Work in the browser

Open app.seifprotocol.com and create an account. A new account needs a password of at least 12 characters. The free workspace holds up to three projects, up to 10 MiB of authoring files and ten Context Builds per project. Files are limited to 1 MiB each; JSON project imports are limited to 10 MiB. The master workspace has no project or Build count limit. Accounts and projects are isolated from each other.

The browser now has the two workspaces shown in the desktop app: Agents for the guided product journey and Context engineering for precise control of context. They use the same agent factory and context engine as the Workstation. Studio projects actual agents, releases and observations into a live scene. The cloud keeps account/session data, agent records and Builds on its server; your portable authoring files can be exported at any time.

Create an agent

1. On the home screen, choose Customer support, Game character, Tutor, App assistant or Start from scratch. If you already have a project, open it and choose + New agent. Every starting point is editable.
2. In What should this agent do?, set its name, job and instructions. Create the agent. It starts as a draft; your users cannot call it yet.
3. Overview shows the same sequence as the desktop: define, connect, try, add knowledge or capabilities, validate, release, publish and integrate. A completed step refers to actual records, not just a visual checklist.
4. Behavior lets you edit identity, purpose, instructions, response type and advanced execution limits. Save changes before testing. A later change makes earlier test evidence stale.

The agent's ID is the stable identifier in the integration URL. Knowledge and capabilities have separate per-agent access; creating an item never grants it to unrelated agents.

Connect and check a model

Open Model → Connect a model. In the browser you can use labelled Rehearsal (deterministic responses, no AI call) or OpenAI with your own API key and model ID. Select the environment before saving. The API key is encrypted on the server, is never placed in a project export, and is not returned by the browser API. Use Check connection to verify a response; a configured connection alone does not prove it works. A real model may incur costs with your provider.

The local Workstation also supports installed subscription CLIs, Ollama and other compatible endpoints. Those are local connection choices, not cloud credentials. A rehearsal evaluation can be published in development or staging, but production requires a real model evaluation and live context binding.

Try, inspect and improve

Try it is a draft conversation. Nothing there is published to your application. Send a representative message and read What happened in this turn: response status, model or rehearsal, latency, knowledge delivered, capabilities offered and technical evidence. Start a new conversation after a major change to compare behavior. A draft preview is separate from application performance metrics.

In Knowledge, use Write information for an editable note or Add a document for a text file. The browser accepts bounded text, Markdown, JSON and CSV documents. For each knowledge layer, Use or Stop using changes this agent's access. See what the agent would receive resolves the context for a sample message without calling a model. Check the selected material and budget there before judging a model response. A folder outside the project cannot be connected from the shared cloud service; use the desktop for that.

In Capabilities, define a typed read or write contract, its purpose and input fields. The agent can only ask for a capability that its profile authorizes. Defining and authorizing a contract does not connect an executor. The browser currently displays that missing step; connect an external API or local executor in the desktop Workstation, where the endpoint and credentials can run in an isolated environment. A write request requires explicit review of its exact arguments before execution. The application still owns business authorization and idempotency.

Test, release and publish

The Tests screen contains behavioral scenarios. Add a realistic message plus phrases the answer must mention or must never mention, then choose Run tests. The result belongs to the exact saved draft, knowledge, model connection and environment. A rehearsal pass proves only that the deterministic rehearsal met the declared checks; use a real model to assess answer quality. If behavior, knowledge, capabilities or the connection changes, run tests again.

In Releases, choose Create release from this draft after a current passing evaluation. A release is immutable. Compare with draft shows changes in behavior, knowledge, capabilities, model and tests since that release, so you can review the next version before publishing it. Choose Publish / roll back for a verified release and select a traffic share. A partial rollout needs an existing baseline. New conversations follow the published route; conversations already in progress keep the release they started with. Production refuses rehearsal releases. The release screen shows history and whether each release is serving. You can export an individual release package from there.

Integrate shows a stable URL and a server-side curl example. Use Show application credential and store that credential in your backend, never in browser JavaScript. It is scoped to that project's published agent service: it can start published turns and report outcomes, but cannot edit, publish or approve actions. The agent-turn URL has this form:

POST https://app.seifprotocol.com/v1/projects/{projectId}/agents/{agentId}/environments/{environment}/turns
Authorization: Bearer {project application credential}
Content-Type: application/json

Send participantId, a unique requestId, message, and optionally sessionId to continue a conversation. The response includes the answering release and a turnId. Performance shows published conversations and reported outcomes. Report task success or failure through /v1/projects/{projectId}/turns/{turnId}/outcomes; a fluent answer alone does not prove the task succeeded.

Control the exact context

Switch to Context engineering for Design → Test → Build → Deploy → Observe. In Design, the map connects consumer, objective, profile, layers and sources. Select a layer to inspect it. The Perspective Mixer changes that profile's layer participation: Auto, Include or Exclude. Context opens the underlying authoring files; the file editor detects conflicting saves. Playground resolves a question and shows rendered messages, selection reasons and estimated tokens without a model call.

A new project starts as Markdown/YAML files (seif.yaml, agents/assistant/AGENT.md, agents/assistant/context.yaml, knowledge/welcome.md). When its Agent Workspace first opens, SEIF compiles that design to the portable context-project.json contract and retains a local backup of the original manifest. Edit the active context-project.json after this conversion. The export carries the active contract, knowledge files and agent-workspace.json drafts. Environment credentials, sessions and factory database files stay out of exports.

Add context scenarios in tests/*.ctx.md. For example:

---
id: refund
name: Refund question includes the policy
consumer: assistant
objective: answer
required: true
query: Can I request a refund?
must:
  - knowledge/welcome.md#refunds
---
# Refund context guard

Test checks context selection; it is different from the model-backed scenarios in Agents → Tests. Build validates and freezes a Context Build. In Design, Compare shows semantic changes between the working files and the most recent verified Build, including changed sources, layer rules and materials. Create a Build first to establish that baseline. In Observe, enter a question to see pieces in four rings: in hand, within reach, out of view and sealed. Select a piece to read why it landed there; this inspection does not call a model. Deploy offers the portable project handoff. A standalone runtime, registry promotion and investigation of standalone runtime traces still require the local Workstation.

In Studio, select an agent in the scene and send a draft message in the inspector. The reply and new interaction appear there without publishing a release. Select any node to inspect its recorded facts and relationships, or an interaction in the timeline to replay its request, release, context, model, tool activity and response frames. The investigation panel separates observed facts, inferences and unknowns; replay does not run the model or repeat external actions. Preview and published agent turns are included while their records are retained.

Use Propose change to prepare a behavior rule, written knowledge, test scenario or capability contract. The proposal shows effects, warnings and a behavior diff before you apply it. A scenario requires you to review its expectations, and a change affecting a shared audience requires explicit confirmation. Applying changes the draft and makes previous tests stale; it never silently republishes a live release. You can discard the proposal instead. The AI collaborator and live multi-user collaboration remain local workflows.

Export, import and local handoff

From Context engineering → Export project, download a ZIP of ordinary authoring files. Open the extracted folder in the local Workstation. The ZIP includes the agent-workspace.json draft definitions, so the local Agent Workspace can recover those agents. On the browser home screen, Import a project accepts that SEIF ZIP or a seif.cloud-project.v1 JSON export and creates an independent cloud copy. Individual immutable agent releases have their own export action in Releases. The project ZIP does not include published route state, credentials, sessions, operational traces, or Build directories. If you want to publish a copied project, check its model connection, run tests, create a new release and publish it deliberately.

Cloud and local folders do not continuously sync. Export/import is an explicit transfer of a revision. See open an exported project locally and agent workflow.

Browser and desktop availability

The shared cloud service refuses project sources outside the project and arbitrary remote connectors to protect other tenants. Team sharing, billing, email verification and account recovery are not available yet.