browserbase · skippyom · May 8, 2026
diff --git a/skills/autobrowse/.env.example b/skills/autobrowse/.env.example
@@ -1,3 +1,9 @@
 ANTHROPIC_API_KEY=sk-ant-...
+# Optional: use OpenAI-compatible Chat Completions instead of Anthropic.
+# AUTOBROWSE_PROVIDER=openai
+# AUTOBROWSE_MODEL=gpt-4.1
+# OPENAI_API_KEY=sk-...
+# OPENAI_BASE_URL=https://api.openai.com/v1
+# For OpenRouter/LiteLLM/etc, point OPENAI_BASE_URL at that provider's /v1 endpoint.
 BROWSERBASE_API_KEY=bb_live_...
 BROWSERBASE_PROJECT_ID=your-project-id
diff --git a/skills/autobrowse/README.md b/skills/autobrowse/README.md
@@ -13,7 +13,7 @@ The output is a `skill.md` — a site-specific playbook any agent can follow. On
 - Node.js 18+
 - [Claude Code](https://claude.ai/code)
 - `browse` CLI: `npm install -g @browserbasehq/browse-cli`
-- `ANTHROPIC_API_KEY` in your environment
+- `ANTHROPIC_API_KEY` in your environment, or `AUTOBROWSE_PROVIDER=openai` with `OPENAI_API_KEY`
 - For bot-protected sites: `BROWSERBASE_API_KEY` + `BROWSERBASE_PROJECT_ID`
 
 ## Setup
@@ -25,6 +25,23 @@ npm install
 cp .env.example .env   # fill in your API keys
 ```
 
+By default, the inner agent uses Anthropic. To use an OpenAI-compatible provider instead:
+
+```bash
+AUTOBROWSE_PROVIDER=openai \
+OPENAI_API_KEY=sk-... \
+node scripts/evaluate.mjs --task my-portal --model gpt-4.1
+```
+
+For OpenRouter, LiteLLM, or another Chat Completions-compatible gateway, set `OPENAI_BASE_URL`:
+
+```bash
+AUTOBROWSE_PROVIDER=openai \
+OPENAI_API_KEY=sk-or-... \
+OPENAI_BASE_URL=https://openrouter.ai/api/v1 \
+node scripts/evaluate.mjs --task my-portal --model anthropic/claude-sonnet-4.5
+```
+
 ## Your project structure
 
 Create this in your working directory before running `/autobrowse`:
@@ -80,7 +97,7 @@ Inspired by [Karpathy's autoresearch](https://github.com/karpathy/autoresearch)
 outer agent (Claude Code + /autobrowse skill)
   └── reads trace → improves strategy.md → repeats
 
-inner agent (scripts/evaluate.mjs → Anthropic API)
+inner agent (scripts/evaluate.mjs → Anthropic API or OpenAI-compatible Chat Completions)
   └── browse open → snapshot → click → snapshot → ...
   └── writes traces/ with summary, full trace, screenshots
 ```
diff --git a/skills/autobrowse/REFERENCE.md b/skills/autobrowse/REFERENCE.md
@@ -10,14 +10,22 @@ node ${CLAUDE_SKILL_DIR}/scripts/evaluate.mjs --task <name> [options]
 |------|---------|-------------|
 | `--task <name>` | required | Task name — matches `tasks/<name>/` directory |
 | `--env local\|remote` | `local` | Browser environment |
-| `--model <model>` | `claude-sonnet-4-6` | Claude model for the inner agent |
+| `--provider anthropic\|openai` | `anthropic` | Model provider for the inner agent |
+| `--model <model>` | provider-specific | Model for the inner agent (`claude-sonnet-4-6` for Anthropic, `gpt-4.1` for OpenAI-compatible) |
 | `--run-number N` | auto-increment | Force a specific run number |
 
 ## Environment variables
 
 | Variable | Required | Description |
 |----------|----------|-------------|
-| `ANTHROPIC_API_KEY` | Yes | Claude API key |
+| `AUTOBROWSE_PROVIDER` | No | `anthropic` or `openai`; same as `--provider` |
+| `AUTOBROWSE_MODEL` | No | Default model when `--model` is omitted |
+| `ANTHROPIC_API_KEY` | Anthropic only | Claude API key |
+| `OPENAI_API_KEY` | OpenAI-compatible only | API key for OpenAI, OpenRouter, LiteLLM, etc. |
+| `OPENAI_BASE_URL` | No | OpenAI-compatible `/v1` base URL; defaults to `https://api.openai.com/v1` |
+| `OPENAI_ORGANIZATION` | No | Optional OpenAI organization header |
+| `OPENAI_SITE_URL` | No | Optional `HTTP-Referer` header for gateways such as OpenRouter |
+| `OPENAI_APP_NAME` | No | Optional `X-Title` header for gateways such as OpenRouter |
 | `BROWSERBASE_API_KEY` | Remote only | Browserbase API key |
 | `BROWSERBASE_PROJECT_ID` | Remote only | Browserbase project ID |
 
@@ -29,7 +37,7 @@ Each run writes to `traces/<task>/run-NNN/`:
 |------|-------------|
 | `summary.md` | Duration, cost, turn-by-turn decision log, final output |
 | `trace.json` | Full tool call log — every command and response |
-| `messages.json` | Raw Anthropic API message history |
+| `messages.json` | Raw normalized message history |
 | `screenshots/` | Visual captures saved during the run |
 
 `traces/<task>/latest` is a symlink to the most recent run.
@@ -41,6 +49,8 @@ Each run writes to `traces/<task>/run-NNN/`:
 | `claude-sonnet-4-6` | $$ | Default — good balance of speed and accuracy |
 | `claude-opus-4-6` | $$$$ | Hardest tasks, complex multi-step workflows |
 | `claude-haiku-4-5-20251001` | $ | Simple tasks, high-volume iteration |
+| `gpt-4.1` | $$ | Default for OpenAI-compatible mode |
+| `gpt-4.1-mini` | $ | Lower-cost OpenAI-compatible iteration |
 
 ## Skill lifecycle
 

diff --git a/skills/autobrowse/SKILL.md b/skills/autobrowse/SKILL.md
@@ -2,7 +2,7 @@
 name: autobrowse
 description: Self-improving browser automation via the auto-research loop. Iteratively runs a browsing task, reads the trace, and improves the navigation skill (strategy.md) until it reliably passes. Supports parallel runs across multiple tasks using sub-agents. Use when you want to build or improve browser automation skills for specific website tasks.
 license: See LICENSE.txt
-compatibility: "Requires Node.js 18+, browse CLI, and ANTHROPIC_API_KEY. Run from the autobrowse app directory."
+compatibility: "Requires Node.js 18+, browse CLI, and either ANTHROPIC_API_KEY or AUTOBROWSE_PROVIDER=openai with OPENAI_API_KEY. Run from the autobrowse app directory."
 allowed-tools: Bash Read Write Edit Glob Grep Agent
 metadata:
   author: browserbase
@@ -96,14 +96,17 @@ Check that `./autobrowse/tasks/<task>/task.md` exists (scaffold it from the temp
 
 ### Requirements
 
-- `ANTHROPIC_API_KEY` must be in the environment (or in a `.env` file in CWD — `evaluate.mjs` auto-loads it). If missing, the harness prints a clear error and exits; don't hunt for keys in other paths.
+- By default, `ANTHROPIC_API_KEY` must be in the environment (or in a `.env` file in CWD — `evaluate.mjs` auto-loads it). If missing, the harness prints a clear error and exits; don't hunt for keys in other paths.
+- To use OpenAI-compatible Chat Completions instead, set `AUTOBROWSE_PROVIDER=openai` or pass `--provider openai`, then set `OPENAI_API_KEY`. For OpenRouter, LiteLLM, or another compatible gateway, also set `OPENAI_BASE_URL` to that provider's `/v1` endpoint.
 
 ### Run the inner agent
 
 ```bash
 node ${CLAUDE_SKILL_DIR}/scripts/evaluate.mjs --task <task-name> --workspace ./autobrowse
 # or for bot-protected sites:
 node ${CLAUDE_SKILL_DIR}/scripts/evaluate.mjs --task <task-name> --workspace ./autobrowse --env remote
+# or with an OpenAI-compatible provider:
+AUTOBROWSE_PROVIDER=openai OPENAI_API_KEY=sk-... node ${CLAUDE_SKILL_DIR}/scripts/evaluate.mjs --task <task-name> --workspace ./autobrowse --model gpt-4.1
 ```
 
 This runs the browser session and writes a full trace to `./autobrowse/traces/<task>/latest/`.

diff --git a/skills/autobrowse/package.json b/skills/autobrowse/package.json
@@ -1,7 +1,7 @@
 {
   "name": "autobrowse",
   "version": "0.1.0",
-  "description": "Self-improving browser agent via skill learning — autoresearch pattern + Browse CLI + Anthropic API",
+  "description": "Self-improving browser agent via skill learning — autoresearch pattern + Browse CLI + Anthropic/OpenAI-compatible APIs",
   "type": "module",
   "scripts": {
     "evaluate": "node scripts/evaluate.mjs"