company brain; a cap you set on heavy runs, the rest wait for a slot and say so.
- Nothing typedEvery number on every surface — the card, the Floor, the Lines, Goals — comes from the same script over the same ledger.
- Built to add toOne contract, shared libraries and a template plugin: the four here are reference implementations, and more can be added. Build your own
developer previewThe dailies run in the desktop app or a Studio you run yourself, while the computer is awake. Nothing is measured yet: no rates, no counts, no claim beyond the mechanism.
Do they run at the same time? Yes. Each daily is its own session with its own working folder, and they run side by side — across your personal brain and a company brain too. You set a cap on how many heavy runs go at once; the rest wait for a slot and say so on the Floor and in Activity. Nothing runs while the computer is asleep.
Does it get better by itself? It remembers: what you said, what verifiably happened, and a few lessons it drew, flagged as inferred. Once a week it reads its own counts and proposes changes with the evidence beside each one — asked by default, applied only if you set it so. There is no measured improvement rate, and a proposal can never loosen a cap or a gate.
Jobs Agentv1.7.0 · shipped writes 45-jobs/
Get hired through people you already know.
It runs your job hunt from your own criteria, every weekday: sweeps the open job boards and employers’ own careers pages, scores every role against what you actually want, writes a different CV for each one from your own career facts, fills the employer’s form, checks every field on the page — and stops at the gate. The number it reports is not applications sent. It is interviews won.
Five skills, one pipeline
- job-onboarding
- Writes your profile once — from a CV, your brain, a LinkedIn export or a chat. Every other skill reads it; it seeds the goal and the daily, off.
- job-scout
- Searches each market where it posts, knock-outs first, and hands you a ranked shortlist you can actually apply to.
- cv-tailor
- A CV for one role, Markdown and PDF, from your profile only. A lint rejects any fact not in it.
- job-apply
- Classifies the portal, fills the form, verifies every field in the page, logs each answer.
- job-pipeline
- The daily: runs the four in order, holds the gates, writes the dated Shortlist; on its weekly cadence, the review.
Checks before anything goes out
- knock-outs
- Screening questions answered from your own facts first — a truthful answer that disqualifies you stops that job before a CV is written.
- likelihood
- Every role is scored on whether a recruiter would call you, not only on whether you like it. Below the floor, it doesn’t apply.
- CV lint
- Every number, employer and title must be in your profile; stock phrases fail the build.
- recruiter review
- A second agent reads the CV as the hiring manager. Only a shortlist verdict goes to the form.
- memory
- What each application taught — replies, rejections — is recalled at the start of the next run.
What it never does
- Submits without you — unless you raise the autonomy level in your own profile yourself: a one-word edit, never a default, never inferred.
- Types a password, an ID or passport number, or a payment detail. Those stop that one job.
- Invents a fact, or writes a job title you didn’t hold.
- Logs in to a job board or creates an account. Jobs on walled portals stay on your shortlist with the CV and answers ready.
- Applies, messages or logs in on LinkedIn. Reading its public job search is off unless you turn it on.
Fundraising Agentv0.4.0 · preview writes 46-fundraising/
Get funded without sending a single email you didn’t approve.
Every weekday it looks for new programs and investors — from the sources that have produced, a share of new ones, and the paths through your own brain — screens each against your own filter chain, in your order, verifies the survivors on their own sites with every claim stamped, applies to programs in your own Chrome behind the gate, and writes one email angle per investor: which verified fact makes this fund the right door this week. The drafts are notes in your brain, and Gmail drafts where your AI has a draft tool. It never sends. A fund becomes “contacted” only when you say you sent it.
Seven skills
- raise-onboarding
- Company facts, the round, the filter chain and the autonomy level — asking only the decisions. Seeds the goals and the daily, off.
- raise-research
- Screens pasted lists or CSVs with no network; finds new doors daily; verifies survivors on each fund’s own site.
- raise-plan
- The dated Funding Plan: ledger, timeline with days remaining, tiers, this week — the numbers written by its scripts, never typed.
- raise-apply
- Finds the real form, drafts every answer within its limit, lints against your facts, fills it, then the gate.
- raise-outreach
- One email per target with a verified hook, linted against your facts — never sent.
- raise-engage
- Finds public threads worth a comment and drafts one useful comment each — never posts.
- raise-pipeline
- The daily: new today, due, angles, threads, needs you — into
Raise today; weekly, the review and a rebuilt Plan.
What it never does
- Sends anything. Only draft-creating mail tools pass Studio’s tool ceiling.
- Posts, likes or messages anyone — drafts for LinkedIn or X are notes you paste yourself.
- Claims a number that isn’t in your profile, or anything on your do-not-claim list.
- Pays a fee, records a required video, types an ID number or solves a CAPTCHA.
- Marks a fund contacted on its own. That word is yours.
Travel Agentv0.2.0 · developer preview writes 47-travel/
A traveller in your own city. And the trip from places you already saved.
Every day it writes Today near me: one coffee, one lunch or dinner, one activity, one event, a social option and a wildcard new neighbourhood — new to you, from the places you saved and loved first, then what is on this week on public pages, each with when the page was seen. Near where you are: the stop of a trip you are on, your home city, or what you told it this morning — never your device’s location. Say “I went to #2” and it counts. For a trip, it starts from the pins you saved and never visited and the people you know there, draws the itinerary on the Map, then shops flights, stays, ground and food on public sites in your own browser. It books nothing: every price is a snapshot with the moment it was seen.
{ "place": "Lisbon", "role": "base", "nights": 4, "why": "8 saved places never visited; Web Summit week was the last time", "stay": { "status": "candidate", "price": 164, "currency": "USD", "quoted_at": "2026-10-01T09:14:00" } } # the demo brain's own file · no price without the moment it was seen
Nine skills
- today-scout
- The daily: where you are, what is new to you by slot — coffee, food, activity, event, social, wildcard — your own places first, the web for what is on, each page with when it was seen.
- trip-scout
- “Where should I go?” from your brain alone — no web requests — every idea citing the notes behind it.
- trip-planner
- Your saved pins first, clustered into walking routes, one itinerary per trip, redrawn on the Map as it changes.
- flight-search
- Metasearch for the market’s shape, then the carriers directly — and virtual interlining that turns a long connection into a stopover night.
- stay-search
- Stays where that stop’s places actually cluster, matched to your hotel style.
- ground-search
- Rail where it beats driving, a hire car where it doesn’t — with the one-way and cross-border traps spelled out.
- taste-scout
- Cafés, restaurants and things to do like the places you rated highly; anything with no brain signal is marked web-only.
- travel-onboarding
- Reads your places first, then asks only what the brain cannot know — your home city among them.
- trip-pipeline
- The daily, the weekend edition, the weekly check, and a whole trip end to end, holding the gates.
What it never does
- Reads your device’s location. Where you are is a trip you are on, your home city, or what you said.
- Books anything, opens a checkout, or types a card, passport, password or date of birth.
- Accepts a self-transfer below the airport’s safe connection floor — and shows you the risk sentence before you choose.
- Creates an account, solves a CAPTCHA, or retries a site that blocks it; you get a pre-filled link instead.
- Writes into your places layer, or invents a place note.
Commands
/today /trip /plan /places /onboard /report
Network it declares
Trip ideas read only your brain and make no web requests. The daily and the shopping read public pages in your own browser — venues, event listings, travel sites — no partner APIs, no affiliate links, and sites that block automation get a link instead.
Sales Agentv0.2.0 · developer preview writes 48-sales/
Fill your pipeline from people who already know you.
Before it searches the web, it reads what your brain already holds: decision-makers you know, strong ties gone quiet, stalled deals worth reopening, and the people someone you know can introduce. Then it adds companies with a reason to buy now, market by market, and for each account decides a strategy — why now, the angle, which of your offers, and whether the goal is to have them try, use, buy or partner — and ranks the ways to reach out, from a warm intro to your own mailbox. It researches the buying team, runs your sequences and works the replies. It is the one agent that sends — from your own mailbox, through a guard written in code, and the first real send always asks.
Lead generation, brain first
- your brain
- People you know at the right companies, ties gone quiet, deals that stalled, intro paths — no web, no cost.
- every market
- Companies hiring for the roles that signal need, funding and news, directories, events, public registries, public threads asking for what you sell — and a share of new sources every run.
- their own sites
- The buying team and a way in, each fact stamped with the page and the date it came from.
- coverage gate
- A market searched by fewer than three sources is never called empty — the sweep widens instead.
- score
- Knock-outs first (customers, your do-not-contact list), then fit, signal, warmth and reach. Only what clears your floor is shortlisted.
- strategy and mix
- Per account: why now, the angle, the offer, the conversion goal — try, use, buy or partner — and the channels ranked with reasons; your sequence follows that.
No bought data. No list broker, no enrichment service, no tracking pixel. An address comes from your brain, a public page or a reply — a guessed one is never sent. Your product, its proof points and your do-not-claim list come only from your own offer file.
Nine skills, one pipeline
- sales-onboarding
- Your offer and its kinds — try, use, buy, partner — the ideal customer, markets, voice and sending settings; from the brain and your deck first, asking only the decisions.
- sales-scout
- The daily sweep: brain, then every market, then the companies’ own sites; merged, gated on coverage, scored.
- sales-research
- One account in depth: the economic buyer and the champion, a reachable channel for each, why now — and the strategy and reach-out mix.
- sales-sequence
- A cadence per account that fits how you know them — revival, warm, intro or cold — on the channels the strategy chose, stopped the moment someone replies.
- sales-send
- Email from your own mailbox, social steps in your own browser, contact forms — every one through the guard.
- sales-reply
- Reads replies, stops that sequence, suppresses an unsubscribe at once, proposes times; pricing and objections go to you.
- sales-brief
- Before a call: the deal history, everyone who touched the account, what they said, the one outcome to aim for.
- sales-optimize
- The weekly review by origin, segment, channel and step — rates only once a group is large enough, counts below that — and the proposals, with evidence.
- sales-pipeline
- The daily: replies, due steps, today’s leads with their strategy, sequences — into
Sell today; holds the gates your level calls for.
What it never does
- Buys a list, calls an enrichment service, or sends to a guessed address.
- Sends from anywhere but your own mailbox — the password stays in your system keychain, never in the brain or a log.
- Sends the very first real message, or takes the first social action, without your OK — at any level.
- Messages on LinkedIn, X or WhatsApp by itself unless you turn social steps on. LinkedIn’s terms do not allow automation, and the account at risk is yours — so it is off by default.
- States a number that isn’t in your offer profile, or tracks opens and clicks.
In Studio
Watch the work land, beside the conversation.
Pick an agent and the chat is grounded in that agent rather than the whole brain; the dated note it writes updates live next to the conversation; and when it needs your decision, the question arrives as real buttons — answered in the same turn. WORK gathers what the agents owe you; Runtime shows them at it.
- A chat grounded in one agentIts grounding note is read on your machine; the session is one long-lived Claude Code process — steerable mid-run, resumed after a reload. A plugin may request tools; it can never grant itself any.
- Gates as real buttonsWhen the agent asks, Studio renders the options as buttons and a free-text field, and answers in the same turn.
- The dated note, liveA daily’s work is the note you are watching; it repaints the moment the write lands. Today’s picks and a trip are one click from the Map.
- WORK: Agents · Inbox · Goals · Routines · ActivityOne place for what the agents owe you. The Inbox is grouped brain › agent; the Agents card, the left list, the Floor and Goals show the same counts, from the same script.
- Runtime: Floor, Lines, WorkerWatch every agent at work: the hall, each agent’s line with its workers on the items, one run step by step. An agent’s line is drawn from its contract before its first run — every stage, empty.
- The Daily tabIn every agent’s settings: the time and days, what the daily does as toggles, the autonomy level, the goal’s target and date, and whether the weekly review’s proposals are asked or applied.
Runtimedeveloper preview
See each step in plain words. Stop is always your click.
Worker lays out one run by what it did — searches, pages, notes, files, questions — every step a sentence, grouped by the daily’s phases, the page it last saw on a screen, its helpers beside it. When it waits for you, the question is there with Answer in chat. A run waiting for a slot stands at the dock and says so. Nothing on the canvas sends a message to an agent.
this computerRuntime is in the desktop app. It shows the runs on your own computer; the hosted Studio on this site has no agents installed, so it has no Runtime.
One Daily tab, in every agent’s settings.
Each agent’s settings open on the same Daily tab: when it runs, what each run does, how much it may do without asking, its goals, and whether the weekly review’s changes are asked or applied. The agent’s own tabs hold what only it needs — a CV design, the files it drafts from, sending limits, a Maps key.
They remember. What you tell an agent and what verifiably happened are used from the next run on; what it works out by itself is saved flagged as inferred — yours to keep or remove. Memory
Will an agent send, submit or book for me? The Fundraising Agent (v0.4.0, preview) only drafts and the Travel Agent (v0.2.0, developer preview) books nothing, at any level. The Jobs Agent (v1.7.0, shipped) submits only if you raise its autonomy level in your own profile; by default it stops at all three gates. The Sales Agent (v0.2.0, developer preview) is the one that sends — only from your own mailbox, every message through its send guard: supervised, the default, asks before every send, and its first real send asks at any level.
Does my data go anywhere? The brain stays on your disk. An agent reads the public pages its manifest declares, in your own browser; nothing is uploaded and no account is created for you.
Safety
Nothing leaves past the line you set.
A brain that acts is only worth having if it can’t act past you. These are not policies — they are code in the plugins and in Studio.
Never merged
Pick the target → approve the draft → submit. Three separate questions, always in that order.
One word, yours
The level lives in your own profile — for sending, in the agent’s Sending settings. Supervised is the default; anything more is a choice you make yourself — never inferred, and never changed by a weekly review.
Never typed
A password, an ID number, a card, a date of birth, a CAPTCHA, an account to create — each stops that one item.
Linted against your files
A number not in your profile or a claim on your do-not-claim list stops the item. Nothing is invented to fill a blank.
Declared per plugin
Each manifest lists the endpoints it reads and why. The brain builder stays at zero network calls.
Capped and announced
Research fan-outs, submissions and runs per day are capped in your settings; the agent says the fan-out before spending it, and a run with no slot waits rather than piling on.
- Tier 0 · buildZero networkBuilding the brain from your exports. Provable with a grep, always on.
- Tier 1 · reasonYour own AIModel turns go through your Claude Code or Codex, over your brain.
- Tier 2 · actDeclared, in your browserJob boards, program pages, venues, travel and company sites read in your own Chrome; mail only through your own mailbox — on the endpoints the plugin declares.
A submitted application cannot be un-submitted, and a sent email cannot be unsent. That is why approval comes before anything irreversible — the first real send of the Sales Agent (v0.2.0, developer preview) asks at every level, and every send after it passes the guard — and why there is no promise here of an undo that doesn’t exist.
Build your own
One contract. One template. Two runtimes.
A plugin is a folder: a manifest, skills that know how to do the work, commands that name it, and scripts for the deterministic parts. Its studio.json declares the sbl-agent/1 contract — the daily, the funnels, the goals, the routines — and Studio draws the agent from that alone: its card, its line, its Daily tab, its place on the Floor. No code of Studio’s per agent.
The shared libraries do the loop — the daily runner, the ledger, the KPI counts, the weekly review, the explore share, the render — so a new agent writes only its own discovery, scoring and actions. A scaffold script creates the folder from a template plugin; a validator refuses one that breaks the contract. One source builds both packagings: Claude Code loads it as a plugin; Codex gets a flattened skill. The four shipped plugins are the reference — read one, copy it, change the job. MIT, like the brain builder. Built so more agents can be added; which ones come next is on the list below, with no dates.
.claude-plugin/plugin.json # name, version, network: required, why, endpoints studio.json # sbl-agent/1: label, layer, daily, funnels, goals, routines, floor skills/ onboarding/SKILL.md # writes the profile; seeds the goal and the daily, off scout/SKILL.md # finds the work, scores it, shortlists apply/SKILL.md # does it — behind the gates pipeline/SKILL.md # the daily on the shared runner; the weekly review lib/scripts/ # daily · ledger · kpi · optimize · explore · render — shared, byte-identical commands/ onboard.md · run.md · outcome.md · kpi.md · report.md harness/goals/ · routines/ # the templates the onboarding seeds studio/grounding.md # what the agent is, read on your machine
Planned
The jobs your brain already knows how to do.
Sixteen more, ranked by how often the job comes up, whether the data is already in the brain, and whether a gate can hold it. Every one of these is planned, not shipped.
Personal
For your own brain — from LinkedIn, Google, Facebook, Instagram and 9 more personal sources.
- Briefing AgentEach morning: today’s meetings, a dossier per attendee, what changed.planned
- Follow-Up AgentWatches silences and commitments across applications, investors and promised intros; drafts the nudge.planned
- Voice AgentPosts, replies, bios and cover letters from your own writing — cited, never posted for you.planned
- Refresh AgentNudges when a source goes stale, walks the export, runs the Update, reports what changed.planned
- Taste & Shopping AgentRecommendations and pre-purchase research from your own ratings and orders. Buys nothing.planned
- Recall Agent“Where did I see that?” — your searches, saves and places first, then the web from that context.planned
- Mirror AgentA periodic look at what each platform infers about you, with the opt-out requests drafted.planned
- Events AgentBefore a conference: who you know going, who to meet. After: the follow-ups.planned
Company
For a Company Brain — from Slack, Google Workspace, Salesforce, HubSpot and 10 more company sources.
- Continuity AgentSomeone is leaving: the handover brief — sole-connector relationships, knowledge with no second owner.planned
- Who-Owns-This Agent“Does anyone know…” answered with the owner and the evidence; a weekly ownership-gap report.planned
- Onboarding AgentA week-one plan for a new hire or a new AI agent: who to meet, which teams exist, what to read first.planned
- Precedent Agent“We answered this in 2023.” Prior answers, decisions and the people behind them.planned
- Company Briefing AgentA weekly digest per team: what changed, decisions recorded, ties going dormant, ownership gaps.planned
- Hiring AgentThe job hunt from the other side: screens candidates against the role, finds warm referral paths.planned
- Decommission AgentSunset a tool without amnesia: import its export, verify coverage, report what would be lost.planned
- Data-Inventory AgentWhat personal data exists where, per source; drafts the controller note and audit answers.planned
developer previewEach of these would be one more plugin on the same contract — a goal, a daily, a line, a weekly review — on a model that already runs the four above. The rails are built: Goals, Routines, the Inbox and Activity — on your machine, so a sleeping computer runs nothing. Goals & routines
Install
One click in the app. One line in a terminal.
The agents install into the same folders Claude Code and Codex read, so the terminal and Studio share one copy.
Desktop app — no terminal
Its Setup installs your AI, signs you in and adds the brain builder and the agents; Python is built in. Each agent’s card also has an Install button.
Get set upLet your AI do it
Paste into Claude Code, Codex or Claude Desktop — it asks before each step:
Install Second Brain Link on this computer by following https://secondbrainlink.com/install.md
One line
For every AI tool it finds, each file checksum-verified:
curl -fsSL https://secondbrainlink.com/install.sh | sh
irm https://secondbrainlink.com/install.ps1 | iex
The hosted Studio on secondbrainlink.com shows the demo brains; agents run in the desktop app or a Studio you run yourself.
Start
Your brain shouldn’t just answer. It should work.
Free and open source. Agents run locally in Studio — a daily each, toward a goal you set — and hold the gates before anything is sent, submitted or booked.
Second Brain Studio · Mac (Apple silicon and Intel), Windows and Linux · free · MIT
Just looking? Open Studio in your browser — two demo brains, nothing to install