AEOS
--:--:--

Claude

ConnectedInstalled

AI Providers · Anthropic · v1.0.0 · manifest schema 1.0

The default model connector. Every AI employee in the fleet ultimately resolves its reasoning through a provider like this one, and the Decision Engine's advisor seam is written against the same `completion.create` capability. It ships installed because an enterprise of AI agents with no model provider is not an enterprise of AI agents.

Last sync
4h ago
Health
84/100
Rate ceiling
1000/min
Webhooks
not supported

Add to the estate and begin the authentication handshake. Every operation is a state transition inside AI-EOS — no external service is contacted.

Provider Runtime

3/3 served

Implements the three capabilities the Claude manifest declares. Operations are described rather than performed — the request shape, the returned figures and the topics raised are the ones a real implementation has, so becoming real is a change of body rather than of contract.

v1.0.0
Runtime
plugin version
1.0
Contract
provider schema
3
Operations
3 declared
2
Topics raised
onto the event bus
Capability coverage100% of the manifest
Operations
  • Create completioncompletion.create

    Runs one inference on behalf of an agent or an advisor, inside the configured token ceiling.

    Reads: defaultModelmaxTokensRaises: runtime.completion
  • List modelsmodel.list

    Enumerates the models this provider can serve and their context windows.

    Reads: defaultModelRaises: nothing
  • Create embeddingsembedding.create

    Vectorises memory records so the retrieval layer can rank them.

    Reads: no configurationRaises: runtime.embedding
Readiness checks
  • Default modeladvisory

    claude-opus-5 is in this runtime's catalogue.

  • Token ceilingadvisory

    Completions are capped at 8,192 tokens, inside the 64,000 this runtime accepts.

  • Credentialadvisory

    The Auth Manager reports a usable credential reference — operations may be dispatched.

A blocking failure stops the runtime dispatching and costs health points. The plugin states the fact; the Health Monitor prices it, alongside every other penalty.

Configuration in effect
defaultModel
claude-opus-5
Provider default — the operator has set no value.
maxTokens
8192
Provider default — the operator has set no value.

Operator values over provider defaults. A blank field takes the default, which is declared by the plugin and printed here rather than hidden inside an operation.

Health

84/100

84/100 — 1052ms mean round trip, above the 800ms threshold. 1 further penalty applied.

  • −8Latency1052ms mean round trip, above the 800ms threshold.
  • −8Rate ceilingPeak consumption reached 95% of the 1000/min ceiling.

Authentication

Authenticated
Method
API Key
Credential
sk_····3bba
Last verified
1d ago
Principal
—
Expires
Does not expire
Scopes held
2/2
Scopes
  • messages:write
  • models:read
Handshake · API Key
  1. 1.InitiateThe Hub prepares a API Key request for Anthropic.
  2. 2.Supply api keyEntered once. Held by reference — the Hub stores no credential material.
  3. 3.VerifyThe Hub performs one authenticated read to confirm the grant, then records the reference.

Nothing is contacted. The Hub stores no credential material — only a masked reference, the granted scopes and an expiry. There is no field on the auth record a secret could go in, which is a stronger guarantee than a policy of not putting one there.

Capabilities3

  • Create completionscompletion.create

    Run an inference on behalf of an agent or an advisor.

  • List modelsmodel.list

    Enumerate available models and their context windows.

  • Create embeddingsembedding.create

    Vectorise a memory record for retrieval.

Modules request capabilities, never connectors. A capability offered by two connectors gives the requesting module a fallback it does not have to know about.

Usage · 24h

273,600
Calls
1,094,400
Records
1052ms
Mean latency
0.2%
Error rate
1000/min
Rate ceiling
declared by the manifest
95%
Peak usage
saturates at 80%
Peak consumption against ceiling

Event Subscriptions2

  • Critical Recommendationdecision.critical

    The advisory layer raised something at critical priority.

    Raised by decision-engine · see the fan-out chain

  • Memory Writtenmemory.written

    A record entered the retention layer and may need filing externally.

    Raised by memory · see the fan-out chain

Publisher

First party
Name
AI-EOS
Publisher ID
aeos
Platform vendor
Anthropic
Support
operations@ai-eos.internal

Published by the platform itself. These manifests ship with the build rather than being fetched.

Anthropic owns the platform this plugin reaches. AI-EOS authored the manifest.

Version and changelog

Up to date
Installed
1.0.0
Published
1.0.0
Minimum platform
1.0.0
Compatibility
Compatible
Manifest schema
1.0
First published
10 months ago

Requires platform 1.0.0 or newer. This build is 1.0.0.

  • 1.0.0major · 5 weeks ago
    • First stable release, and the first connector with a provider runtime behind its declaration.
    • completion.create, embedding.create and model.list are dispatchable rather than advertised.
    • The model catalogue lives in the provider, never in the Hub's registry.

Permissions and scopes

operator+

AI-EOS permissions are what this plugin may do insidethe platform, resolved against the API Gateway’s own table. Scopes are what the external platform is asked to allow. They are different systems and neither substitutes for the other.

  • memory:read4 roles

    Memory records across every scope the caller may see.

  • analytics:read4 roles

    Enterprise Intelligence read models and the Decision Engine's outputs.

  • agent:execute3 roles

    Run a registered agent through the AI Runtime. Required by runtime.execute — the reservation this permission was written for, now taken up.

Requested external scopes
  • messages:write
  • models:read

Dependencies0

Declares no dependencies. It can be installed against an empty estate.

Whether each is satisfied is resolved against the live estate at install time, not here.

Security

Authentication
API Key
Credential storage
Never stored
External scopes
2
AI-EOS permissions
3
Minimum role
operator
Inbound deliveries
Not accepted
Tenant isolation
Installation state is tenant-scoped

A plugin is a data literal. Nothing in the marketplace executes plugin code, and a plugin cannot bypass the Authentication Manager, the Integration Hub, the API Gateway or a permission check, because it has no code path — the host performs every transition.

The credential is referenced by a masked hint (sk_····3bba) and nothing else. The auth state type has no field a credential fits in, so the guarantee is enforced by the type checker rather than by convention.

Privacy

Data leaving AI-EOS
2 capabilities
Data entering AI-EOS
1 capabilities
Modules served
4
Event topics received
2

Direction is declared per capability, so what this plugin can send and what it can receive are both enumerable rather than a matter of trust. It reaches only the modules listed above, and only through capabilities the Integration Hub resolves.

Documentation2

Logs3

  • Sync39m agoSynchronisation completed — 1,094,400 records reconciled in 24h.
  • Gateway23m ago273,600 gateway calls in 24h at 1052ms mean latency.
  • Registry5d agoConnector v1.0.0 installed against manifest schema 1.0.

Roadmap

  • Per-agent token accounting feeding the Agent Fleet's cost column
  • Streaming responses once the AI Runtime owns request lifecycle
  • Model routing by task class rather than a single default

What this connector would gain next. Mirrors the Future Integrations list every module carries in the Navigator registry.