Model Context Protocol · production ready

Give your AI a workspace.
Keep the steering wheel.

Fliox turns client operations into explicit tools an agent can discover, understand, and use safely. It can search, analyze, draft, update, and coordinate — while scopes, approvals, tenancy, and audit stay under human control.

Connect an agent Explore capabilities

ScopedRevocableAuditedApproval-backed

fliox / agent-session

LIVE

# orient before acting

› connection.whoami

scope projects:read · projects:write
tenant Immigrate Technologies
policy human approval for bulk writes

› workspace.search "renewal cases at risk"

12 resultsprojects · tasks · documents

› projects.bulk_update_status.preview

PROPOSAL 8F213 projects → in progress

Human approval

› waiting for approval

01 / QUICKSTART

From zero to a useful agent
in six deliberate moves.

Each step narrows uncertainty before it increases authority.

Create an Agent Connection

Open Agent Access in Fliox. Name the connection for the client and purpose, choose the smallest useful scopes, restrict company access, and set an expiration.

Point your client at Fliox

Use the Streamable HTTP endpoint with the one-time Fliox bearer token. Claude, Codex, Cursor, and custom MCP clients can use the same server.

Let the agent inspect its boundaries

Start with connection.whoami, then resources/list. The agent learns its scopes, allowed companies, operating guide, safety rules, and current capability catalog.

Discover, then read

Use workspace.search to find typed results, then load the exact project, task, document context, template, or page before planning an action.

Preview risky changes

Bulk updates and deletion begin with a preview. Fliox freezes the target ids and diff in a proposal and gives the agent an approval link for the human.

Execute, verify, audit

After approval, the agent executes the current proposal version, re-reads affected objects, and reports the result. Reversible operations expose a compensating rollback.

CLIENT CONFIGURATION

One endpoint. One revocable token.

Use the same connection in any MCP-compatible client. Restrict it to the companies and scopes needed for that agent's job.

mcp.json

{
  "mcpServers": {
    "fliox": {
      "url": "https://fliox.io/v1/mcp",
      "headers": {
        "Authorization": "Bearer fliox_mcp_••••••"
      }
    }
  }
}

02 / CAPABILITY MAP

From “find it” to “fix it”
without leaving the system.

Deterministic tools use the external agent's intelligence and Fliox's operational guardrails.

Find the work

Orient the agent inside the exact personal or company workspace before it acts.

connection.whoami
companies.list
workspace.search
workspace.overview
company.overview

Understand operations

Read projects, tasks, readiness signals, and management metrics with links back to Fliox.

projects.list
projects.get
tasks.list
documents.context
company.analytics

Run client work

Create and update bounded objects with retry-safe, auditable writes.

clients.create
projects.create_from_template
projects.update_details
projects.steps.create
projects.steps.move
projects.reminders.create

Manage knowledge

Draft, validate, translate, publish, and connect reusable templates and pages.

templates.validate_draft
templates.create_draft
templates.translation_upsert
pages.create
template_pages.replace

Handle documents

Work with requirements and review metadata without placing private file bodies in MCP context.

projects.documents.list_requirements
documents.requirements.create
documents.files.create_upload
documents.files.review
smart_forms.generate_pdf

Change safely at scale

Freeze an exact target set and diff, ask a human, then execute or compensate.

projects.bulk_update_status.preview
projects.bulk_assign_responsible.preview
tasks.bulk_update_status.preview
changes.execute
changes.rollback

FOR AI CLIENTS

The documentation is part of the protocol.

Call resources/list, then read fliox://capabilities/catalog. The response contains every tool schema, required scope, side-effect flag, approval policy, and the recommended operating sequence.

initialize→resources/list→resources/read

03 / HUMAN CHECKPOINT

AI prepares the change.
A person releases it.

Bulk and destructive operations never jump directly from intent to mutation. The preview stores the exact targets, current versions, visible diff, and rollback plan in Fliox.

Frozen target setExecution uses reviewed IDs, never a fresh broad search.

Stale-state protectionScopes, roles, versions, and ownership are checked again.

Explicit compensationRollback is available only when a safe inverse exists.

AGENT

Preview

Build exact targets and a human-readable diff.

*.preview

PERSON

Review & approve

Open the deep link and release the proposal.

Agent Access

AGENT

Execute

Apply the approved version after all checks pass.

changes.execute

SYSTEM

Verify or compensate

Re-read state; roll back only when supported.

changes.rollback

04 / WORKED EXAMPLES

Small calls. Clear intent.
Inspectable outcomes.

READ / NO SIDE EFFECT

Find cases that need attention

Typed search returns match reasons and deep links without crossing the connection's workspace boundary.

{
  "jsonrpc": "2.0",
  "id": 12,
  "method": "tools/call",
  "params": {
    "name": "workspace.search",
    "arguments": {
      "company_id": "018f…",
      "query": "renewal",
      "types": ["project", "task", "document_requirement"],
      "limit": 25
    }
  }
}

READ / DEFINED METRICS

Build a management briefing

Project, task, document, and workload metrics include their definitions and calculation timestamp.

{
  "jsonrpc": "2.0",
  "id": 13,
  "method": "tools/call",
  "params": {
    "name": "company.analytics",
    "arguments": {
      "company_id": "018f…"
    }
  }
}

WRITE / HUMAN APPROVAL

Prepare a portfolio change

A stable retry key and exact project IDs produce a proposal. They do not update the projects yet.

{
  "jsonrpc": "2.0",
  "id": 14,
  "method": "tools/call",
  "params": {
    "name": "projects.bulk_update_status.preview",
    "arguments": {
      "idempotency_key": "renewals-2026-07-16-v1",
      "company_id": "018f…",
      "project_ids": ["0190…", "0191…", "0192…"],
      "status": "in_progress"
    }
  }
}

05 / SAFETY MODEL

Access is a chain of checks,
not a single secret.

Tenant isolation

Every company operation requires an allowed company and an active owner or assistant role. Cross-tenant IDs resolve as not found.

Least-privilege scopes

Read and write scopes are issued per capability family. A token can be company-limited, expiring, and revoked instantly.

Idempotency

Every mutation carries a stable key. Identical retries replay safely; a changed payload with the same key is rejected.

Human approval

Bulk and destructive changes stay frozen until a person approves the exact target set and diff inside Fliox.

Private document boundary

documents.context returns readiness metadata and risks — never file bodies, storage keys, or private download URLs.

Audit by default

Tool, connection, role, object references, outcome, duration, and safe summaries remain visible in Agent Access.

Give your AI a workspace.Keep the steering wheel.

From zero to a useful agentin six deliberate moves.

Create an Agent Connection

Point your client at Fliox

Let the agent inspect its boundaries

Discover, then read

Preview risky changes

Execute, verify, audit

One endpoint. One revocable token.

From “find it” to “fix it”without leaving the system.

Find the work

Understand operations

Run client work

Manage knowledge

Handle documents

Change safely at scale

The documentation is part of the protocol.

AI prepares the change.A person releases it.

Preview

Review & approve

Execute

Verify or compensate

Small calls. Clear intent.Inspectable outcomes.

Find cases that need attention

Build a management briefing

Prepare a portfolio change

Access is a chain of checks,not a single secret.

Tenant isolation

Least-privilege scopes

Idempotency

Human approval

Private document boundary

Audit by default

Give your AI a workspace.
Keep the steering wheel.

From zero to a useful agent
in six deliberate moves.

From “find it” to “fix it”
without leaving the system.

AI prepares the change.
A person releases it.

Small calls. Clear intent.
Inspectable outcomes.

Access is a chain of checks,
not a single secret.