How do you run Paperclip or Orca agents on one cost-optimized API?
Neither Paperclip nor Orca has a model endpoint of its own, so you set the API on the agents they run. In Paperclip, Claude Code takes ANTHROPIC_BASE_URL in the agent's environment variables, and Codex and OpenCode take a gateway through PAPERCLIP_CODEX_PROVIDERS and PAPERCLIP_OPENCODE_PROVIDERS. Orca launches each agent CLI as it is, so you configure Claude Code or Codex the usual way. Pointed at Standard Compute, every agent sends each step to the cheapest model that can do it, on one plan from $19 a month.
# Claude Code agent in Paperclip (claude_local adapter)
ANTHROPIC_BASE_URL=https://api.stdcmpt.com
# Bind your Standard Compute key as a secret (Company Settings → Secrets)
ANTHROPIC_API_KEY=<secret: Standard Compute key>
# Model field on the same page: the router picks per step
# anthropic/claude-standardcompute
Paperclip values from its Claude Code, Codex and OpenCode adapter docs; Orca follows its own GLM guide: configure the harness, then let Orca launch it. Export STANDARDCOMPUTE_API_KEY in your login shell for Codex in Orca. Checked October 15, 2026.
Where each orchestrator takes the endpoint
| Agent | Paperclip | Orca |
|---|---|---|
| Claude Code | Agent → Harness / Runtime → Environment variables: ANTHROPIC_BASE_URL=https://api.stdcmpt.com and the key as a secret | ~/.claude/settings.json env block, read by every Claude Code session Orca opens |
| Codex | PAPERCLIP_CODEX_PROVIDERS on the Paperclip host, wire_api: "responses", key as the agent's OPENAI_API_KEY | [model_providers.standardcompute] in ~/.codex/config.toml; Orca mirrors it into managed Codex accounts |
| OpenCode | PAPERCLIP_OPENCODE_PROVIDERS with an explicit models map, plus PAPERCLIP_OPENCODE_SMALL_MODEL | A provider in your OpenCode config, as you would without Orca |
From the Paperclip adapter docs and Orca's agent docs, checked October 15, 2026. Base URL https://api.stdcmpt.com for Anthropic-format agents, https://api.stdcmpt.com/v1 for OpenAI-format ones.
Paperclip: what to know before you set it
- Environment variables on an agent's Harness / Runtime page are passed to that agent's runtime on every run. Bind the key as a secret from Company Settings → Secrets; with PAPERCLIP_SECRETS_STRICT_MODE=true, inline values for key-like names are rejected.
- Codex has no flag or variable for a custom endpoint, so Paperclip writes PAPERCLIP_CODEX_PROVIDERS into its managed Codex home for the run and restores the file afterwards. An agent that sets its own CODEX_HOME is left alone.
- OpenCode only resolves a provider/model that is listed in the provider's models map. The OpenCode settings apply when dangerouslySkipPermissions is on (the default) and the agent runs locally.
- With an API key set, Paperclip's Claude Code model dropdown reads /v1/models at your ANTHROPIC_BASE_URL, so Standard Compute's model IDs appear there.
Orca: configure the harness, then launch it
Orca's docs say it works with any CLI agent: the agent picker launches a process in a terminal, and local panes open your shell as a login shell. Its own guide for third-party models gives the rule to follow: configure the provider wherever the harness stores provider and model settings, then let Orca launch that harness in the right worktree.
So every Claude Code session Orca starts reads ~/.claude/settings.json, and every Codex session reads ~/.codex/config.toml; for managed Codex accounts Orca mirrors that file into the active runtime home. Orca's usage rows marked inferred pricing come from its local price table, not from your provider's bill.
The catch
Every agent on one key shares that key's request limits (300 a minute on Starter) and one compute budget. Ten agents in parallel use the budget ten times as fast, and at the budget new requests from all of them stop with HTTP 402 until renewal or an upgrade; running requests finish and nothing is billed on top. Paperclip's Costs page and Orca's usage rows show what the agent CLIs estimate, not your plan's bill.
Where Standard Compute fits
Standard Compute is one OpenAI- and Anthropic-compatible API for every agent in the fleet: the model standardcompute (anthropic/claude-standardcompute for Claude-format agents) sends each step to the cheapest model that can do it, and for every model we vet providers for price, uptime and speed, with automatic fallback. Opt-in Europe or US Compliance rules apply per key, team or company, and we don't train on your prompts or store them. Plans start at $19 a month for a $20 compute budget.
FAQ
Can Paperclip use a custom OpenAI-compatible endpoint?
Not for itself, since Paperclip does not call models directly. It passes the endpoint to the agents it runs: PAPERCLIP_OPENCODE_PROVIDERS for OpenCode agents, PAPERCLIP_CODEX_PROVIDERS for Codex agents (which speak the Responses API), and environment variables such as ANTHROPIC_BASE_URL for Claude Code agents.
How do I set ANTHROPIC_BASE_URL for a Claude Code agent in Paperclip?
Open the agent, go to Harness / Runtime and add it under Environment variables: ANTHROPIC_BASE_URL=https://api.stdcmpt.com (no /v1) and ANTHROPIC_API_KEY bound to a secret holding your Standard Compute key. Paperclip passes these to Claude Code on every run.
Does Orca support custom model providers?
Orca launches the agent CLIs you already use, so a provider works when the agent supports it. Set it where that agent keeps its settings, such as ~/.claude/settings.json for Claude Code or ~/.codex/config.toml for Codex, and every session Orca starts uses it.
Will running many agents in parallel cost more on Standard Compute?
Not more than the plan price. Parallel agents use the compute budget faster, and when it is used up, new requests stop with HTTP 402 instead of billing overage. Plans start at $19 a month for a $20 compute budget; larger plans carry larger budgets and higher per-key request limits.
Is Paperclip or Orca free?
Both are free and open source under the MIT license. Paperclip is self-hosted with no account required; Orca is a desktop app for macOS, Windows and Linux. You pay only for the models the agents call.