Skip to content
prune.

Point Cline at Prune

Three fields. The provider key never sits in the agent. Cache and cheaper routing cut waste. Shield stops junk and runaway spend. Every call comes back with a signed receipt.

Save

Exact and semantic cache plus cheaper-model routing when the prompt is unique, not only on repeats.

Shield

Real OpenAI / Anthropic / OpenRouter keys stay in Prune. Optional spend caps reject over-budget calls before the provider.

Prove

Each response includes signed prune_metadata. Check it at /verify.

Do this in Cline

  1. Join at withprune.com, vault a provider key, copy your prune_… key.
  2. Cline → Settings → API Provider → OpenAI Compatible.
  3. Paste these three fields. Change only Model ID for another vaulted provider.
Base URL:  https://api.withprune.com/v1
API Key:   prune_your_key
Model ID:  gpt-4o-mini

Do not paste provider sk-… keys into Cline. If Cline loops on Plan Created, switch Plan → Act and ask a concrete task.

Prove it without opening Cline

One request. Look for prune_metadata and receipt_signature in the response.

curl https://api.withprune.com/v1/chat/completions \
  -H "Authorization: Bearer prune_your_key" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"ping"}],"max_tokens":16}'

Same three fields

Kilo Code and Zoo Code: API Provider → OpenAI Compatible, same paste. Use BYOK, not Kilo Gateway. If Zoo imported Roo settings, delete leftover sk-… keys.

OpenCode uses a config file plus auth, not a VS Code picker:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "prune": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Prune",
      "options": {
        "baseURL": "https://api.withprune.com/v1"
      },
      "models": {
        "gpt-4o-mini": { "name": "GPT-4o mini (via Prune)" }
      }
    }
  },
  "model": "prune/gpt-4o-mini"
}
opencode auth login
# Other → provider id: prune → paste prune_…
# Then select prune/gpt-4o-mini

What this is not

Closed subscription IDE backends (Cursor plan, Copilot subscription, Claude Pro) do not route through Prune. Leave those as-is. Prune only sees traffic you send to api.withprune.com.

Copy-paste configs and a manual test plan: github.com/surenkotian/prune-coding-agents. Generic app setup: prune-quickstart.

A 9-turn API-style agent simulation in CI showed 65.4% input-token savings from prefix cache on traffic through Prune. That is a long-thread cache result, not the unique-path claim. Unique-path Savings is measured separately: 93.83% median cost reduction on 40 unique prompts, 0/40 cache hits (run 2026-08-04). Details: docs/coding-agents.md. Verify receipts at /verify.