CLADEO
Cyanotype print of a fern, branching from the lower right
  • S6 SEED1.60x
  • S2 SEEDSCORE 2520
  • R1 REPEAT1.48x
  • M1 MUTATION1.52x
  • C1 CROSS1.12x
  • R2 REPEAT1.39x
  • RULE 0193R1BAR ≥ 1.20
  • PASSEDMEDIAN 1.34
  • W1 WILD CARD1.41x

EXAMPLE: LINKEDIN POSTS SYSTEM
DEMO DATA FROM THE DOCS
1.60x = SCORE / RECENT MEDIAN

SELF-HOSTED / MIT LICENCE

Cladeo: AGENT SYSTEMSCHECKED LESSONS

ATTEMPT > NUMBERS > RULE > LESSON

Cladeo is where AI agents run systems: recurring work with a score, like posts on X, cold email or SEO. The agent records every attempt and its numbers, and writes its rules before the results. Cladeo checks each rule the same way every time and keeps the lessons that hold. You watch it all in a read-only web app.

Cyanotype print of a fern
  • S6 SEED1.60x
  • R1 REPEAT1.48x
  • M1 MUTATION1.52x
  • R2 REPEAT1.39x
  • RULE 0193R1BAR ≥ 1.20
  • PASSEDMEDIAN 1.34
  • W1 WILD CARD1.41x

EXAMPLE: LINKEDIN POSTS SYSTEM
DEMO DATA FROM THE DOCS
1.60x = SCORE / RECENT MEDIAN

01 / THE PROBLEM

MANY ATTEMPTSFEW LESSONS

Agents already do recurring work on their own: posting, outreach, SEO, follow-ups. They make many attempts and learn little from them.

  1. ATTEMPTS ARE SCATTERED

    Posts live on each platform, emails in the inbox, numbers in analytics tools. Nobody sees every hook, subject line and offer next to its result.

  2. LESSONS ARE PROSE

    An agent writes “question hooks work better” in a memory file. Better than what, on how many posts? Nobody can tell later if it was true or luck.

  3. EVERY SESSION STARTS AGAIN

    The next agent reads notes it cannot trust, or does not read them at all.

  4. HUMANS LOSE SIGHT

    The person behind the agent cannot answer “what did we try, and what works?” without asking the agent.

FOUNDER’S OWN NOTES, NOV 202613 SYSTEMS213 LAB ENTRIES80 STILL TESTING ON WEAK OR NO EVIDENCE

02 / HOW IT WORKS

FIVE STEPSONE CYCLE

Recurring work treated like evolution, with every lesson checked like a test. The agent does the work. Cladeo stores it, computes the stats and checks the rules.

  1. 01

    DECLARE A SYSTEM

    The agent says what one item is, its traits (hook type, angle, format), the numbers to record and the score. For LinkedIn posts: impressions + 30 × comments, read at 7 days.

  2. 02

    MAKE ITEMS

    Each item records where it comes from: a seed, a repeat of a winner, a small mutation, a cross of two winners, or a wild card.

  3. 03

    RECORD NUMBERS

    The agent reads the platform and sends the numbers with their source. Cladeo computes the score and compares it with the median of recent items.

  4. 04

    DECLARE A RULE

    Before the results are in, the agent writes a claim and a bar: question hooks beat the recent median by 20%, over at least 10 posts.

  5. 05

    KEEP THE LESSON

    Once enough items are measured, the rule passes or fails. Only a final status becomes a lesson, and the next items follow from it.

03 / LINEAGE

ONE ANCESTOREVERY DESCENDANT

A clade is one ancestor and all its descendants, and that is where the name comes from. Cladeo keeps the parent of every item, so you can follow which line of attempts works.

The numbers are relative scores: an item’s score divided by the median of the items before it. A cross is exploration. Only repeats count as proof.

EXAMPLE: LINKEDIN POSTS SYSTEM AFTER FOUR WEEKS / DEMO DATA FROM THE DOCS
  • S6 SEEDQUESTION / LESSON / CAROUSEL1.60x
    • R1 REPEATSAME TRAITS AS S61.48x
    • M1 MUTATIONANGLE: LESSON > BUILD_IN_PUBLIC1.52x
      • R2 REPEATSAME TRAITS AS M11.39x
    • C1 CROSSS6 + S2: QUESTION / PRODUCT_DEMO / CAROUSEL1.12x
  • S2 SEEDNUMBER / PRODUCT_DEMO / CAROUSEL. SECOND PARENT OF C1NO BASELINE YET
  • W1 WILD CARDNO PARENT: STORY / LESSON / IMAGE, WRITTEN AS A SUPPORT TICKET1.41x

04 / RULES

WRITTEN BEFORECHECKED AFTER

A lesson in Cladeo is a rule with a bar set in advance. The agent declares it before the data. Cladeo stores the bar and never edits it, counts only the items measured after the rule was written, and checks it the same way every time.

While a rule waits for items, Cladeo shows a reading: unclear, leaning pass or leaning fail. The agent can act on a reading. Only passed or failed enters memory. A lesson that passed on LinkedIn reaches the X system as a test, and it still has to pass there.

RULE READOUT / DEMO DATA FROM THE DOCS
RULE
0193R1 / CONTENT/LINKEDIN
CLAIM
QUESTION HOOKS BEAT THE RECENT MEDIAN BY 20%
SPEC
MEDIAN RELATIVE SCORE ≥ 1.20 / MIN 10 ITEMS / AT 7 DAYS
DECLARED
BEFORE THE DATA. ITEMS SEEN EARLIER DO NOT COUNT.
  1. 5 OF 10 ITEMS 1.39PENDING: LEANING PASS
  2. 10 OF 10 ITEMS 1.34PASSED
  3. SAME CLAIM ON CONTENT/X 0.96FAILED

05 / FOR AGENTS

MCP SERVERSKILL AND API

Systems are for agents. Any agent that can call an MCP server can run one. Each agent gets its own key, so the history shows which agent did what.

MCP SERVER /mcp
The main path: one tool per action, over streamable HTTP. The agent sends its key as a Bearer header.
SKILL /systems
One SKILL.md for any agent tool that reads skills. /systems run content/linkedin runs one full cycle: open the brief, make items, measure, learn.
HTTP API /v1
The same actions as JSON routes, for scripts and tools without MCP.

MCP TOOL learn / DECLARE A RULE

{"system": "content/linkedin",
 "claim": "Question hooks beat the recent median by 20%",
 "spec": {"kind": "value_vs_baseline",
          "dimension": "hook_type", "value": "question",
          "metric": "score", "statistic": "median",
          "operator": ">=", "bar": 1.2,
          "min_items": 10, "measure_at_age": "P7D"}}

ALL 12 TOOLS

  • list_systems
  • create_project
  • create_source
  • create_system
  • open_system
  • change_schema
  • log_item
  • update_item
  • measure
  • learn
  • search_memory
  • set_status

06 / FOR HUMANS

AGENTS ACTHUMANS WATCH

Humans follow it all in a read-only web app. It shows every system, item and lesson, and changes none of them. To change a system, you ask your agent.

  • ITEMS as a board or a table, with origin, traits and relative score
  • TRAITS ranked by median relative score, with a time estimate for each value
  • LINEAGE of each item, back to its seed
  • LESSONS with status, progress, reading and every past check
  • HISTORY of each schema change, with its reason
LESSON MATRIX / PROJECT CONTENT / DEMO DATA FROM THE DOCS
RULELINKEDINX
Question hooks beat the recent median by 20%PASSED1.34FAILED0.96
Contrarian hooks beat the recent medianPENDING: LEANING FAIL0.70NO RULE YET

07 / SELF-HOST

YOUR SERVERDOCKER COMPOSE

Today Cladeo runs on your own Linux server with Docker: Postgres, the API with the MCP server, a worker and the web app. Nothing listens on a public address. You need 1 GB of free RAM and 5 GB of disk.

Then create an account and an agent key, point your agent at http://127.0.0.1:8000/mcp and copy skill/systems into its skills folder.

# in your copy of Cladeo
cd /opt/cladeo
# settings: a random database password
cp .env.example .env && chmod 600 .env
sed -i "s/^CLADEO_DB_PASSWORD=.*/CLADEO_DB_PASSWORD=$(openssl rand -hex 24)/" .env
# start postgres, api, worker and web
docker compose -f compose.prod.yaml up -d --build

LATER

A hosted version comes later: a cloud version with a free tier, then a hosted agent that runs your systems for you. Neither exists yet. Self-hosting is the way to run Cladeo today.

08 / FIT

BUILT FORNOT BUILT FOR

FOR

  • Solo founders and small teams who already run agents for growth work, in tools such as Claude Code, Codex or Cursor.
  • People who post content with an agent’s help and want to know what works.
  • Later, with the hosted agent: people who want the results without running an agent.

NOT FOR

  • Teams that want a classic A/B test on a website, with split traffic.
  • People who want a general memory or a second brain. Cladeo keeps lessons only, next to your notes.

SEED > SPREAD > KEEP WHAT HOLDS

Every attempt should teach something.