What Is Ultracode in Claude Code?
Ultracode is an effort setting in Claude Code, Anthropic's terminal coding agent. When it's on, Claude automatically plans and runs multi-agent workflows — fanning out parallel subagents to explore, implement, and verify — instead of handling your request in a single pass. It's also a keyword: type ultracode in a prompt to run just that one task as a workflow.
Ultracode is the user-facing switch for dynamic workflows, a Claude Code capability where Claude writes a JavaScript orchestration script, and a background runtime executes it. The script can spawn subagents, run them in parallel, and hold intermediate results in variables — so the messy middle never touches your context window. Only the final answer flows back.
On this page
The two meanings of "ultracode"
The main source of confusion: "ultracode" names two different switches that share one engine.
| Keyword | Effort setting | |
|---|---|---|
| How you invoke it | Type ultracode inside your prompt |
/effort ultracode, Tab toggle in /effort, or launch with claude --effort ultracode |
| Scope | That single task only | The whole session |
| Who decides to orchestrate | You did — the task runs as a workflow | Claude decides per task |
| Good for | Occasional heavy tasks | Long sessions of audit-grade work |
One request can even trigger several workflows in sequence — for example, one to understand the codebase, then another to implement the change, then a third to verify it.
How it works under the hood
When Claude decides a task warrants orchestration, it writes a workflow script — real JavaScript, executed by a background runtime while your session stays responsive. The script gets a small API:
// Claude writes and runs this for you
const findings = await parallel(
files.map(f => agent(`Audit ${f} for injection risks`))
);
const report = await agent(
`Synthesize into one prioritized list`,
{ schema: { priority: "number", issue: "string" } }
);
log(`Found ${report.length} issues`);agent()spawns one subagent with a fresh context window (structured JSON output supported)parallel()andpipeline()fan out one agent per list item, concurrently or in stagesphase()groups agents in the progress view; script variables hold intermediate results
The key property is context economics: a hundred agents can read a hundred files, but your main conversation only receives the distilled result. Effort levels scale how deep Claude thinks in one context; workflows scale how broad it works across many.
You can watch any run in the /workflows progress view — pause it, restart individual agents, or inspect per-agent token usage.
How big can workflows get?
- Concurrency: 16 agents at once by default (1–256 via
CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS) - Ceiling: 1,000 agents per run; 4,096 items per
parallel()/pipeline()call - Guardrail: a warning appears past 25 agents or a projected 1.5M tokens — suppressed while the Ultracode effort setting is on
- Size guideline:
small(<5 agents),medium(<10),large(<50),unrestricted— default issmallon Pro plans
Requirements
- Plans: all paid Claude plans (Pro, Max, Team, Enterprise), plus API access via Anthropic, Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry
- Pro extra step: enable the "Dynamic workflows" row in
/config - Version: any recent Claude Code; v2.1.284 (Sept 2026) made effort and orchestration independent dials — before that,
/effort ultracodeforcedxhighreasoning
The keyword only fires on input typed by a human at the prompt. It is inert from claude -p, SDK calls, scheduled tasks, and webhook/PR-comment payloads — a deliberate safety line so automated pipelines don't accidentally fan out hundreds of agents.
What it costs
There's no separate Ultracode fee — you pay in usage:
- Anthropic's docs warn that Ultracode consumes meaningfully more tokens than normal modes, since each task may spawn many agents
- On subscriptions, usage limits drain noticeably faster; a single request can chain multiple workflows
- A real published example: one pre-check workflow burned 471,480 tokens across 5 agents and 84 tool calls in ~4 minutes — and caught a wrong technical proposal a single-pass review had accepted
Mitigations: run on a small slice of the task first, check /model before big runs, request smaller models for weak stages, and watch per-agent usage in /workflows.
When to use it (and when not)
Worth it:
- Codebase-wide audits and security reviews
- Large migrations and refactors touching many files
- Cross-checked research where one agent's claims get verified by another
- Synthesizing long threads — feedback piles, feature requests
Skip it:
- Single-file edits and quick questions (pure overhead)
- Tasks that need your input mid-run — workflows can't pause to ask
- Open-ended writing (PRDs, docs) where breadth doesn't help
Reported at the extreme end: teams running Ultracode sessions around the clock on multi-week rewrites — one widely shared example rewrote on the order of a million lines over eleven days. That's the ceiling, not the norm; most valuable runs use 3–10 agents.
FAQ
Is Ultracode free?
No separate fee, but it's limited to paid plans and draws down usage limits faster. See What it costs.
Does Ultracode force xhigh reasoning?
Not anymore. Since Claude Code v2.1.284, effort and orchestration are independent dials. Exception: the launch flag claude --effort ultracode still sets effort to xhigh.
Can I turn the keyword trigger off?
Yes — disable "Ultracode keyword trigger" in /config, or disable workflows entirely with "disableWorkflows": true in settings or CLAUDE_CODE_DISABLE_WORKFLOWS=1.
Does it work on Opus 4.6 or Sonnet 4.6?
The Ultracode toggle only appears for xhigh-capable models, which excludes Opus 4.6 and Sonnet 4.6. Use a current-generation model.