taprun 0.15.3

Compile your AI agent's browser-acting trajectory into a deterministic program that replays at zero LLM tokens. CLI + Python SDK.

taprun

Compile your AI agent's browser-acting trajectory into a deterministic program that replays at zero LLM tokens, with built-in drift verification.

pip install taprun

After install you get two surfaces in one package:

1 · CLI

The tap command. First invocation downloads the platform binary from the npm @taprun/cli-* packages; subsequent calls exec the cached binary. Equivalent to brew install LeonTing1010/tap/tap for Python users on Linux/Windows where Homebrew isn't standard.

tap forge https://news.ycombinator.com/        # compile a tap (Tier 0, no AI)
tap run hackernews/top                         # zero LLM tokens
tap doctor hackernews/top                      # cross-validate vs authoritative
tap mcp start                                  # MCP server for Claude Code / Cursor / Cline

2 · Python SDK

from taprun import forge, run, doctor. Each call shells out to the same tap CLI; ~30-100 ms warm spawn cost. For hot loops use tap mcp start and call via MCP instead.

from taprun import run, doctor

rows = run("hackernews/top")              # execute a compiled tap
print(rows[0]["title"])

verdict = doctor("hackernews/top")        # 'ok' / 'broken' / 'stale'

Compile from a browser-use trajectory

from browser_use import Agent
from taprun import forge

agent = Agent(task="...", llm=...)
result = await agent.run()

forge(
    trajectory=result.model_dump(),       # AgentHistoryList dict, JSON string, or path
    site="example",
    name="dashboard",
)
# writes ~/.tap/taps/example/dashboard.tap.json
# replay with: tap run example/dashboard

v1 trajectory compile covers the navigation skeleton (go_to_url, click_element, input_text, scroll, wait, done). Navigation steps replay at 0 LLM tokens. For destination pages exposing a Tier 0 source (RSS / JSON-LD / agents.json / OpenAPI), pair with tap forge <url> to compile the extraction half — that combination delivers end-to-end zero-token replay. Stagehand and raw Anthropic tool-use formats land in a follow-up.

How it works

1. Forge: inspect the site → compile a deterministic .tap.json plan (one-time cost; free for pages with a Layer 1 source).
2. Run: replay the plan, no LLM in the loop.
3. Doctor: independently fetch the authoritative source and diff. Catches drift the agent would miss on a self-replay.
4. Heal: cached patches replay at 0 tokens; LLM only invoked when the patch cache misses (Pro tier).

Configuration

# Override the binary path (e.g. to use a brew-installed `tap`)
export TAPRUN_BIN=/opt/homebrew/bin/tap

# AI key for the forge pipeline (BYOK; Hacker tier and above)
tap config set ai.key sk-ant-...

Pricing

• Free: 65+ community taps, run, doctor, Tier 0 forge (the deterministic compile path when Layer 1 is available).
• Hacker ($9/mo, BYOK): full forge pipeline with Layer 4 AI fallback.
• Pro ($29/mo): heal + refresh + scheduling. 100% local.

The MCP server / CLI / forge / doctor binary is closed-source. The public Chrome extension runtime (https://github.com/LeonTing1010/tap) is MIT.

Links

• Homepage: https://taprun.dev/?utm_source=pypi&utm_medium=package&utm_campaign=taprun
• Docs: https://taprun.dev/docs?utm_source=pypi&utm_medium=package&utm_campaign=taprun-docs
• Issues: https://github.com/LeonTing1010/tap/issues
• Skills: https://github.com/LeonTing1010/tap-skills

MIT License (the public extension; the tap binary distributed by this package is proprietary).