Connect your tools

Every recipe assumes the bridge is running on the same computer. See Serve API to start it. The base URLs are:

• OpenAI: http://127.0.0.1:8080/v1 (Codex, opencode, Continue, Open WebUI, LiteLLM, the openai SDK)
• Anthropic: http://127.0.0.1:8080, no /v1 (Claude Code, the anthropic SDK)

model can be any string. Any API key works, unless you started the bridge with a token; then use the token as the key.

Context settings for coding agents

Codex, Claude Code and opencode don’t know how big a model called pooled is. Without the room’s context they never compact, and the room refuses the first prompt that is too long. The bridge prints the right settings for your room once its model is ready. GET /health returns the context as ctx.

Room	Context	Codex compact limit	Claude Code output	opencode output
Qwen3 1.7B	16384	13107	4096	4096
Qwen3 1.7B, room short of memory for 16K	8192	6553	2048	2048
Qwen3.8 27B	16384	13107	4096	4096
Qwen3.8 27B, ?ctx=32768	32768	26214	8192	8192
Qwen3.6 35B MoE	32768	26214	8192	8192
Qwen3.6 35B MoE, ?ctx=65536	65536	52428	16384	8192

The host sets the context with ?ctx= on the room page before starting the model. Use the MoE or the 27B at 32K for agents. The 1.7B is too small.

Codex CLI

Add a provider to ~/.codex/config.toml, with the numbers from the banner (this is a 32K room):

~/.codex/config.toml

model = "pooled"
model_provider = "pooled"
model_context_window = 32768            # the room's context
model_auto_compact_token_limit = 26214  # 0.8 of it

[model_providers.pooled]
name = "Pooled"
base_url = "http://127.0.0.1:8080/v1"
wire_api = "responses"

Then run codex in your project.

• model = "gpt-5-codex" instead silences the “model metadata not found” warning. Codex then sends apply_patch as a custom tool, which works too.
• Codex gives up on a stream silent for 5 minutes. The bridge sends an event every 10 s while it waits, but on a very slow room add stream_idle_timeout_ms = 900000 to the provider.
• With a token, add env_key = "POOLED_TOKEN" to the provider and export POOLED_TOKEN.
• Codex always sends a web_search tool. The bridge skips it with one warning.

Claude Code

Run this in your project (the banner prints it with your room’s numbers):

ANTHROPIC_BASE_URL=http://127.0.0.1:8080 ANTHROPIC_API_KEY=pooled \
  CLAUDE_CODE_MAX_CONTEXT_TOKENS=32768 CLAUDE_CODE_MAX_OUTPUT_TOKENS=8192 \
  CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 claude --model pooled

• CLAUDE_CODE_MAX_CONTEXT_TOKENS is the room’s context, so Claude Code compacts in time.
• CLAUDE_CODE_MAX_OUTPUT_TOKENS is a quarter of the context. Without it, Claude Code says “Prompt is too long” before sending anything on a room under about 32K.
• CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 stops side requests, such as titles, from queuing between the agent’s steps.

MCP servers work (up to 1024 tools). Anthropic’s own server tools, like web search, are skipped.

opencode

The banner prints an opencode.json for your room. Save it in your project, or in ~/.config/opencode/opencode.json. For a 32K MoE room:

opencode.json

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "pooled": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Pooled room",
      "options": { "baseURL": "http://127.0.0.1:8080/v1", "apiKey": "x" },
      "models": {
        "pooled/qwen3.6-35b-moe": { "name": "Qwen3.6 35B MoE · Q4", "tool_call": true, "limit": { "context": 32768, "output": 8192 } }
      }
    }
  },
  "model": "pooled/pooled/qwen3.6-35b-moe"
}

Then run OPENCODE_DISABLE_CLAUDE_CODE=1 opencode. That variable stops opencode from loading ~/.claude/CLAUDE.md and your skills into every prompt. Pass --title to opencode run, or set a small_model, to skip its title request.

Continue

In ~/.continue/config.yaml, then pick Pooled room in Continue’s model menu:

~/.continue/config.yaml

models:
  - name: Pooled room
    provider: openai
    model: pooled
    apiBase: http://127.0.0.1:8080/v1
    apiKey: pooled
    roles: [chat, edit]

The key must not be empty. Leave autocomplete to a local model: a room is too slow for completions as you type.

Open WebUI

Open WebUI also uses port 8080, so start the bridge on another port:

npx @pooled/cli serve 4TKG9P --port 8081

In Open WebUI, add a connection under Settings → Connections → OpenAI API: URL http://127.0.0.1:8081/v1, any key. The room’s model shows as pooled/<model>.

The bridge only answers on 127.0.0.1, so Open WebUI in Docker can’t reach it. Run Open WebUI on the host, or its container with --network=host.

OpenAI and Anthropic SDKs

Change the base URL and nothing else:

from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8080/v1", api_key="pooled")

from anthropic import Anthropic
client = Anthropic(base_url="http://127.0.0.1:8080", api_key="pooled")

The SDKs also read OPENAI_BASE_URL / OPENAI_API_KEY and ANTHROPIC_BASE_URL / ANTHROPIC_API_KEY, so existing code runs unchanged. Tools, streaming, JSON output and the Responses API all work. See Tool calling.

LiteLLM

config.yaml

model_list:
  - model_name: pooled
    litellm_params:
      model: openai/pooled
      api_base: http://127.0.0.1:8080/v1
      api_key: pooled

Run the proxy on the same computer as the bridge. If others use your proxy, their prompts go to the room too.

curl

curl http://127.0.0.1:8080/health

curl http://127.0.0.1:8080/v1/chat/completions -H 'content-type: application/json' \
  -d '{"model": "pooled", "messages": [{"role": "user", "content": "Hi"}], "stream": true}'

curl http://127.0.0.1:8080/v1/messages -H 'content-type: application/json' -H 'anthropic-version: 2023-06-01' \
  -d '{"model": "pooled", "max_tokens": 200, "messages": [{"role": "user", "content": "Hi"}]}'

With a token, add -H "Authorization: Bearer $POOLED_TOKEN". Status codes are on HTTP endpoints.