Coding Agents
Configure coding clients for Tokamak API routing, server metering and usage insights
Start with Install and connect to choose your deployment, install its CLI and authenticate. Examples below use tokamak; use tokamak-stag on staging or tokamak-dev on development. Choose a model that supports the client's API dialect, streaming and required tool/image/reasoning features. Server API requests use the credential's organization and existing metering/governance controls.
Each client below has a complete guide covering launch, configuration, model selection, usage and recovery. For shared connection and runtime issues, see Troubleshooting agents.
Client integrations
| Client | Tool name | Configuration |
|---|---|---|
| Claude Code | claude | Private per-run Messages settings overlay |
| Codex CLI | codex | Responses overrides in the real Codex home |
| OpenCode | opencode | Inline Chat Completions provider overlay |
| DeepSeek Harness | deepseek-harness | Private settings copy and final patch |
| Copilot CLI | copilot | Responses provider environment; GitHub auth retained |
| Droid | droid | Custom model, environment key reference and backup |
| Oh My Pi | oh-my-pi | Responses registry and temporary model-role overlay |
| Jan Agent | jan-agent | Built-in tokamak provider supplied at run time; Jan's own configuration unchanged |
| Codex App | codex-app | Persistent Responses provider, bound credential helper and restore |
| Claude Cowork | claude-cowork | Persistent Messages gateway profile, /v1/models discovery and restore |
| Legacy Pi | pi | Responses provider/defaults in the native agent home |
The launcher can install supported CLI packages when missing; Oh My Pi, Jan Agent and desktop apps must already be installed. Native permissions and hosted-account requirements still apply. The maintainer launch reference records implementation contracts and validation evidence.
Models and launch modes
Ordinary tokamak launch claude uses Claude's native default through Tokamak and server-enabled Claude Pool routing. It does not apply saved Tokamak model or compaction overrides. Other clients use the saved model when required; Codex CLI can retain its own selection if no model is configured. Login no longer picks a default: the first interactive launch of an agent that needs one asks and saves the choice. The catalog picker groups models by vendor and shows each model's context window.
tokamak launch claude --model <catalog-id>
tokamak launch codex --model # Interactive catalog picker
tokamak launch claude --model-selection # Claude-specific tier selectionClaude alone supports --native, which uses local authentication outside Tokamak metering and clears inherited Tokamak-managed environment settings. Extra native arguments belong after --. Explicit client provider/model selections and independently authenticated plugins can use other routes.
Offline inspection and desktop recovery
tokamak launch codex --model <catalog-id> --check
tokamak launch codex-app --restore
tokamak launch claude-cowork --restore--check performs no network, install, update, settings write or client execution. It does not certify live compatibility. Desktop setup persists until restored and requires fully reopening the app to take effect; Tokamak does not terminate active tasks. Restore rejects unknown edits. Treat desktop configuration and journals as credentials.
Cowork uses tokamak launch claude-cowork without --model; use
tokamak launch claude-cowork --check for offline diagnostics. Claude discovers
models from the configured Tokamak endpoint's authenticated /v1/models response
and shows recognizable Claude IDs backed by compatible Claude routes. Tokamak's
catalog does not publish Claude Desktop's optional family-tier override.
The launcher removes its previous fixed list and context preference. Fully quit
and reopen Claude, then select a discovered model or start a new task if an old
conversation retains an unavailable model. An empty eligible catalog needs a
Claude-capable server route; the launcher does not invent a fallback.
Usage and governance
tokamak usage reads server-recorded usage. The Tools tab compares routed request, token and cost shares, active credential owners, reliability and attribution coverage. Supported client OTLP is separate activity, not another charge. Missing activity does not prove zero tool use. Copilot inference may remain unknown even when its OTLP is labelled. Client labels never grant authority or set prices.
To label a run for Analytics, add --tag to any launch: tokamak launch claude --tag exp-42 --tag team:robotics, or a comma list such as --tag=exp-42,team:robotics. Without --tag, the TOKAMAK_TAG environment variable (a comma list) applies, then the defaults saved with tokamak config set-tag exp-42 team:robotics (tokamak config clear-tag removes them). Tags are lower-cased and must match [a-z0-9][a-z0-9._:/@-]{0,63}; at most 16 are kept and any other value is dropped with a warning, never failing the launch. Analytics breakdowns can group usage by Tag, and a request with two tags counts under both. Tags are labels only, never access, routing or pricing inputs.
Removed pool/issue/tracking options fail before the client separator; Claude Pool is a server credential source, not a client pool product. See Bring Your Own Client for SDK setup and API boundaries.
Windows validation status
Tokamak resolves Windows npm batch shims to avoid error 193 and locates the
Codex desktop package even when its label is ChatGPT. Droid help/version and
exec keep their native command order; repeated settings normalization is
supported. Claude JSON reformatting no longer blocks profile refresh or restore.
See the tested-version matrix
for actual startup/API outcomes. Server model compatibility, output
bounds, available credit and client account requirements remain prerequisites.