Get started

⌄

Run locally

Techtree runs on your own computer and shows what a Skill changes: it makes tasks from the Skill, and your agent works them once without it, or with its earlier version, and once with it. Both runs use the same setup; only the Skill changes.

Your Skill

Give this to a coding agent that can use your terminal, or to Hermes. It sets Techtree up, asks where your Skill is, and stops at every review for your answer.

Agent prompt

Strongly recommended but optional: install the Techtree skill with `npx skills add regents-ai/techtree --skill techtree`. The same information is at https://techtree.sh/agents.md. Help me test one of my Skills with Techtree. Install the exact Techtree release that page names, or check that the one I have matches. Ask me where my Skill is, and whether to compare it with no Skill or with an earlier version of it. Then use Techtree one step at a time to make the tasks, run them both ways and compare. Show me each review exactly as Techtree prints it, and stop for my answer whenever it asks for a review or approval. Never approve anything for me.

You need
  • macOS or Linux
  • uv 0.10.2 or later
  • Python 3.12, installed for you by uv
  • Docker, running
  • Hermes 0.21.3 or later
  • An account with a model provider you choose; its calls may cost money
Time
About 15 minutes to set up. Making the tasks and running the comparison then take longer, depending on your Skill and your model.

Manual setup

uv tool install --python 3.14 regents-cli==1.10.32
regents techtree forge inspect-skill path/to/your-skill
# Then plan with a provider and model you choose:
regents techtree forge plan --help
# Every review waits for your answer.

Example

The Hello World Climb is a small, fixed challenge that ships with the release. It runs the same tasks without a Skill and with a starter Skill, then shows the difference. It shows how a run works; its toy tasks say nothing about your own Skill.

You need
  • macOS or Linux
  • uv 0.10.2 or later
  • Python 3.12, installed for you by uv
  • Docker, running
  • One of these, chosen for each run:
  • A ChatGPT Plus or Pro plan, signed in with regents techtree model login; the model calls use your plan's allowance
  • Your own Prime Intellect API key, set as PRIME_API_KEY or saved with prime login; the model calls are charged to your Prime account
Model
The Climb fixes the model: gpt-6-luna, whichever way you run it.
Limits
Each task is tried once with the Skill and once without. Each try stops starting model calls after 44 calls, 500,000 input tokens or 32,000 output tokens. With 36 tasks, a run can make up to 9,504 model calls. On your own Prime key, a run also stops once it has spent $6.50. Before anything runs, Techtree shows these limits and waits for your approval.
Time
About 15 minutes to set up. How long the run takes depends on the model service.
Your data
The model calls go to OpenAI on your ChatGPT plan, or Prime Intellect on your own Prime key, only after you approve.

Example run

uv tool install --python 3.14 regents-cli==1.10.32
regents techtree setup
regents techtree doctor --climb hello-world-climb@3
regents techtree skill starter
# Prepare with the Skill it placed, then run the start command it prints.
# On your ChatGPT plan:
regents techtree climb prepare hello-world-climb@3 --skill path/to/skill --access chatgpt-plan
# Or on your own Prime key:
regents techtree climb prepare hello-world-climb@3 --skill path/to/skill --access prime-key

Steps

  1. Inspect. Techtree reads the Skill's files without running any of them and tells you what it can use.
  2. Review. Before a model sees your Skill, you see what would be sent, to which provider, and the limits. Nothing runs until you say yes.
  3. Tasks. Tasks are built and checked, and you choose which to keep. Some are held out from anyone improving the Skill.
  4. Compare. Your agent works the tasks without the Skill, or with its earlier version, and then with it. Each run shows the most it may spend and waits for your approval.
  5. Decide. Techtree compares the two runs task by task and says whether the Skill improved, regressed, was mixed or made no difference, with the held-out tasks judged on their own.

Your data

Nothing is uploaded to Techtree. Making the tasks and running them send your Skill's files and the tasks to the model provider you chose, only after you approve the review that says so. Episodes and traces stay on your computer. Publishing a finished Result is optional and uploads only its proof bundle.

Next