From Outsourcerer
Offloads grunt work to cheaper engines (GLM, Devin, Codex) while Claude orchestrates. Auto-conserves tokens when Claude's window runs hot, with configurable driving modes.
How this skill is triggered — by the user, by Claude, or both
Slash command
/outsourcerer:outsourcererThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
**Created by**: [Alex Greenshpun](https://linkedin.com/in/alexgreensh)
evals/evals.jsonreferences/agentic-local.mdreferences/effort-and-tiers.mdreferences/jobs-and-safety.mdreferences/lanes-and-models.mdreferences/mechanism.mdreferences/model-advisory.mdreferences/parallel-and-fanout.mdreferences/second-opinion-and-parity.mdreferences/tab.mdscripts/anthropic-openai-shim.pyscripts/outsourcerer.cmdscripts/outsourcerer.ps1scripts/outsourcerer.shscripts/run-or-codex.shscripts/run-or-model.shscripts/tests/conformance.shscripts/tests/test_advise.shscripts/tests/test_agentic_local_cert.pyscripts/tests/test_autodetach.shCreated by: Alex Greenshpun
Delegate the work Claude would hand its own subagents to a cheaper or different engine instead. Model-agnostic; default glm-5.2 (a free Devin lane). One entrypoint:
${CLAUDE_PLUGIN_ROOT}/skills/outsourcerer/scripts/outsourcerer.sh <subcommand> [args]
Global flags (--provider, --cloud-ack) work anywhere, before or after the subcommand.
You drive this script in natural language; the user should never learn a flag. If you're explaining a command instead of running it, you're doing it wrong.
Every session begins by silently running brief — it returns the ready lanes, the user's live limits (e.g. claude5h=95), and a conservation call in one shot. Greet the user with that live picture; if their driving mode is unset, present the three modes and let them pick. The choice is remembered (~/.outsourcerer/mode) — never ask twice.
advise, see below), offload proactively, conserve automatically. Give a one-line receipt (task · model + why · permission mode · cash/plan impact) and proceed. Still stop for: cloud consent, a mutating/dangerous escalation, an ambiguous goal, or real cash spend.Set it with mode auto|manual|hybrid (A/B/C accepted). The user re-steers anytime in plain language ("take the wheel", "ask me first"). No mode weakens safety, consent, or permission mode.
Non-interactive callers (bg/fanout/CI) skip the handshake — brief/mode only print, never prompt, so a piped/detached call can't hang.
brief reads the real Claude 5h/7d window and the ChatGPT/Codex plan. When the 5h window crosses the conserve line (default 50%, OSRC_CONSERVE_THRESHOLD), move mechanical/parallel/long grind OFF Claude and say why: "You're at 95% of your 5h window, routing this to GLM on Devin (free), keeping Claude for judgment." Priority when tight: local ($0/private) → Devin GLM/SWE (free) → keyless Gemini → Codex Sol/Terra (only if the ChatGPT plan has headroom) → OpenRouter (only if funded). Adapt to what the user has — a Codex-only user never gets a Devin tour. Auto/Hybrid do this automatically; You-drive recommends and waits. Conservation changes ROUTING only, never quality/safety/consent.
When the user hasn't named a model, always run advise — it classifies the task, pulls live benchmark scores, and returns the best value that clears the bar. This is the model-selection brain in every mode; the mode only decides whether you ask first (auto-pilot: take the winner + proceed; you-drive/hybrid: present it and wait). Cheap ≠ dumb: glm/hy3/deepseek are capable tier (frontier capability, budget price) and get flagship-grade prompting. Details: references/model-advisory.md, references/effort-and-tiers.md.
run/explore (read-only) · research (sandboxed exec, devin/codex) · edit (auto-accept edits) · yolo (all tools, no sandbox — sparingly). Wrap any verb in bg for supervised background work (poll status), or fanout for N parallel jobs.
session = YOU supervise a delegate live (not a spectator mode). It's a persistent tmux TUI you drive programmatically: session read shows the delegate's actual work as it happens, session send "…" steers it mid-flight, session model <name> switches its model, session stop ends it. This is the one mode with a real feedback loop — headless run/bg/fanout fire-and-forget (you get progress markers or a final result, never the chance to course-correct mid-run). Prefer session for long, complex, exploratory, or high-stakes delegations where catching a wrong turn early, answering the delegate's clarifying question in the moment, or iterating with it beats one blind shot. The read→steer→read loop is exactly what watched sessions get and headless ones miss. (tmux-based → Mac/Linux; on Windows fall back to bg+status.) Full semantics + tiers: references/mechanism.md, references/jobs-and-safety.md.
--provider needed)| Alias(es) | Lane |
|---|---|
sol / terra / luna / gpt-5.5 | codex-native (ChatGPT sub); add --provider claudex to run it inside the Claude Code harness via the user's own CLIProxyAPI (offer only when doctor says claudex READY; unofficial bridge, detect-only) |
fable / opus / sonnet / haiku | claude-native (Claude sub) |
gemini-pro / gemini-flash / gemini-flash-lite | gemini (agy keyless) |
glm / hy3 / deepseek / any OpenRouter id | OpenRouter (--provider cc/codex); capable tier |
swe / swe-1.7 / kimi / any Devin id | devin (default); glm/deepseek self-heal here when OpenRouter is out of credits |
ollama:<m> / local:<m> | local (keyless, PRIVATE, $0) |
gpt-image / nano-banana | image only (image subcommand) |
| any model the USER configured | droid / cursor (`--provider droid |
Full table + image backend order: references/lanes-and-models.md.
consent grant (or --cloud-ack) — remembered in ~/.outsourcerer/cloud-consent. Never retry a gate refusal blindly; the error names the fix.references/tab.md.~/.outsourcerer. The error spells out the two fixes (allowWrite, or OSRC_HOME=). Don't rerun unchanged.outsourcerer.cmd/.ps1 launch it. Only tmux session is unavailable (bg/fanout cover it).Core: brief · mode · consent · run/explore/research/edit/yolo · bg/fanout (+ status/watch/result/cancel) · advise · doctor · models. More: suggest/deals (cheap now) · estimate (cost quote) · tab (the ledger) · second-opinion · image · continue · tap (capture live limits without token-optimizer) · parity/parity-codex/parity-droid/parity-cursor (two-way bridges) · cleanup/gc. Failure states map to one-line user messages (done/done?/blocked/timeout/wedged): references/jobs-and-safety.md.
mechanism.md — setup, provider flags, verb/tier semantics, session modelanes-and-models.md — full alias/lane table, native lanes, image backendsjobs-and-safety.md — background jobs, watchdog, exit codes, failure messagesparallel-and-fanout.md — fanout, gauntlet recipe, self-healeffort-and-tiers.md — --effort per lane, capability-vs-price tierstab.md — cash vs plan-limit accounting, estimatesmodel-advisory.md — advise: classification, scoring, thresholds, fallbacksecond-opinion-and-parity.md — second-opinion, parity, install paths, env varsnpx claudepluginhub alexgreensh/outsourcerer --plugin outsourcererDelegates implementation and review tasks to external AI CLI tools (Codex, Gemini) with cross-model adversarial review for cost savings and improved accuracy.
Spawns cheap headless Claude Code workers on cheaper models (GLM/Sonnet/Haiku) with full tool harness and isolated config, fanned out by an Opus orchestrator. Workers produce branches for review.
Decides when and how to delegate tasks to Gemini or Qwen with quota tracking and error handling. Use for token-heavy operations or when tasks exceed context window.