SEIF One project, any runtime
A SEIF project is the source of truth for what a business's agents know (knowledge, profiles, tests) and for how they work together (the team: members, routines, flows). It is edited, tested and versioned in SEIF; each runtime receives an export, and every export says what it could not carry. Moving from one runtime to another, or using several at once, never means rebuilding the team.
runtime/engine/portable-team.mjs describes the team once (seif.portable-team.v1):
| SEIF | What it is |
|---|---|
| agents | the team's members, each an agent of the project with its own profile; one is the lead |
| routines | work that starts by itself: at a time on given days (at, days, in the team's time zone) or when a connected API reports an event (on) |
| flows | one member's accepted result (or each item of it) becomes the next member's work, with a brief |
and each adapter maps it:
| Trinity | n8n | OpenBot | |
|---|---|---|---|
| agent | one Trinity agent per member (trinityTeam): CLAUDE.md with its instructions and the team, the seif-context skill, the SEIF MCP server (hosted, or the engine carried with only the knowledge its profile reads) | an AI Agent node per member, with a model node and the SEIF context as an MCP Client Tool (n8nWorkflows) | an OpenBot agent per member in OpenBot's own import format (openbotImport, openbot-agent-import v1): its instructions, the seif-context skill, and under files/ (copied to imported/) the engine and only the knowledge its profile reads |
| routine on a schedule | a Trinity schedule (cron, IANA zone, enabled as in SEIF) | a Schedule Trigger | an OpenBot routine (custom cron, IANA zone, active as in SEIF) |
| routine on an API event | a Trinity schedule, disabled, marked [PENDENTE · configurar após importar] with the webhook step (Trinity has no event feed); SEIF keeps running it meanwhile | a Webhook trigger | an OpenBot routine, inactive, marked [PENDENTE] with the webhook step; SEIF keeps running it meanwhile |
| flow | the sender may call the receiver (permitted_agents) and its CLAUDE.md says what to hand on | the next AI Agent chained after the first | the team's channel instructions say who takes what |
| lead | may call every member | — | the channel's lead |
On the web: Contexto → Conectar a um agente. Trinity: Publicar no Trinity (one agent, or the whole team) with the Trinity address and an API key made there, used once and never kept; or download the ZIP. n8n: Baixar workflows (Workflows → Import from file). OpenBot: Baixar para importar no OpenBot (OpenBot → Server settings → Import → the file); the earlier seif.openbot-pack.v1 and the live bridge (Conexões → OpenBot) still work.
Nothing is dropped for needing setup
A routine, connection or capability that needs a step in the destination still travels, marked pending, with the step written next to it (runtime/engine/portability-report.mjs). Before downloading, Conectar a um agente shows per runtime what travels as is, what arrives pending and what stays only in SEIF. The OpenBot archive carries the checklist in the lead's files (agents/<lead>/files/SEIF-PENDENCIAS.md, copied to imported/); the Trinity publish result lists every pending item.
Importing a folder into SEIF works the same way. The import report lists what became knowledge, agents, materials (images, PDFs) and runtime extras (.claude/agents, .claude/commands), and every file skipped with the reason. A credential found in a file is replaced by a reference, never stored. Contexto → Configuração is the checklist after import (runtime/engine/setup-checklist.mjs): each connection's secret (stored in the vault from the item), a connection test, each agent's model connection, capabilities that wait on a connection, and proveniência left on for an external audience. Every item says whether it was checked or still waits, and why.
Verified (2026-10-29)
- Trinity v1.0 (local and on
seif-prod-armbehindtrinity.seifprotocol.com): the Missão CNH team published as five agents, eachverified: truewith 0 compatibility problems and started; the 7 scheduled routines became schedules inAmerica/Sao_Paulo(disabled, as in SEIF); the lead may call the four members and the prospector may call sales through the flow; the four event routines were reported as not carried. A Trinity that keeps a name reserved (a deleted agent) or versions a taken name makes SEIF undo the attempt and publish the whole team under one release suffix, so everyCLAUDE.mdand permission names the agents that exist. - n8n 2.45: the 11 workflows imported (
n8n import:workflow) and opened with every node recognised: schedule and webhook triggers, AI Agent 2.2, Anthropic model, MCP Client Tool 1.2 over streamable HTTP, and the prospecting → sales chain. Credentials (the model and the SEIF token) are picked in n8n; nothing secret travels in the files. - OpenBot: the archive follows the published
openbot-agent-importv1 rules (nightly-labs/openbot,resources/agent-import/*/SKILL.md), checked bytests/engine/portable-team.test.mjs. OpenBot imports it from its own screen (Server settings → Import); its automation API on 127.0.0.1 does not take imports (the team server that does needs a paired OpenBot client), so the import itself is a step in OpenBot, not a call from SEIF.