AttexaEcosystem Reference
The Claude ecosystem · baseline orientation

What "we use Claude" actually means

A plain-English map of what Anthropic ships and how the pieces relate, so a client's "we use Claude" can be placed precisely. One model family underneath; many surfaces on top. Companion to the Attexa audit & governance matrix.

Verified November 13, 2026 6 reference labels · 1 model family Internal amotivv reference
01 · The fast orientation

Six common meanings of "we use Claude."

The differences that matter in an engagement are almost always about the surface (where the work happens, what it can touch, what gets logged), not the model.

When a client says…They likely meanWhere it runs
"We use Claude"Chat · the conversational assistantWeb, mobile, or Desktop; unified agentic conversations by plan, rollout, and Enterprise admin/role settings
"Claude is on our machines"Claude Desktop · the app shell; layout varies by rolloutmacOS / Windows / Linux (beta)
"Claude for our docs / ops work"Claude Cowork · autonomous knowledge-work agentDesktop, web, and mobile; separate Cowork or unified conversations where enabled
"Our devs use Claude Code"Claude Code · the coding agentTerminal, Desktop, IDEs, web, Slack, CI/CD
"We're building agents on Claude"Managed Agents / the Agent SDKManaged Agents: Anthropic-hosted or self-hosted service (private beta); Agent SDK: your runtime
"We hit the Claude API"Claude API · Console / Messages APIDirect Claude API; Anthropic-managed AWS or Google Cloud; partner-operated Bedrock or Google Cloud; Microsoft Foundry. Confirm the operator, not just the billing cloud.
02 · The mental model

One model family. Multiple execution paths.

There is one Claude model family underneath everything; the surfaces on top decide what it can reach and what is recorded. The current lineup is Fable 5.2 for demanding reasoning and long-horizon work, Opus 5.5 for long-running coding and knowledge work, Sonnet 5.5 for speed and intelligence, and Haiku 5.5 for high-volume, latency-sensitive tasks. Model access depends on the surface and plan.

Chatconversation
Desktopthe app shell
Coworkwork agent
Codecoding agent
Managed Agentsbuild your own
APIprogrammatic
One model family
Fable 5.2claude-fable-5-2demanding reasoning · long-running Opus 5.5claude-opus-5-5coding · knowledge work Sonnet 5.5claude-sonnet-5-5speed · intelligence Haiku 5.5claude-haiku-5-5fast · high-volume
Common confusion: "Claude" the model vs. "Claude" the app. A model upgrade does not establish new app controls. Fable 5.2 is generally available; its standard monitoring retention is 30 days. Eligible customers can receive a ZDR exception while transitioning to Enterprise Frontier Safeguards, now in private beta, with monitoring data stored in customer-controlled infrastructure. Confirm the agreement; customer custody is not independent proof of agent actions.
03 · Plans & pricing

Which plan a client is on changes what we can work with.

Several governance and audit features (the Compliance API, custom RBAC) require Enterprise (excluding Public Sector) or Claude Platform / API tiers. Most current pipeline runs on Team plans, always confirm the tier before assuming a feature is present.

PlanUS price (November 2026)Notes
Free$0Limited usage; no Cowork
Pro$20 / mo · $200 / yrIncludes Claude Code; Sonnet 5.5 + Opus 5.5; temporary included Fable allowance through November 16, then usage credits
Max$100 / moFormerly Max 5×; 5× Pro usage per five hours and per week; includes Code; claim $100 monthly API credits
Ultra$200 / moFormerly Max 20×; 10× Pro weekly usage; no five-hour limit for verified accounts, temporary limit otherwise; includes Code; claim $200 monthly API credits
Team Standard$25 / seat (mo)$20 / seat / mo billed annually · min. 2 seats · Code + Cowork · $20/seat monthly API credit
Team Premium$125 / seat (mo)$100 / seat / mo billed annually · 5× Standard usage · $100/seat monthly API credit
Enterprise$20 / seat / mo + usageAnnual seat billing plus usage at API rates; optional shared monthly group budgets alongside member limits. Self-serve: 20-seat minimum; sales-assisted: 50. SCIM, RBAC, audit logs, retention controls, Compliance API; security-center reports and centralized parent management in beta. Separate Code Review offer: eligible paid organizations can claim $10,000–$50,000 by January 15, 2027, for up to 30 days of eligible reviews. Exclusions apply; connected GitHub Enterprise Server and GitHub Enterprise Cloud with data residency (.ghe.com) are eligible. Not API credit; self-hosted GitHub Action and CLI reviews are excluded. Legacy contracts may differ.
Claude APIPay-as-you-goAPI billing through Console; eligible Max/Ultra/Team credits apply after linking an organization
Separate subscription limits from API credits: Max/Ultra/Team API credits require a linked Console organization and seven days on an eligible plan; availability is rolling out. Existing Max 20× subscribers move to Ultra automatically with no price change. Ultra verification is automatic and separate from phone verification; it removes the five-hour limit, not weekly or model-specific limits. Team pools are capped at $500/month, and unused credits expire each billing cycle. They cover API and Managed Agents usage, plus self-run Agent SDK/claude -p with that organization's API key, not interactive Code or extra app usage. The June SDK change remains paused: plan-signed-in SDK/claude -p still uses subscription limits. Fable normally uses paid usage credits on Pro/Team Standard. The November 9 launch offer includes all Fable models for up to 50% of those plans’ weekly limits through November 16, 2026 at 11:59 pm PT; usage credits, when enabled, apply above that allowance and after the offer ends. Max/Ultra/Team Premium retain the 50% weekly Fable cap. Eligible Pro/Team Standard accounts also receive one Fable reset expiring with the offer; eligible Max/Ultra/Team Premium and legacy Enterprise Premium seats receive one to three resets expiring March 8, 2027 at 11:59 pm PT. Resets require paid billing and, for Team/Enterprise, seat membership before November 9 at 10:00 am PT, with trial-specific exceptions; they restore usage rather than raise limits, and exclude usage-based, self-serve, legacy Standard, and HIPAA-ready Enterprise plans. These offers are not API credits. Consumer Usage Credits are distinct from the monthly API grant: purchases from December 7, 2026 will expire after 12 months; earlier purchases do not expire. Regional exceptions apply: Canada/Taiwan credits do not expire, and Japan purchases from September 10, 2026 expire after six months. Do not apply consumer-credit terms to Team or Enterprise contracts. Usage-based Enterprise seats include no token allowance and are not eligible for these API credits. Enterprise pooled group budgets are spending caps, not credits: a member stops at their own limit or when all applicable pools are exhausted. On Enterprise Standard/Premium seat plans, pools govern usage credits, not seat-included usage. Pools are configured in Organization settings, not yet through the Admin API. Cowork requires a paid plan. When enabled, Enterprise Efficient mode restricts unified work to Sonnet/Haiku and currently starts new sessions on Sonnet 5.5 at low effort. Confirm regional prices, legacy contracts, and current limits before quoting.
04 · Claude Desktop

One app. A changing split between chat and agent work.

Claude Desktop brings conversation, Cowork capabilities, and Code together on macOS and Windows; current Cowork guidance also lists Linux in beta. Unified Chat/Cowork is rolling out to Pro/Max/Ultra and is generally available for eligible Enterprise organizations. In existing Enterprise organizations it is off by default; unchanged settings are scheduled to enable from November 16, 2026 when Chat and cloud code execution are on. Explicit enable/disable choices are preserved. Team unified experience is now documented as available: new organizations start enabled; unchanged existing settings enable December 1, 2026 when cloud code execution is on, unless an owner keeps it off. Separate Team modes are scheduled to retire February 1, 2027. ZDR organizations are excluded; Team has no Efficient mode. Confirm owner settings before rollout. The Code tab remains. Sheets, Whiteboards, and Data are on by default for Team. Enterprise Sheets/Whiteboards are off by default; Data is scheduled to turn on by default November 19, 2026. Data’s certified queries and live sharing remain beta. Confirm rollout and artifact/role settings.

Chat

Conversation on web, mobile, and Desktop. Where unified Chat/Cowork is enabled, the same conversation can continue into cloud execution and connected tools. Paid plans add shared Sheets and Whiteboards (beta) and Claude Data (GA, formerly Dashboards), subject to admin settings.

Cowork

Agentic knowledge work, available as a separate mode or within the unified conversation. Cloud sessions follow the account across devices; local sessions remain device-bound.

Code

Interactive coding assistant with direct access to local files. Same engine as the CLI.

Ask about execution and enforced settings, not just tabs. Enterprise owners can now manage selected Desktop controls in Organization settings (beta, Desktop 4.29.0+), including workspace folders, local MCP, extensions, and allowed organizations. Matching device policy wins even when less strict. The account-based block on outside accounts starts after organization sign-in and an app restart; it is removable and expires after 30 days without use of that organization. Read-only folders constrain file tools, not Code commands, MCP servers, or hooks. Disabling user-added local MCP does not disable extension-provided servers. Confirm the actual execution path and effective policy.
05 · The agentic surfaces

Cowork, Code, and Managed Agents.

All three support agentic work, but their hosting and control boundaries differ. Cowork capabilities serve knowledge workers and are merging into Claude conversations; Code serves developers; Managed Agents supplies a configurable harness, hosted by Anthropic or customer-run in private beta.

Claude Cowork
Knowledge workers · ops, legal, finance
Agentic non-coding work without a terminal, available through Desktop and cloud sessions on web and mobile.
  • Run code in an isolated environment; reach connected local folders through the open Desktop app
  • Build Excel/PPTX/docs and collaborate in native Sheets/Whiteboards (beta) and Claude Data where enabled; synthesize research and extract data
  • Connect to internal systems via MCP; run scheduled tasks
Runs New Pro/Max/Ultra tasks: cloud; Team/Enterprise: local or cloud by policy; existing local tasks stay local
Plan Paid plans · Pro/Max/Ultra new tasks cloud-only; Enterprise unified experience GA with admin/role controls, unavailable with ZDR/HIPAA/CMEK; Team unified access follows owner settings; standalone Team cloud remains documented as beta
Check the execution and capture path. Enterprise has stable Cowork session-transcript export and beta unified-chat export, including cloud tool work. Use the matching session or chat endpoints; vendor-recorded content is not independent observation of every action.
Claude Code
Developers
One coding engine across terminal, Desktop, web, IDEs, Slack, and CI/CD, shared CLAUDE.md, settings, and MCP.
  • Understand whole codebases; coordinated multi-file edits
  • Git workflows and tests in Code; hosted Code Review (GA, Team/Enterprise) reviews GitHub PRs without running the code or deciding whether to merge
  • Automate with skills, hooks, mods, sub-agents and MCP. Experimental Code Mode batches tool calls through JavaScript. Session/Desktop schedules and cloud scheduled tasks (research preview) remain separate
Runs Terminal · Desktop · Web · IDEs · Slack · CI/CD
Plan Pro / Max / Ultra / Team / Enterprise, or API. Hosted Code Review: Team/Enterprise, owner-enabled and separately metered; unavailable with HIPAA, ZDR, or CMEK.
Ask where and which controls are active. Code Mode (experimental, off by default, v2.1.323+) is first-party interactive-terminal only, not SDK, print, Desktop, cloud, gateway, or third-party-provider use; HIPAA/ZDR configurations are excluded. It cannot start with OTel export enabled. Auto mode reviews the JavaScript, not each call’s computed values; individual deny/ask rules and PreToolUse hooks still apply. An admin can block new/resumed Code Mode sessions, not end one already running. Version 2.1.323 also refuses invalid SDK/print sandbox settings at startup and starts Manual mode when requested auto mode is unavailable. The working-directory restriction still covers file-write tools, not shell, MCP, or hook programs. The Agent SDK now offers onFailure: "block" for failed policy callbacks (TypeScript v0.3.317+, Python v0.2.177+, compatible Code required); defaults and event exceptions still matter. Server-managed policy supports Enterprise group configurations outside HIPAA: each member receives only the highest-ranked match, not a merge with other groups or organization-wide settings. Without a catch-all, unmatched local users receive no server-managed policy; endpoint-managed policy can still apply. Cloud/remote Cowork model limits use the catch-all or documented organization-wide fallback, not the member’s group list. Team remains organization-wide. Cached-policy verification and application of fetched policy without user approval remain separate controls. Code v2.1.319 adds permissions.denyInsteadOfPrompting: calls that would wait for permission are denied and the turn continues. On first-party/gateway paths with telemetry or feature flags off, auto mode now asks or denies calls the server cannot check instead of silently falling back to the built-in classifier; CLAUDE_CODE_AUTO_MODE_SERVER=0 retains that classifier, and Bedrock/Vertex/Foundry behavior is unchanged. These are client controls and automatic decisions, not human consent or action evidence. Mods still run outside the Bash sandbox. By default, the built-in guard loads with managed settings, a managed-mcp.json file, or active Team/Enterprise /login credentials; an API key or long-lived token alone does not load it. By default it preserves deny/ask rules and PreToolUse decisions against user-mod approvals. The managed allowModsToOverrideDenyRules option can relax more than deny rules, so verify the effective policy. The guard also skips selected user-mod prompt, settings, compaction, and transcript hooks. An organization-enabled marketplace plugin does not necessarily qualify as an organization-owned mod. Mods’ own file/process calls remain outside those rules. Code v2.1.321 also makes mod-initiated $.tool.call and $.mcp.call run without permission prompts, even under user ask rules or in dontAsk/headless runs; deny rules and organization policy still apply. Mods Claude writes can turn on without asking in auto/bypass modes, even with managed allowUnsandboxedCommands: false or allowManagedDomainsOnly, and do not load in dontAsk mode. Treat mod approval, mod-initiated calls, and mod installation as separate controls. Disabling the WebFetch tool does not block mod HTTP requests; organization-level availability policy is separate. Restrict user mods or use a policy hook with fail-closed error handling. Mod subprocesses retain the user's network access. Local-mode HIPAA configuration requires separate enablement. Its BAA coverage is limited to eligible terminal/Desktop use with Enterprise sign-in directly to Anthropic, not IDE extensions, cloud sessions, gateways, or plugins. Check the agreement and implementation guide.
Managed Agents
Developers building custom agents
A configurable harness, distinct from the Agent SDK. Anthropic hosts the service directly or through its AWS/Google Cloud offerings. New self-hosted Managed Agents (private beta) moves the service, loop, credentials, and session history to your infrastructure, not just its sandboxes. Model calls still go to your chosen provider.
  • Hosted: subagents and dynamic workflows (beta). Self-hosted service: predefined subagents only; no dynamic workflows or advisor
  • Hosted: cron deployments start sessions. Self-hosted service: bring your scheduler; no native scheduled deployments, dreams, or outbound webhooks
  • Vault-backed credentials; hosted MCP tunnels in research preview; tool policies always_allow, always_ask, or server-evaluated auto
  • Outcomes (public beta): define success criteria for iterative work; a completed workflow run is not proof the work passed
  • Hosted: events and best-effort OTel export (beta), with workflow code outside the event stream. Self-hosted: session history in your database, stderr logs, and opt-in OTel; you manage retention and integrity.
Runs Hosted service or self-hosted private beta · sandbox hosting is a separate choice
Plan Hosted API/cloud access: beta; tunnels and dreaming require research-preview access. Self-hosted service: request private-beta access from your account team. Anthropic-managed Google Cloud still excludes tunnels, pre-built self-hosted workers, and memory on self-hosted environments.
Hosting changes the control boundary. Self-hosted service requires a deployment per tenant and trust level, customer-secured workers, and an optional egress gateway for vault environment-variable credentials. It can use your secret manager and sign short-lived MCP identity tokens; those authenticate agents, not completed actions or per-call human consent. The server sends conversation content to the configured inference provider; provider agreements and covered-model exceptions govern retention. Self-hosting is not blanket ZDR or HIPAA coverage. Hosted direct-API IP allowlisting gates incoming requests, not runtime traffic; it excludes cloud-provider offerings. Hosted web controls, tool-domain lists, and limited networking remain distinct. Tool policies do not gate hosted workflow starts; turn interrupts and Google Cloud IAM revocation do not stop all existing work. Customer operation and logs are not independent, signed action evidence.
06 · Side by side

The agentic surfaces compared.

Claude CoworkClaude CodeManaged Agents
Who it's forKnowledge workersDevelopersDevs building agents
What it isFinished productFinished productBuilding-block platform · hosted or self-hosted private beta
Where it livesDesktop, web, mobile, and Chrome side panel; access varies by plan and admin settingsTerminal, Desktop, Web, IDEs, Slack, CI/CDAnthropic-hosted service, or customer-run service/loop/storage; self-hosted scheduling is external
Code executionNew Pro/Max/Ultra tasks run in Anthropic's cloud; Team/Enterprise can run locally or in cloud by policy. Local resources require Desktop open.Local (remote/cloud option)Hosted: managed or self-hosted sandbox. Self-hosted service: your workers; vault env-var isolation requires its optional egress gateway
Logging / auditTeam/Enterprise: opt-in OTel with device-policy and execution-path limits, plus a beta integration for a customer-chosen EDR sensor inside the local Cowork VM. No EDR installation in cloud sessions. Enterprise Compliance API separately exports stable session transcripts and beta unified-chat cloud tool work.Standard sessions: opt-in OTel and Enterprise local transcripts. Experimental Code Mode cannot run with OTel export; do not assume capture parity. Version 2.1.312+ has no safety_event. Cloud/scheduled, API-key, and third-party-cloud sessions remain excluded from Compliance capture.Hosted: events and beta OTel; custom/self-hosted tool spans need SDK tracing. Google Cloud API-activity logs have Data Access off by default. Self-hosted service: your event database, stderr log, and opt-in collector; not a signed execution ledger.
PlanPro / Max / Ultra / Team / Ent.Pro / Max / Ultra / Team / Ent. or APIHosted access · beta; self-hosted service · private beta by request
07 · Around the edges

The connectivity pieces.

Cowork on 3P

Routes inference through the customer's cloud (Vertex, Bedrock, Foundry, or a gateway); the agent still runs locally. Config moves to MDM."No data to Anthropic" applies here to the Google-managed/Vertex or Bedrock route, not Foundry. Do not extend that claim to the new Anthropic-managed Google Cloud API: Anthropic is its data processor.

Claude in Chrome

Browser extension for navigation and form actions. Cowork also has a built-in Desktop browser; admin settings govern the two separately.Cowork network-egress limits do not apply to web-fetch, web search, or Chrome.

Cloud work, Duet & Meetings

New Pro/Max/Ultra work is cloud-based; linked local resources still require Desktop. Duet remains a proactive cloud agent in US Pro/Max/Ultra waitlist research preview, sharing plan limits. Claude Meetings is a separate Team/Enterprise beta rolling out for Google Meet: a visible participant saves transcripts, summaries, and screen descriptions, not audio, and can build artifacts or use connected tools with permission. Team defaults on; Enterprise requires owner enablement. ZDR/HIPAA/CMEK are excluded. Connector use consumes the user’s allowance; meeting notes/transcripts and non-connector answers are included, with docs/whiteboards free through February 10, 2027. Confirm guest policy and participant consent. Neither Duet nor Meetings is assessed by the companion four-context matrix.

MCP (Model Context Protocol)

The open standard (Anthropic-created, now Linux Foundation) that connects Claude to tools and data. Supporting interactive connectors can now open full apps beside unified Claude on web/Desktop and let Claude make approved changes. Rollout and existing connector-tool policy govern access; developer support is public preview. Local MCP configuration is not automatically a cloud connector.

08 · Audit, compliance & the governance gap

Where native coverage ends.

The Compliance API · Enterprise + Platform

  • Activity Feed: identity, administrative and resource events, retained for six years; ordinary activity recording starts at Compliance API enablement. For enrolled organizations, Access Transparency events are recorded even while that API is off and can predate its first enablement. The separate audit-log CSV export covers the past 180 days.
  • On-demand content retrieval and explicit API deletion are available. Retention-policy deletion of chats, projects, and published artifacts emits no corresponding deletion activity; changes to the retention policy are logged separately. Missing deletion events do not prove content still exists. Enterprise beta also retrieves current members' saved Claude and Cowork instructions; this is configuration visibility, not proof of instructions used for each action.
  • Enterprise: the Primary Owner enables access through standard setup. For eligible parents enrolled in centralized management (beta), parent owners/admins manage the Compliance API switch and parent-level keys at admin.claude.com. Child organizations retain separate data boundaries; parent-level keys can reach across them. Eligible standalone Console organizations: admins enable Activity Feed access; Enterprise content endpoints are separate.

Cowork governance caveats

  • Enterprise exports cover stable Cowork session transcripts and beta unified chats with cloud tool calls/results. Customers can retain the exports; they are not signed agent-action records.
  • Enterprise cannot delete local sessions through the API. Unified conversations and their cloud sessions follow conversation retention; active-member account/project memory now follows the corresponding policy. Artifacts have separate retention rules. Standalone cloud sessions, schedules/runs, and local Cowork history remain outside custom retention; deleting a conversation does not stop its schedules.
  • OTel, Compliance API capture, and the new EDR integration are separate. Team/Enterprise can deploy an approved vendor's sensor inside the local Cowork VM on macOS or Windows with Desktop 4.4212.0 or later; the sensor independently reports VM activity to that vendor, not Anthropic. It can send content outside Cowork's network allowlist. Cloud sessions are excluded. Package-signature verification is not signing of the activity records; coverage, retention, and evidence integrity depend on the chosen sensor. Cowork OTel needs an admin-configured collector. Current Desktop-started sessions send metadata unless device policy enables content; older versions may send content with the policy unset. Web/mobile sessions, cloud schedules, and unified chats moved to cloud stay metadata-only. Tool-input strings can be truncated; metadata still includes connected-folder paths and, on first-party deployments, user email. Check content coverage and privacy separately. HIPAA-ready local Cowork/Code transcripts have 30-day or shorter retention; ZDR local sessions are excluded. Capture alone is not BAA coverage. The unified Enterprise experience still excludes HIPAA, ZDR, and CMEK; its Compliance chat coverage remains documented as beta.
The load-bearing gap

Anthropic provides native audit records, session exports, beta unified-chat exports, and Inference hooks for prompt and pre-tool-call gates. Enterprise security center (beta) adds configuration checks, Owner attestations, a downloadable PDF, and scoped Compliance API access. The Activity Feed records attestation changes, with documented omissions; it is not a complete history of every control-status change. The current-results endpoint omits some controls and attestations, and Passed can mean an accepted risk rather than a working control. These are posture records, not certification or signed evidence of executed actions. Enterprise IP allowlisting and Admin API member sign-out are also in beta: an IP block can leave chat work running for up to 24 hours; sign-out ends account sessions and the organization’s OAuth tokens, not API keys, Code cloud sessions/schedules, or linked Cowork computers. Members can sign in again. Cowork’s customer-chosen local-VM EDR remains a separate observer, with sensor-dependent evidence. Assess configuration, access, observation, and action proof separately.

Access Transparency supplies signed, append-only vendor-access/preservation records on eligible API workloads, not agent-tool execution receipts. Inference hooks send governed Enterprise transcripts and, with tool validation on, proposed tool calls to a customer security endpoint. That endpoint can retain frames and return allow/deny; deny can prevent execution. Configured HMAC authenticates delivery with a shared secret, not third-party-verifiable execution proof. Coverage depends on role scope, rollout sampling, and failure settings. In Only selected roles mode, only members set to Custom roles with a selected role are inspected; built-in roles, including owners, are not, even if assigned that role. No qualifying members means no member inspection. Platform/API organizations, including Managed Agents, remain outside hook scope. Assess enforcement, coverage, integrity, independence, and retention separately.

This is the gap the Attexa audit & governance matrix maps, surface by surface.
09 · Quick reference

Which surface for which task.

If the goal is to…The surface is
Have a conversation with ClaudeClaude conversation (web, mobile, or Desktop)
Do non-coding work: files, docs, researchCowork capabilities (Desktop, web, mobile, or unified Claude where rolled out)
Write/fix code, review PRs, automate dev tasksClaude Code (Terminal, Desktop, IDE, Web, Slack)
Build a custom agent on Claude's stackManaged Agents for a hosted or self-hosted private-beta harness; Agent SDK for a runtime you operate
Run inference inside your own cloudCowork on 3P; Code on Bedrock / Vertex / Foundry
Get programmatic audit access (claude.ai + API)Compliance API (Enterprise + Platform)
Connect Claude to internal tools/databasesMCP servers (across all surfaces)