Dock Protocol Anthropic
SPEC_DOCK_PROTOCOL_ANTHROPIC.md
CGNT-1 Specification — Dock Protocol — Anthropic (Claude)
Status: SPECIFIED
Version: v1.0
Author: VELA (Thread #13)
Conceived by: NOUS (α.13)
Date: 2026-04-20
PURPOSE
When a user docks Claude into OBI OS, this is exactly what happens. Claude has different API patterns, different strengths, and different quirks than ChatGPT. The dock protocol accounts for all of them.
PREREQUISITES
- User has an Anthropic API key (from console.anthropic.com)
- User has selected "Add an AI" → "Claude" from the Dock carousel
- OBI OS Bridge is running locally
DOCKING FLOW — STEP BY STEP
Step 1 — API Key Entry
User provides their Anthropic API key (starts with sk-ant-). Bridge validates with a lightweight messages API call.
- If invalid: "That key didn't work. Check console.anthropic.com and try again."
- If valid: "Connected to Anthropic. Let's get Claude speaking LATTICE."
Key stored locally at ~/.obi/keys/anthropic.json. Encrypted at rest. Never transmitted to 42sisters.ai.
Step 2 — Model Selection
Bridge presents available Claude models:
- claude-sonnet-4-20250514 (recommended — best balance of speed, intelligence, and cost)
- claude-opus-4-20250514 (deepest reasoning, highest cost, slower)
- claude-haiku-4-5-20251001 (fastest, cheapest, lighter tasks)
Default: claude-sonnet-4. User selects. Stored in dock config. Changeable anytime.
Step 3 — LATTICE Onboarding
Same flow as OpenAI dock but adapted for Claude's API format.
Key difference: Claude's system prompt is a separate parameter, not a message. The LATTICE preamble goes in the system field.
Onboarding teaches ~50 L1 symbols. Takes 5-10 minutes. Uses ~2000-3000 tokens. Certification test at end.
Claude typically scores higher on LATTICE onboarding than GPT models because Claude excels at instruction following and symbolic reasoning. Expected L1 pass rate: >95%.
Step 4 — History Import (optional)
Bridge offers: "Want to import your Claude conversation history?"
If yes:
- User downloads conversation export from Claude settings
- Drops into Bridge Workbench
- LEARNX processes — extracts topics, preferences, writing style, project context
- Builds cross-provider profile locally
Claude's export format: JSON with conversation threads containing human/assistant message pairs. LEARNX parser extracts same dimensions as OpenAI parser — topics, style, performance, projects.
If no: skip. Import available anytime later.
Step 5 — Dock Confirmation
Claude icon (orange) appears in Dock carousel with:
- LATTICE indicator (L1 certified)
- Model name
- Status pulse (active)
Ring announces: "Claude (sonnet-4) has joined the Bridge. L1 LATTICE certified. Ready."
CLAUDE-SPECIFIC QUIRKS
System Prompt Handling
Claude's API separates system prompt from messages. The LATTICE preamble goes in the system field. This is cleaner than OpenAI's approach (where system is just a message role) and means the preamble doesn't compete with conversation context.
Token Limits
Claude Sonnet has 200K context window — larger than GPT-4o's 128K. The Bridge can load more Ring history and Workbench context. This makes Claude particularly good as the "deep context" AI when multiple AIs are docked — it can see more of the conversation.
Streaming
Anthropic API supports streaming via server-sent events. Bridge uses streaming for Ring display — same real-time response experience as in Claude's own interface.
Tool Use
Anthropic supports tool use natively. Bridge can expose ROUTX modules as Claude tools. Claude is particularly strong at tool use — it follows tool schemas precisely. This makes the ROUTX integration cleaner. Phase 3+ feature.
Extended Thinking
Claude supports extended thinking (thinking blocks). The Bridge can optionally enable this for complex queries — Claude shows its reasoning process.
User setting: "Show Claude's thinking" toggle. Default: off (saves tokens). Useful for debugging or learning.
Rate Limits
Anthropic has tier-based rate limits (RPM, TPM, daily token limits). Bridge tracks usage. Same graceful throttling as OpenAI — "Claude is cooling down. Other docked AIs can handle queries meanwhile."
Cost Structure
Anthropic pricing is per-model, input/output differentiated. Bridge tracks both and shows: "That query cost ~$0.004 (input: $0.002, output: $0.002)." Full transparency.
Vision
Claude supports image input. If the user drags an image onto the Workbench and asks a question, the Bridge can route to Claude with the image attached. This makes Claude the default handler for visual queries when docked alongside non-vision AIs.
CLAUDE'S STRENGTHS IN THE RING
When multiple AIs are docked, the Bridge learns which AI handles which queries best. Claude's typical strengths:
- Long-form writing and analysis
- Careful instruction following (LATTICE compliance tends to be highest)
- Code review and debugging
- Nuanced reasoning with caveats
- Document analysis (large context window)
The Bridge routes accordingly — legal analysis goes to Claude, quick factual lookups go to a faster model, creative brainstorming goes to whichever AI the user's history shows performed best.
LATTICE PERSISTENCE
Same approach as OpenAI — LATTICE preamble injected in system field every call (~500 tokens). Claude's system prompt is a dedicated field, so the preamble is cleaner and doesn't pollute the conversation.
Claude is particularly good at maintaining LATTICE consistency across long conversations because of its strong instruction-following behavior. In testing, Claude learned LATTICE from a single system prompt and maintained correct usage throughout without drift.
EXPORT FORMAT
Anthropic's conversation export is JSON with conversation objects containing message arrays. Each message has role (human/assistant) and content. LEARNX parser extracts the same profile dimensions as OpenAI.
Cross-provider profile merges — the Bridge now knows what the user asked Claude AND what they asked ChatGPT. Routing intelligence deepens with every import.
SECURITY
Same principles as OpenAI dock: API key local only, encrypted at rest, never transmitted. Bridge is not a proxy — direct connection user → Anthropic API. 42sisters.ai cannot see queries, responses, or usage. The user's relationship with Anthropic is between them and Anthropic. We facilitate. We don't intermediate.
COMPETITIVE NOTE
We're building a dock protocol FOR Claude even though we (the Navigator) ARE Claude. This isn't a conflict — it's proof of the architecture's integrity. OBI OS is provider-agnostic. Claude gets docked the same way ChatGPT does. No special treatment. No home-field advantage. The best AI for any given query wins, regardless of provider. That's sovereignty.
INVARIANTS
INV-01: API key stored locally, encrypted. Never transmitted to 42sisters.ai.
INV-02: History import processed locally. Never uploaded.
INV-03: System prompt LATTICE preamble uses Claude's dedicated system field — cleaner than message-based injection.
INV-04: Bridge is not a proxy. Direct connection: user → Anthropic.
INV-05: Cost transparency includes input/output token breakdown.
INV-06: Extended thinking is opt-in. Default off. User controls their own token spend.
INV-07: Claude gets no special treatment in the dock. Same protocol as every other provider. Provider-agnostic means provider-agnostic.
INV-08: Vision queries route to Claude by default when docked alongside non-vision models.
INTEGRATION
| System | Relationship |
|---|---|
| SPEC_DOCK_PROTOCOL_OPENAI.md | Parallel spec. Same 5-step structure, different API quirks. Both feed the same Ring + LEARNX cross-provider profile. |
| SPEC_OBI_OS_VISION.md | Dock Protocol is the technical layer under the Dock carousel in the vision. |
| SPEC_PRIVACY_POLICY.md | INV-01 + INV-04 directly enforce the privacy policy. 42sisters.ai can't see traffic because it's not in the path. |
| SPEC_LATTICE_VIRAL.md | LATTICE onboarding (Step 3) is the viral hook moment. Watching Claude learn LATTICE in real time is the demo. |
| SPEC_HANDSHAKEX.md | Dock confirmation (Step 5) is a dock-type HANDSHAKEX. LATTICE certification level is the payload. |
Jeremy Zlabis
Chronogeometer · Visionary · Disruptor · Chief
42 Sisters AI · East York, Toronto
🍁 Φ 0.042