Outsourcerer
Turns the AI subscriptions and local models you already own into one orchestrated team, with a single supervising session and a real receipt.
- Source repo
- alexgreensh/outsourcerer
- Stars
- ★ 199
- Last updated
- today
- License
- NOASSERTION
- Primary language
- Shell
- FA score
- 67/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeChatGPT · OpenAI API · Claude API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer with both Claude Code and Codex subscriptions who wants the frontier model reserved for judgment calls while a budget lane runs tests and mechanical refactors.
- Not a fit if
- Larger companies expecting ordinary open-source commercial terms
- Users who only want model swapping inside one harness
- Website
- alexgreenshpun.com ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v0.17.6 (2026-11-20)
- Source review
- 67/100 · Some gaps
What does this agent do, and when should you use it?
Outsourcerer is a self-contained bash orchestrator that installs into the agent harnesses you already run — Claude Code, Codex, Antigravity, Devin, Cursor, Droid, Hermes, Cline, Pi, Grok Bot. The session you talk to becomes the lead: it splits work, picks a lane and model from live benchmark scores, and dispatches through the local CLIs (claude, codex, devin, agy). Long jobs live on disk as a task graph via plan and flow, with prebuilt shapes such as council-build, spec-build and best-of-n, git-worktree isolation, and a serialized merge gate that only lands a change after your tests pass on the merged result. Cost is tracked in a local JSONL ledger and shown by tab in two honest columns: real cash on API lanes and finite subscription-window consumption. Security defaults are concrete: keyless on subscription lanes, exactly one key read from ~/.env when a paid lane needs it, and a hard block that refuses cloud dispatch if a credential file sits anywhere in the working tree. Installation is per host, e.g. the Claude Code plugin marketplace commands.
Under the hood it is a readable bash script exposing verbs: run/research/edit/yolo for read, sandboxed investigation, code change and approvals-off change; bg/status/watch/result/logs/cancel for background jobs plus a liveness watchdog; second-opinion for consensus-gated re-checks; image for text-to-image (GPT-image first, then nano-banana, then OpenRouter); tab/estimate for the ledger and a pre-flight quote; suggest/deals for live cheap and free models per platform; doctor/models for health checks and a live model list; parity, parity-codex, parity-hermes to mirror into Devin/Antigravity/Hermes or reverse-bridge into Codex/Droid/Cursor; and plan/task/flow/land for the software factory. fanout runs many specialists in parallel, each in its own git worktree and supervised to a classified end (done / blocked / timed-out). fleet lists every session with a status read from the live process and its transcript, and Mission Control (outsourcerer mc or hub serve) shows projects, running work, decisions and lane limits on one page.
- A developer with both Claude Code and Codex subscriptions who wants the frontier model reserved for judgment calls while a budget lane runs tests and mechanical refactors.
- A feature that spans days: advisors propose in parallel, a cheap model builds, another reviews against the plan, and the merge only lands after tests pass.
- Running an existing role library (e.g. agency-agents) as a squad unmodified, routing security roles to one model by name with
--route. - Sensitive code that must stay on the machine: point the local lane at Ollama or LM Studio for $0, no tokens burned, nothing leaving the box.
- Many AI sessions open at once and no single view of which is working versus waiting on your approval.
- Verifying claimed savings: separating real OpenRouter cash cost from finite Codex/Claude subscription-window usage.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Config files
~/.env
How do you install or deploy this agent?
You need bash and at least one of the CLIs you plan to drive (claude, codex, devin, agy). Install per host. Claude Code:
/plugin marketplace add alexgreensh/outsourcerer
/plugin install outsourcerer@outsourcererAntigravity:
agy plugin import claude-codeCodex, Devin and Hermes use the bridge/parity commands:
outsourcerer parity-codex
outsourcerer parity
outsourcerer parity-hermesPaid OpenRouter lanes need one key in ~/.env; subscription lanes (Claude/Codex/Antigravity/Devin) are keyless. First run executes its own health check (doctor) and tells you the exact command for anything missing.
How do you use this agent?
Once installed, drive it in plain language; the session becomes the orchestrator. On first use it shows what is ready, how much token budget is left, and asks which mode to drive in (auto-pilot, you-drive, hybrid). Core shapes:
outsourcerer plan new rate-limit --repo .
outsourcerer flow run council-build --plan rate-limit --var goal="Add per-user rate limiting" --var check="make test"
outsourcerer plan run --plan rate-limitParallel squads and receipts:
outsourcerer fanout --agents ./engineering --route 'security-*=glm-5.2, *=haiku' --task "audit auth" --worktree
outsourcerer fanout status <id> --json
outsourcerer fanout collect <id>
outsourcerer tabPrivate local lane (doctor auto-detects a running server):
outsourcerer doctorMission Control page:
outsourcerer mcWhat are this agent's strengths and limitations?
- Crosses agents, not just models: one dispatch layer drives Claude Code, Codex, Antigravity, Devin, Pi and Grok Bot harnesses.
- Durable plans on disk survive crashes and session loss;
plan runresumes, and held decisions carry defaults and deadlines. - Honest accounting: cash versus subscription-limit columns, and OpenRouter cost is read back per generation from the provider rather than trusting the harness estimate.
- Concrete security defaults: keyless subscription lanes, one scoped key read, a fail-closed credential-file block before cloud dispatch, and
:freelanes flagged as may-train. - Runs any agent library unmodified via
fanout --agents, with name-based--routeand per-specialist git worktrees for parallel edits.
- Depends on third-party CLIs being installed and authenticated; a missing one removes that lane entirely.
- Licensed PolyForm Noncommercial 1.0.0 with a written small-business carve-out (<5 people, <$20k/month); other commercial use requires a deal.
- Worktrees are explicitly not a security sandbox; lanes without OS isolation run with your user rights.
- The Grok Bot lane is labeled beta and the README says it is not yet proven on a live bot; the Cursor plugin direction is also new.
- Claude multi-account usage rests on Anthropic terms the README says it has not confirmed.
How does this agent compare with similar options?
The README contrasts itself with Claude's built-in subagents (Claude-only, no cross-session plans, no advisor consensus) and with routers/proxies such as claude-code-router and LiteLLM (model swapping inside one harness, no durable task graph, no cost ledger). It concedes server-side routing at scale to the proxy category and positions itself deliberately as a local tool.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Outsourcerer This agent | 67 · Some gaps | Agent plugin / skillFree + model costs | ★ 199 | today | Shell | Codex · Claude Code |
| Gas Town — Multi-Agent Workspace Manager | 62 · Some gaps | CLIFree + model costs | ★ 19k | 7d ago | Go | Codex · Claude Code |
| AGRO Sandbox Runtime | 76 · Good | CLIFree + model costs | ★ 43 | 4d ago | TypeScript | Codex · Claude Code |
| pitwall | 81 · Good | — | ★ 22 | today | Go | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
SECURITY.md gives concrete, specific handling: per-variable key reads from ~/.env, keys never written to argv/logs/job files, a credential hard-block before any cloud call that fails closed, local lanes sending nothing, and full disclosure before each cloud dispatch — data_flow_transparency and sensitive_data_handling score 3. However the README explicitly states worktrees are not a security sandbox and delegates run with the user's rights, so least_privilege is only 2. User confirmation exists via auto/manual/hybrid modes and --cloud-ack, but auto-pilot deliberately reduces asking, so 2. Dependency security shows SHA-pinned CI actions and a file-size gate, but no dependency manifest or vulnerability scan evidence, so 2. External effects have merge-after-test and escalation for irreversible actions, but Grok Bot branch pushes and similar are thinly specified, so 2. Rollback has preserved worktrees and cleanup refusing unmerged work, but no overall undo/recovery procedure, so 2. Source attribution is only author bylines in README/SKILL.md with no versioned provenance, so 1.
SKILL.md and README are broadly consistent on modes, escalation rules and lane priority, so self_consistency 2; but the tool depends on many external CLIs and subscription states with no version constraints or degradation matrix in-repo, so dependency_availability 1. Failure messages are described (classified done/blocked/timed-out, retry policy, loud failure reporting) but no concrete error text samples are shown, so 2.
Audience and scenarios are very broad (cost saving, parallelism, long jobs, local privacy, multi-model review), so 3; capability boundaries are stated clearly (worktrees not a sandbox, Grok Bot beta unproven, text lanes don't transfer scripts, OpenCode paid consent), so 3. Trigger precision relies on natural language plus a skill description with many aliases and env-var-dependent behavior, so 2. Environment fit claims macOS/Linux/Windows Git Bash but Windows-without-WSL details and validation evidence are thin, so 2.
README is well structured but extremely long, so information_architecture 2; install notes are scattered across badges and command examples with no standalone complete install section, so 1. Naming stability has a version badge and version-gate tests but no changelog, so 2. Examples are plentiful but there is no formal FAQ, so 2. Known limitations have dedicated passages (sandbox, beta, uncovered sessions), so 2. The PolyForm Noncommercial 1.0.0 license text is present but metadata says NOASSERTION and commercial use requires contact, so 2. Versioning has a release badge but no CHANGELOG file, so 1. Maintenance responsibility points only to an individual author and email with no governance or support commitment, so 1.
Output usability includes JSON status envelopes, receipts and a Tab ledger, but these are documented rather than demonstrated, so 2; marginal value comes from cross-subscription orchestration and cost tracking, which is differentiated but crowded, so 2; cost benefit has conserve thresholds, free-lane priority and local zero-cost lanes, but savings figures are self-asserted, so 2.
claim_traceability is weak: many performance and savings claims ("$0 cash", "500+ patterns") have no supporting data or verifiable artifacts, so 1; cross_source_corroboration is limited to SECURITY.md echoing README, with test files present but no shown results, so 1; fact_inference_separation is reasonable, as SECURITY.md separates core engine, local shim and test harness scan results, so 2.
- Worktrees are not a security sandbox: without OS isolation a delegate runs with the user's full rights and can reach files outside its worktree.
- Auto-pilot mode deliberately reduces asking; confirm its escalation boundaries (direction, cash spend, irreversible actions) match expectations.
- Heavy dependence on many external CLIs and subscription states; any missing or changed lane can affect availability, and no version constraints are in-repo.
- Cost-savings and security-scan figures in the README are self-reported and not independently verified.
- License is PolyForm Noncommercial; commercial use requires contacting the author, and metadata reports NOASSERTION.
- Grok Bot integration is labeled beta and the author states it is not yet proven on a live bot.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.17.6 Outsourcerer 0.17.6
Release notes excerpt
Fixed - Model picker reads files as UTF-8 on every system, and Mission Control's server does too. - Limit and exclude lists show up in advise again. A lane at its limit is named as parked with its quota, and excluded models are never scored. - Test suite runs on every platform again, including the new release checks. Known issue: on the GitHub macOS runner, one composer check (a typed draft read back as empty) still fails; it passes on Linux, Windows and local macOS. Tracked for 0.17.7.View on GitHub ↗ -
v0.17.5 Outsourcerer 0.17.5
Release notes excerpt
Changed - Project pages look like the build board. Done so far, release time and what waits on you up top, then the decisions, the ship steps and the work grouped by theme, all in plain words. Home and the sidebar show projects and decisions only, never lane or session lists. - Stale is labelled stale. A project whose session has not reported for 30 minutes says "Last update from its session: N h ago" and hides its old next step. A decision not re-confirmed for 2 hours says so and drops below fresh ones. Quiet projects fold away. …View on GitHub ↗ -
v0.17.3 Outsourcerer 0.17.3
Release notes excerpt
Fixed - Plans are always current. Mission Control checks your plan limits the moment you open it and every 5 minutes after, and the refresh button forces a real check. - Antigravity shows your Gemini quota. The headline reads the Gemini pool, not the Claude and GPT pool that sat at 0. - Refresh actually refreshes. It re-reads every project and redraws the page, instead of timing out quietly. - Done stays done. A project you mark shipped stays shipped in every tab and after restarts, and "running" counts only lanes that are working. …View on GitHub ↗
FAQ
Do I have to pay for model usage?
Can my keys or code leak to the cloud?
.env, id_rsa, credentials, …) at root or nested and fails closed, naming what left. A gitignored .env.local does not block the job because it runs in a clean copy of the checkout; the local lane keeps everything on your machine.What happens when a dispatched job fails?
Does it run on Windows?
Can I just offload one task without the factory machinery?
run/research dispatch, or suggest to find cheap and free models right now. The durable plans, fanout squads and advisor councils are opt-in for longer or wider work.Is Outsourcerer open source?
Is Outsourcerer safe to use?
Is Outsourcerer still maintained?
Who makes Outsourcerer?
What is Outsourcerer's official website?
Thanks — we use this to improve the page.