Skip to main content
Gas Town is a multi-agent workspace manager. It runs other people’s coding agents in terminal sessions, gives each one a role and a piece of work, and merges what comes back. Gas City is the same machinery taken apart into an SDK for building your own arrangement of it. Neither is a coding agent, so almost none of the configuration on this page is theirs. Read coding harnesses first.

​
No wire, no base URL, no key

Gas Town talks to no model. It has no base URL setting and no API key field of its own, and its stated design principle is that it orchestrates agents through terminal sessions and environment variables rather than by linking against them. Its own words for it are that integration is configuration rather than compilation. So the whole of pointing Gas Town here is: pick a child agent this API serves, configure that agent the way its own page on this site says to, and use Gas Town’s environment injection to deliver the pieces that belong in the environment. Every question about the endpoint therefore belongs to the child, including the one that catches people out. Whether the base URL wants /v1 on the end differs between them, and a town running two kinds of agent needs both answers.

​
Which child agent

Gas Town ships presets for claude, gemini, codex, kiro, cursor, auggie, amp, opencode, copilot, pi and omp. Four of those have pages here.

OpenCode

Chat completions. Documented here, and the shortest path to a working rig.

Pi

Chat completions, and the one child with an output ceiling you can set.

Codex CLI

The Responses API. Documented here.

Cursor

An editor with a CLI. Documented here.
Claude Code is the claude preset and needs an endpoint this site does not publish, so a rig running it will not reach this API. That is worth knowing because it is the preset most of Gas Town’s own documentation is written around. We have said nothing about the rest.

​
Registering an agent, and where the environment goes

A preset is JSON, and for an agent Gas Town does not ship you write it yourself at ~/gt/settings/agents.json. The field that matters here is env, a map merged into the child’s environment when Gas Town launches it:
~/gt/settings/agents.json
That variable is the one OpenCode reads through the {env:HOPSCOTCH_API_KEY} interpolation in its own opencode.json, so the provider block stays where it belongs and only the secret is delivered here. ~/gt is outside your repository, which makes this a smaller exposure than the config files the shared page warns about. It is still a plain file holding a key. Reading it from your own environment instead, and leaving env out, works wherever your child agent supports it. There is a lighter path if the only thing you are changing is the command:

​
Gas City names the endpoint once

Gas City has something Gas Town does not, and if you are pointing many agents at one endpoint it is the reason to look at it. Its city.toml carries an [upstreams] table: named endpoint presets, selected per agent, that answer the question of who serves the model separately from the question of which agent does the work. An upstream carries a base_url, a credential as api_key or auth_token, and, where it needs to, the name of the environment variable each harness expects those to arrive in. That last part is the useful bit. A base URL variable is called something different in every harness, which is exactly the mess the shared page describes, and base_url_env and api_key_env let one upstream render itself into whichever names the child wants. A credential may be written as a $VAR reference so the secret stays out of the file, and env on the upstream is a raw escape hatch merged after the rest. Set the base URL there in the form the child agent wants, which for most of the children documented here means with /v1 on the end, and use the $VAR form for the key.

​
What this does to Activity

Both of these exist to run agents in parallel without a person in the loop for each turn, which is the traffic shape the shared page’s warnings are about, at the scale they were written for. Every turn of every child is its own billed request against the same key. A town with several agents working at once, each looping, produces rows in Activity faster than anything else on these pages, and the account spend rate cap counts money reserved rather than money spent, so concurrency is the dimension that reaches it. Where a child agent lets you set a maximum output, setting it is worth more here than anywhere. Rate limits and spend controls covers the cap. Take 402 off any retry list you control. An orchestrator’s job is to keep work moving, and a 402 with the code insufficient_credit is a balance rather than a transient fault, so a supervisor that restarts a failed agent will rediscover that as fast as it can.

​
Beads is not in the way

Gas Town tracks work in Beads, a version-controlled issue database the agents read and write. It calls no model and needs no key, so it is not a second thing to configure here. The same is true of the merge queue and the health patrol: they are shell and git, and the model calls are all inside the child.

​
What we did not verify

  • That a Gas Town or Gas City session against this API succeeds. Nothing was run.
  • Whether an env entry reaches every child unchanged. We read the field and its worked examples rather than tracing each preset’s launch.
  • Whether any part of gt or gc itself calls a model. We found none in their configuration surfaces, and neither carries a base URL or a credential of its own, but we did not audit either codebase for one.
  • Anything about the gemini, kiro, auggie, amp, copilot or omp presets, including whether they can be pointed at a custom endpoint at all.
  • How stable this is. Both projects are moving quickly and Gas City describes its upstream layer as a phase of work rather than a settled surface.