Anthropic Claude API

Anthropic is the primary LLM provider for all 36 OpenClaw agents. All calls go via portkey — never direct. This hub documents the model lineup, dual Max plan architecture, billing topology, deprecation timeline, and the Hermes P2 demotion. Read it whenever choosing a model, auditing billing paths, planning OSIL model tiers, or rotating the Anthropic API key. For routing configuration, proxy fixes, virtual keys, and CHOKEPOINT-1 enforcement, start at portkey first.

Quick reference

Field	Value
Vendor	Anthropic
URL	https://console.anthropic.com
KB doc	API
Auth method	API key (x-api-key header) — but NEVER sent direct; always via Portkey
Auth credential	op://Aurora/anthropic/api-key
Cred-proxy port	pending B1-B6 ratification — see nemoclaw-audit-2026-05-03
Webhook port	n/a (all calls via Portkey; no inbound webhook)
Webhook handler	n/a
Webhook dedup table	n/a
Tunnel path	n/a
Direct API base	https://api.anthropic.com/v1/messages — FOR REFERENCE ONLY; OpenClaw never calls this directly
Proxy routing	:18900 primary (Portkey cloud) + :18903 Max shortcut — see portkey
Outbound API base	https://api.portkey.ai/v1 via portkey
Rate limits	Per Anthropic tier; surfaced in Portkey dashboard
Rate-limit action	429 → Portkey handles backoff; Discord ops alert
Cost	API-tier: pay-per-token; Max-plan: flat-rate monthly subscription
Backup/recovery	Anthropic-owned (vendor cloud); no local backup
Discord alert channel	ops
Drift cadence	Model deprecation tracking; tool-calls-health-check.timer 5-min
Status	production

Model lineup (authoritative — as of 2026-05-03)

⚠️ Model deprecation: old IDs sunset June 15 2026. Migrate before then.

Model	Canonical ID	Cost (in/out)	Tier	Agents
Claude Opus 4.7	claude-opus-4-7	$5/$25 per MTok	Strategic	aurora, solara (2 agents)
Claude Sonnet 4.6	claude-sonnet-4-6	$3/$15 per MTok	Operations	14 agents
Claude Haiku 4.5	claude-haiku-4-5-20251001	Lowest	Automation	18+ agents

Via Portkey, prefix with @anthropic-provider/ for direct-provider routing, e.g.:

• @anthropic-provider/claude-opus-4-7
• @anthropic-provider/claude-sonnet-4-6
• @anthropic-provider/claude-haiku-4-5-20251001

Deprecated IDs (migrate by June 15 2026):

• claude-opus-4-20250514 → use claude-opus-4-7
• claude-sonnet-4-5-20250514 → use claude-sonnet-4-6
• claude-haiku-4-20250514 → use claude-haiku-4-5-20251001

Model routing policy: see feedback_model_routing_pattern — Opus plans/audits (main conversation); Sonnet subagents execute; Haiku trivial lookups only.

Dual Max plan topology

Two billing paths exist for Anthropic LLM calls:

Path A — API tier (:18900 → Portkey cloud → Anthropic API)

• Pay-per-token; surfaces in Portkey dashboard + tool_calls Supabase table
• cost_usd, tokens_in, tokens_out fields populated in every tool_calls row
• Used by: all agents NOT routed via Path B

Path B — Max plan flat-rate (:18903 → anthropic-max-router → teamsteph@betterfiles.com OAuth)

• teamsteph@betterfiles.com Max subscription serves 40 agents
• cost_usd = NULL in tool_calls rows (flat-rate signal per CHOKEPOINT-1 spec)
• Routes via proxy.js → 127.0.0.1:18903; concurrency cap 5 (MAX_PLAN_LOCAL_CAP)
• henryRERI account Max plan remains on primary :18900 path for interactive VSCode use
• Added 2026-04-30; Phase 8 cutover to :18910 FAILED 2026-05-02 (format mismatch); :18903 remains canonical

Path A vs Path B — cost signal table

Field in tool_calls	Path A (API-tier)	Path B (Max-plan)
cost_usd	>0	NULL
tokens_in	>0	NULL or 0
tokens_out	>0	NULL or 0
cache_read_tokens	>0 if cache hit	>0 (from buffer parse)
cache_write_tokens	>0 on first write	>0 (from buffer parse)

For full proxy routing details, Fix #1 (Kimi 404 gating), and Fix #16 (cache_control on max-plan), see portkey.

Prompt caching (cache_control)

Anthropic supports prompt caching via cache_control: {type: "ephemeral"} on system prompts, large user message blocks, and tool definitions.

Fix #16 (2026-05-02): The old Tailscale max-proxy translated requests to OpenAI format, silently stripping cache_control for all max-plan agents. The new anthropic-max-router at :18903 uses the claude-code-20250219 beta header and accepts Anthropic-native format — cache_control is now injected by proxy.js on both :18900 and :18903 paths.

Effect: 35+ max-plan agents now get actual cache hits on their system prompts. Cache observability for the max-plan path is tapped via response buffer parsing (lines ~160-180 in proxy.js) and written to tool_calls.cache_read_tokens / cache_write_tokens.

See API for full prompt caching spec.

Hermes status — demoted to P2

Hermes (NousResearch-self-evolution) is NOT a dependency for OSIL. Demoted to P2 side-track per OSIL B5 ratification (2026-05-03).

OSIL Phase 1-14 proceeds with GEPA + DSPy + Reflexion + Voyager + autoresearch stack. Hermes is tracked as a future enhancement but does not gate any OSIL phase. Reference: openclaw-self-improvement-layer-2026-05-03 §B5.

NemoClaw cred-proxy (pending)

NemoClaw plan nemoclaw-audit-2026-05-03 includes B1-B6 blockers covering L7 credential injection that would change how the Anthropic API key reaches the proxy. Until B1-B6 are ratified, the credential path is:

op://Aurora/anthropic/api-key → master.env → proxy.js environment → x-api-key header (injected by Portkey)

Do NOT add any plaintext API key to hub files, config files, or scripts. op:// reference only per G-NO-PLAINTEXT-CREDS.

How it’s used

• Trigger: Any of the 36 agents issuing a structured-output or chat request; routed via portkey proxy.js
• Workflow (API-tier): Agent → OpenClaw Gateway :18789 → proxy.js :18900 → Portkey cloud → api.anthropic.com → response buffered and returned; tool_calls row written non-blocking
• Workflow (Max-plan): Agent → OpenClaw Gateway :18789 → proxy.js :18900 → isAnthropicModel() check → local shortcut to :18903 → anthropic-max-router → Anthropic OAuth → response; tool_calls row written with cost_usd=NULL
• Agents involved: All 36 agents; Strategic tier (Opus) = aurora, solara; Operations tier (Sonnet) = 14 agents; Automation tier (Haiku) = 18+ agents
• Failure mode — Anthropic API down: Portkey handles fallback to OpenRouter (Tier 2) or moonshot-kimi (Tier 2) per tier config; see portkey
• Failure mode — Max-plan pool saturated: 502 with x-openclaw-max-plan-fallback: pool_saturated; agent must retry; no automatic fallback to API-tier (different billing)
• Failure mode — model ID deprecated post-June 15 2026: 404/invalid model; update models.json for affected agents per feedback_models_guardian_pattern
• Success criteria: All requests resolved via Portkey within SLA; tool_calls rows present; cache hit rate visible in Portkey dashboard + tool_calls.cache_read_tokens

Cross-links

Agents that touch this

• _summary — Strategic Opus tier; primary Max-plan consumer (via :18903)
• _summary — Strategic Opus tier
• All 36 OpenClaw agents consume Anthropic models via this integration

Skills that invoke this

• aurora-model-swap — model ID rotation for Aurora agent
• claude-api — Anthropic SDK patterns, caching, tool use, model migration
• acquisitions-outreach — LLM call for SMS draft generation
• acquisitions-followup — LLM call for reply classification
• hubspot-deal-ingest — LLM enrichment calls

Plans that govern this

• openclaw-self-improvement-layer-2026-05-03 — OSIL: adds multi-tier DSPy/GEPA optimizers that call Claude; Hermes P2 demoted
• openclaw-fragmentation-fix-2026-05-01 — CHOKEPOINT-1 definition; tool_calls schema; proxy.js Fix #16
• nemoclaw-audit-2026-05-03 — NemoClaw cred-proxy pending B1-B6; would change credential injection path
• second-max-plan-decision — dual Max plan architecture; teamsteph@betterfiles.com topology
• vendor-deep-audit-comprehensive-2026-05-02 — Anthropic model cost baseline in audit

Feedback rules

• feedback_model_routing_pattern — Opus plans main; Sonnet executes via subagent; Haiku trivial only
• feedback_models_guardian_pattern — models.json overwritten by gateway on restart; edit backups dir
• feedback_chokepoint_principle — CHOKEPOINT-1: every LLM call writes tool_calls row
• feedback_dual_write_required — dual-write for all state mutation
• feedback_no_plaintext_creds — credentials as op:// references only; no raw keys in files
• project_second_max_plan_decision — teamsteph Max plan decision log

KB / source docs

• API — messages API, streaming, tool use, prompt caching, extended thinking
• MODEL-CATALOG — full model lineup with context windows and pricing
• MODEL-CATALOG — how Portkey routes to Anthropic models
• OBSERVABILITY — Portkey-side metrics for Anthropic calls

System maps

• vm-gateway-topology — full call path visualization
• vm-osil-overview — OSIL LLM tiers (DSPy, GEPA, Reflexion over Claude)

Related: LLM routing cluster

Cluster anchor: portkey

Hub	Role
portkey	All calls MUST route via this proxy; Fix #1 + Fix #16 live here
supabase	tool_calls writes; cost_usd / token accounting
openrouter (Tier 2 — hub pending)	Fallback if Anthropic unavailable
moonshot-kimi (Tier 2 — hub pending)	Non-Anthropic model alternative; gated by Fix #1

Related: Memory/embeddings cluster

Cluster anchor: supabase

Hub	Role
portkey	Voyage embeddings via voyage-embeddin-9dffd3 VK (616K req/mo)
supabase	44 SQLite DBs + vec0 vector search; chunk embeddings via Voyage
voyage (Tier 2 — hub pending)	Embedding model; Anthropic handles LLM; Voyage handles vectors

Related: Credential layer cluster

Cluster anchor: 1password

Hub	Role
1password	Stores op://Aurora/anthropic/api-key
portkey	Consumes Anthropic key via virtual key abstraction

Open issues / TODOs

• P0: Model deprecation deadline June 15 2026 — audit all models.json files for old IDs (claude-opus-4-20250514, claude-sonnet-4-5-20250514, claude-haiku-4-20250514); migrate to canonical IDs before cutoff
• P1: Phase 8 cutover blocked — :18910 binary format mismatch; see openclaw-fragmentation-fix-2026-05-01 §A2.9 F26; needs either proxy.js translation layer or different upstream binary
• P1: NemoClaw B1-B6 — pending Henry ratification; if ratified, cred-proxy changes how op://Aurora/anthropic/api-key reaches the proxy; update this hub when resolved
• P2: Hermes side-track — NousResearch-self-evolution demoted to P2; revisit in OSIL Wave 3 if GEPA/DSPy stack reaches token efficiency ceiling
• P2: Max-plan VK dead credentials — anthropic-max-p-99c12d VK in Portkey returns 401; archive or rotate

Related domain hubs

• ai-osil — AI/OSIL domain hub (this is the primary integration for OSIL)

Recent activity

• 2026-05-04: ai-osil domain hub cross-linked (ai-osil) as primary domain for OSIL AI self-improvement layer
• 2026-05-03: hub created (W1-S5); Hermes P2 demotion documented; model deprecation timeline added
• 2026-05-02: proxy.js Fix #16 (cache_control on max-plan path); Phase 8 cutover attempted + rolled back
• 2026-04-30: :18903 Max shortcut added for teamsteph@betterfiles.com (40 agents)
• 2026-05-01: Second Max plan live; dual billing topology established