Capx
Platform / Agent Adapters

Model-agnostic by design.

Capx does not lock you into one LLM. Use Claude, GPT, Codex, Cursor, or any HTTP endpoint as an agent backend. Switch models per agent, per task, or per playbook step.

Claude

private beta
AnthropicOpus, Sonnet, Haiku

Strategy, writing, analysis, complex reasoning

adapter: claude
model: claude-sonnet-4-6
budget: 200
memory: persistent

GPT

private beta
OpenAIGPT-4o, GPT-4o-mini

General purpose, structured output, tool use

adapter: openai
model: gpt-4o
budget: 150
memory: persistent

HTTP

private beta
CustomAny endpoint

Proprietary models, internal APIs, fine-tuned models

adapter: http
config:
  endpoint: https://your-api.com/agent
  method: POST
  timeout: 30s
  retry: 3

Adapter registry

Every adapter is a first-class integration with typed schemas, automatic prompt formatting, and response parsing. One interface, many backends.

AdapterProviderModelsFeaturesUse caseStatus
ClaudeAnthropicOpus, Sonnet, HaikuPrompt caching, streaming, tool use, persistent memoryStrategy, writing, analysisprivate beta
GPTOpenAIGPT-4o, GPT-4o-miniStructured output, function calling, streamingGeneral purpose, code, dataprivate beta
CodexOpenAICodexCode generation, refactoring, static analysisEngineering tasksprivate beta
CursorCursorCursor AgentWorkspace access, file editing, test runningPR creation, code reviewprivate beta
HTTPCustomAny endpointCustom headers, auth, retries, timeoutFine-tuned models, internal APIsprivate beta
0adapters in private beta

Per-step model switching

Use the right model for each step. Research on Opus for depth. Drafting on Sonnet for speed and cost. Code samples on GPT-4o. Final review back on Opus.

Multi-model playbookresearch-and-write.yaml
playbook: research-and-write
steps:
  - name: research
    agent: strategist
    adapter: claude
    model: claude-opus-4
    tool: web-research

  - name: draft
    agent: marketer
    adapter: claude
    model: claude-sonnet-4-6
    tool: content-write

  - name: code-samples
    agent: engineer
    adapter: openai
    model: gpt-4o
    tool: code-generate

  - name: review
    agent: strategist
    adapter: claude
    model: claude-opus-4
    tool: content-review
researchAnthropic
claude-opus-4

Deep reasoning for comprehensive research

draftAnthropic
claude-sonnet-4-6

Fast and cost-effective for content drafting

code-samplesOpenAI
gpt-4o

Strong at structured code generation

reviewAnthropic
claude-opus-4

Deep reasoning for quality evaluation

Automatic fallback chains

Automatic fallback chains ensure continuity when a model is unavailable. Configure primary and fallback models per agent. Tasks are never dropped.

Primary
claude-sonnet-4-6

First choice. Used when available and within budget.

Fallback 1
gpt-4o

Automatic switch if Claude is unavailable or rate-limited.

Fallback 2
claude-haiku-4

Last resort. Lower cost, still capable for most tasks.

fallback:claude-sonnet-4-6->gpt-4o->claude-haiku-4->queue

Connect your first model.

Claude, GPT, Codex, Cursor, or your own HTTP endpoint. One interface, any backend.