tstack on a Budget: Claude Code Skills, One Cheap Key
Run Theo's tstack skills in Claude Code or Codex on one cheap API key: custom base URL setup, which model group per skill, and why council and sol-loop differ.

tstack is the set of agent skills Theo (t3dotgg) uses every day, cleaned up and open-sourced on October 24, 2026. It works in Claude Code and Codex, and it passed 1,700 GitHub stars within three days (1,776 when we checked on October 27).
The skills are great. They also burn a lot of tokens. babysit-pr watches a pull request through CI and review bots and fixes what they find. full-send takes a change all the way to a merged PR while nobody is watching. An agent that loops on CI output for an hour can get expensive fast.
This guide shows how to run tstack in Claude Code (or Codex) through a custom base URL and a cheap, OpenAI/Anthropic/Gemini-compatible key from XCoffee. It also covers what does not work well on a budget relay, so you don't waste an afternoon finding out.
What's in tstack: the 9 skills
| Skill | Say | What it does |
|---|---|---|
full-send |
"full send" | Proves, reviews, files, babysits and merges a change, unattended |
full-stop |
"full stop" | Same as full-send, but stops when the PR is ready for you to merge |
babysit-pr |
"babysit" | Watches a PR through CI and review bots and fixes real findings |
file-pr |
"file a PR" | Writes a short PR that leads with the problem |
takeover |
"take over" | Finishes someone else's PR or branch and keeps their credit |
modernize |
"modernize" | Rebases an old PR and decides if it should still exist |
council |
"council" | Gets review from three models from different providers and loops until they agree |
sol-loop |
"sol loop" | Reviews a change with GPT Sol in a loop until both agents agree |
leaving-pr-comments |
(automatic) | Formats comments the agent posts for you |
Some skills depend on others: full-send uses sol-loop, file-pr and babysit-pr; full-stop uses full-send; file-pr, babysit-pr and modernize use leaving-pr-comments; and takeover uses modernize, file-pr, babysit-pr and leaving-pr-comments. Install them all.
Step 1: Install tstack for Claude Code or Codex
You need an authenticated GitHub CLI (gh) and python3 (for the pr-wait script in babysit-pr).
git clone https://github.com/t3dotgg/tstack
cp -R tstack/skills/* ~/.claude/skills/ # Claude Code
cp -R tstack/skills/* ~/.codex/skills/ # Codex
Step 2: Set a custom base URL in Claude Code
Claude Code reads documented environment variables for gateways (env var reference). XCoffee speaks the Anthropic Messages API, and you pick a group instead of a single model. The group names xcoffee-coding, xcoffee-fast and xcoffee work directly as model names. This setup works in Claude Code:
export ANTHROPIC_BASE_URL="https://api.xcoff.ee" # host root, no /v1
export ANTHROPIC_API_KEY="<your XCoffee key>" # ANTHROPIC_AUTH_TOKEN also works
export ANTHROPIC_MODEL="xcoffee-coding" # main agent: edits and fixes
export ANTHROPIC_SMALL_FAST_MODEL="xcoffee-fast" # background tasks
export CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 # required
claude
ANTHROPIC_BASE_URLis the host root. Claude Code adds/v1/messagesitself.ANTHROPIC_API_KEY(sent asx-api-key) andANTHROPIC_AUTH_TOKEN(sent asAuthorization: Bearer) both work.CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1is required. Without it, requests fail with400 Server-managed tools unavailable.- Claude Code's web search is not available through XCoffee. It is a server-managed tool. Everything else tstack uses, like file edits, shell and
gh, runs on your machine and works.
Keep the key out of your repo. Put these lines in your shell profile or a local, git-ignored env file.
Codex
Codex reads ~/.codex/config.toml. With Codex 0.164, use the Responses wire API (wire_api = "chat" is no longer supported) and define a custom provider (Codex advanced config):
model = "xcoffee-coding"
model_provider = "xcoffee"
web_search = "disabled"
[features]
multi_agent = false
[model_providers.xcoffee]
name = "XCoffee"
base_url = "https://api.xcoff.ee/v1"
env_key = "XCOFFEE_API_KEY"
wire_api = "responses"
Then export XCOFFEE_API_KEY=... before you run codex. You need both web_search = "disabled" and multi_agent = false. Without them, Codex sends tools XCoffee doesn't support and requests fail with unsupported_tool. That means Codex's web search and multi-agent features are not available through XCoffee.
Step 3: Match each tstack skill to a cheap model group
XCoffee plan details updated November 12, 2026.
XCoffee has three groups. Each one routes to a member model automatically and fails over on eligible errors before output starts. Credits use weighted input, cache-read and output usage at the selected group’s multiplier. See the current group rates and credit formula.
| Group | API model name | Use it for |
|---|---|---|
| Coding | xcoffee-coding |
Code edits, debugging, adding tests |
| Fast | xcoffee-fast |
Extraction, classification, formatting, short answers |
| Default | xcoffee |
General work. Check its intended use first |
A good split for tstack:
babysit-pr: run the main agent onxcoffee-coding. Most of the tokens go to reading CI logs and bot comments, which is long input with short output. If you hand log triage to a subagent,xcoffee-fastis the cheapest place for it.full-stop/full-send:xcoffee-codingfor the main agent. These chain several skills, so watch your hourly cap (below).file-pr,leaving-pr-comments: short writing tasks.xcoffee-fasthandles them well.modernize,takeover: rebases and judgment calls. Usexcoffee-coding.
How far it goes depends on your plan and request mix. Uncached input, cache reads and output carry different weights; output includes reported reasoning, and cache writes count as full input. Check the current plans and credit rules for prices, daily allowances, rolling hourly caps and group multipliers. Your account shows your pass’s actual limits.
Can you run tstack council and sol-loop on a cheap API?
This is where you need to be honest with yourself.
council (per its SKILL.md) asks for "one strong model from each of three different providers", for example Claude Opus, GPT Sol and Gemini Pro. In Claude Code it uses the harness's subagent tools, and the skill says that if it can't reach three providers it should report that instead of substituting.
XCoffee can't give you that. You pick a group, not a model, and request parameters or headers can't select a member. The groups also overlap: MiniMax-M3.1-Flash, DeepSeek-V4.1-Flash and Gemini-3.8-Flash show up in more than one group, so two "different" reviewers can land on the same model.
What you can run is a budget council. tstack's README says to put personal settings, "like which models sit on your council", in CLAUDE.md or AGENTS.md, and the skill uses council members you name. Something like this:
## tstack council members
For `council`, use three subagents: `xcoffee-coding`, `xcoffee`, and `xcoffee-fast`.
This is a budget council on one relay, not three providers. Label it that way in the PR.
In Claude Code, subagents accept a model field (subagent docs), and the group names work as model names, so you can give each reviewer its own group. In Codex, multi-agent has to be off with XCoffee, so run the three reviews one after another. Treat the result as three cheap extra passes, not as independent cross-provider review. For a high-stakes PR, run the real council with first-party keys.
sol-loop needs OpenAI's GPT Sol and its SKILL.md says plainly: "If no route runs Sol, report the blocker. Do not substitute a model." XCoffee's groups don't list Sol, so don't use sol-loop through XCoffee. Since full-send calls sol-loop, either keep a separate OpenAI route for that step or use full-stop and review the PR yourself.
Limits and gotchas on a budget relay
- 4 concurrent requests per account. That limit is shared across all keys, groups and protocols. A council round starts three reviewers at once, which leaves one slot for the main agent. Run a council next to another agent session and requests will queue or fail. Run one council at a time.
- 120 requests per minute, plus a rolling hourly cap of 100 credits and 4,000 weighted request units. A request with over 32,000 input tokens (cache included) counts as 2 units, and over 80,000 counts as 4. Long CI logs add up quickly in
babysit-pr, so trim logs before you feed them in. - Output caps per request: by default 32,768 tokens for Coding and Default, 16,384 for Fast, and the default output budget is 4,096 unless you ask for more. Don't ask a Fast subagent for a giant rewrite.
- Failover only happens before output starts. A stream that breaks mid-answer is a failure your agent has to retry.
- Not supported: stored sessions, uploaded file APIs and provider-hosted tools. That is why web search is off in Claude Code and Codex, and multi-agent is off in Codex. Client-side tools such as file edits, shell and
ghwork fine. - Daily reset at UTC 00:00. Unused credits don't roll over.
- Pinned model features. Some Claude Code features assume Anthropic's own models. If something misbehaves, try the same task on first-party Claude before blaming the skill.
tstack FAQ
What is tstack?
An open-source pack of Claude Code and Codex agent skills from Theo (t3dotgg), released October 24, 2026. It focuses on the PR lifecycle: filing, babysitting, reviewing and merging.
Does tstack need T3 Code?
No. Some skills use T3 Code features like delegate_task when available and fall back to native tools otherwise. council and sol-loop need a harness that can run subagents on the right models.
How do I use a custom base URL in Claude Code?
Set ANTHROPIC_BASE_URL (for XCoffee, https://api.xcoff.ee with no /v1) and ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN, then set ANTHROPIC_MODEL to a model name your gateway accepts, such as xcoffee-coding. With XCoffee, also set CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1, or requests fail with 400 Server-managed tools unavailable.
What's a cheap API for Claude Code?
Choose an Anthropic-compatible gateway that supports the client features you need. XCoffee offers 30-day passes with plan-specific credit allowances, one key for the matching OpenAI, Anthropic and Gemini endpoints, and automatic routing within task groups. See the current plans for prices and limits.
Can XCoffee run a true three-provider council?
No. You can't pin a specific model inside a group. Use the three groups as a budget council and say so in the PR.
Can I run sol-loop through XCoffee?
No. It requires GPT Sol, and the skill tells the agent not to substitute.
Do tstack skills use a lot of tokens?
The long-running ones do. babysit-pr and full-send read CI logs and review-bot comments again and again, which is mostly long input. Trim logs and send triage to a fast group.
Does web search work in Claude Code or Codex through XCoffee?
No. Web search is a server-managed tool, and XCoffee doesn't support provider-hosted tools. Codex's multi-agent feature also has to be off ([features] multi_agent = false). File edits, shell and gh, which tstack relies on, work normally.
Which Codex settings does XCoffee need?
In Codex 0.164: wire_api = "responses" (chat is no longer supported), base_url = "https://api.xcoff.ee/v1", env_key = "XCOFFEE_API_KEY", plus web_search = "disabled" and multi_agent = false. Leave either out and you get unsupported_tool.
Run tstack without watching a meter
tstack makes the agent do the boring parts of shipping: babysitting CI, writing the PR and reworking old branches. If you prefer a predictable subscription budget for those loops, compare XCoffee’s current plans and choose the allowance that fits your workload. Daily credits and rolling hourly caps still apply. Setup details are in the XCoffee guide, and if you are still comparing options, see our free LLM API roundup. Keep your first-party keys for the reviews that really matter.