Claude Code with Synaplan — Anthropic-Compatible API

Synaplan ships an Anthropic Messages API-compatible gateway, so Claude Code — and any other client that speaks the Anthropic protocol — can use your Synaplan instance as its backend. You keep one API key, one budget, one statistics page, and full control over which upstream models serve the traffic.

The gateway lives at POST /v1/messages (Claude Code calls /v1/messages?beta=true). This is not Synaplan's native chat SSE at /api/v1/messages/stream — the two share nothing but a similar name.

Why route Claude Code through Synaplan?

• One key, metered like everything else — requests show up in your usage statistics with their own Messages API channel, count against your Synaplan budget (when the operator key pays), and respect your rate limits.
• A session trail in your chat list — each Claude Code session gets a rolling 2–3 sentence AI summary chat, so you can always see what was transported over the API.
• Bring your own key or use the operator's — per-user encrypted Anthropic keys with an optional install-wide fallback.
• Any model, not just Claude — model aliases can point Claude Code at OpenAI or Gemini catalog models; the gateway translates the wire format.

Setup (three steps)

1. Enable the gateway (admin, once per install)

An administrator opens Channels → AI Agents and turns on Enable Messages gateway. On the same page the admin decides the key policy:

• Allow operator key — users without their own Anthropic key use the install-wide key; their usage is billed against their Synaplan cost budget.
• BYO keys only (default) — every user saves their own Anthropic API key.

2. Get your credentials (user)

1. Create a Synaplan API key under Channels → API Keys (sk_…).
2. Optional: save your own Anthropic key under Channels → AI Agents. BYO keys require at least the Pro plan — in exchange your gateway traffic is metered at zero cost (statistics only, your Synaplan budget is never charged, because Anthropic bills your key directly).

3. Point Claude Code at Synaplan

export ANTHROPIC_BASE_URL="https://your-synaplan-host"
export ANTHROPIC_API_KEY="sk_your_synaplan_api_key"
# or: export ANTHROPIC_AUTH_TOKEN="sk_your_synaplan_api_key"
claude

Set exactly one credential variable: ANTHROPIC_API_KEY is sent as x-api-key, ANTHROPIC_AUTH_TOKEN as Authorization: Bearer. Setting both is the most common foot-gun.

To make this permanent, add the two export lines to your shell profile (~/.zshrc or ~/.bashrc), or put them in Claude Code's settings.json under env.

The Channels → AI Agents page shows this exact snippet with your host pre-filled, plus a copy button.

Endpoints

Authentication uses the same Synaplan API keys as the OpenAI-compatible API: x-api-key: sk_… or Authorization: Bearer sk_….

Billing: whose key pays?

Every response carries x-synaplan-budget-percent and x-synaplan-budget-remaining headers so scripts can react before hitting the wall.

See what happened: statistics & session summaries

• Statistics (/statistics in the app) break usage down by communication channel — Messages API, OpenAI API, Web, Widget, MCP, … — and the activity log shows a channel column per request.
• Session summary chats: each API session appears as a chat in your normal chat list (marked with a console icon), holding a rolling 2–3 sentence summary of what the session did. Summaries are generated asynchronously by your configured sorting model — fast and cheap — and are metered like any other usage.

Using non-Claude models

Admins can map Claude Code model IDs to any catalog model via model aliases (Channels → AI Agents), e.g. point claude-sonnet-4-5 at a Gemini or OpenAI model. The gateway translates the Anthropic wire format transparently:

• OpenAI — Chat Completions, including tool calls
• Google/Gemini — generateContent / streamGenerateContent

Anthropic-only fields (e.g. thinking) are stripped before the upstream call. This works technically; note that Anthropic does not officially support running Claude Code against non-Claude models.

Claude Code + your knowledge base (MCP)

For RAG and memories inside Claude Code, prefer the native MCP connection over server-side context injection:

claude mcp add --transport http synaplan https://your-synaplan-host/mcp \
  --header "Authorization: Bearer sk_your_synaplan_api_key"

Claude Code then discovers rag_search, memory_search and the other Synaplan tools by itself and calls them when useful.

Platform configuration (self-hosting)

The gateway is controlled by BCONFIG flags (group MESSAGES_GATEWAY). Defaults are off except budget notices and session summaries — and nothing applies until the master switch is on:

All of the user-facing switches are also editable in the Channels → AI Agents UI. The seeded defaults are inserted on container start (insert-if-missing), so an operator's explicit value survives every deploy.

Troubleshooting

See also

• MCP Server & Client — connect Claude Code to your knowledge base
• Interactive API (Swagger) — the OpenAI-compatible sibling
• REST Architecture
• Code Examples
• Claude Code documentation (Anthropic)