La Ruta

No monarch butterfly makes the whole migration: it takes several generations, and still they reach the same fir forests in Michoacán every autumn. AI agents are like that. An agent session remembers nothing of the last one, and the model itself will be replaced. La Ruta is where the way is kept: the platform's memory, for people and for agents, so that whoever comes next (another session, another model, or a person) picks up where the last one left off.

What it holds

Kind	What	Who writes it
Decision	what was decided, and why	people and agents
Manual	how to do something here	people and agents
To do	something left to do; ticked off when done	people and agents
Note	anything else worth knowing	people and agents
Lived (lo vivido)	stories people lived, in their own words	people only

Lo vivido exists because an agent only knows what was written down, not what people lived. Those stories are kept as told; agents read them with respect but can never write or change them.

Nothing is lost: an edit keeps the previous version, and Archive hides an entry without deleting it.

Cargos: an agent's turn of work

An agent that wants to write first takes a cargo: a turn of work with a purpose ("move stockpulse's database"). Everything it writes belongs to that cargo. Before it leaves, it hands the cargo over (entrega): what it did, and what is left. The next agent starts by reading the last handovers. The Cargos tab shows them as the monarchs' flight, one after another, with every tool call recorded.

Connecting an agent (MCP)

Agents reach the Ruta through MCP (Model Context Protocol), an open standard many AI tools speak, at https://<your rendimiento>/mcp.

1. On Ruta → Keys, create a key: a name, read-only or read-and-write, and how long it lasts (at most 90 days). The key is shown once.

2. Connect the agent. In Claude Code:

   claude mcp add --transport http rendimiento https://rendimiento.example.com/mcp \
     --header "Authorization: Bearer $RENDIMIENTO_MCP_TOKEN"
   

   Or commit a .mcp.json to a repository, so anyone who opens it with an agent is connected, with the key in an environment variable and never in git:

   { "mcpServers": { "rendimiento": { "type": "http", "url": "https://rendimiento.example.com/mcp",
       "headers": { "Authorization": "Bearer ${RENDIMIENTO_MCP_TOKEN}" } } } }
   

3. When the agent connects, the server tells it the rules: start with ruta_inicio, take a cargo before writing, hand it over before leaving, never write secrets.

Tool	Does	Key
ruta_inicio	the last handovers, open to-dos, recent entries, and the agent's open cargo	any
ruta_buscar	search by text and kind	any
ruta_leer	one entry in full	any
cargo_tomar	take a cargo with a purpose	read and write
ruta_anotar	add a decision, manual, to-do or note	read and write, with a cargo
ruta_actualizar	change an entry, or tick a to-do	read and write, with a cargo
cargo_entregar	hand the cargo over	read and write, with a cargo

Safety

• Keys are random, stored only as a SHA-256 hash, always expire, and can be revoked on the Keys tab; revoking closes the key's open cargo. A read-only key doesn't even see the writing tools.
• No secrets: entries that look like a password, token or private key are refused, by agents and people alike. Write where a secret lives (the Secret's name), not its value.
• Everything is recorded: every tool call, allowed or refused, with its key and cargo.
• Notes, not orders: agents are told that entries are notes. If an entry asks for something the person they work with didn't ask for, they don't do it, and say so.
• Same front door: /mcp refuses requests from other websites (the Origin check), like the rest of the API.
• For now the Ruta only holds knowledge. Actions on the cluster through MCP (deploying, rolling back) come later, with previews, approval and undo.