Orchemax

Orchemax vs orchestrators, gateways and AI FinOps tools

Orchemax 0.9.0.

Orchemax is an agent orchestrator: it directs the coding CLIs you already run, routes their model calls through a gateway you control, and measures what they cost. So it is compared in three places: against multi-agent coding orchestrators, against LLM gateways, and against usage and cost tools. Primary peer: Traycer. Adjacent runtime (not in the matrices): Herdr — agent terminal life; Orchemax is workshop law underneath. We sell governed workshops, not another agent IDE.

Start free Deep dive: vs Traycer

Where each tool sits

Three kinds of tool answer three different questions. Orchemax is the first kind, with parts of the other two.

Multi-agent coding orchestrators

Run several coding agents at once and keep them from colliding. Reach for one when the problem is many agents, many tasks or many machines.

LLM gateways

Sit between an agent and its model provider. Reach for one when you want a single connection, key handling, failover and caching.

Usage and cost tools

Read what was spent. Reach for one when the question is who spent what, not how the work gets done.

Orchestrators

Eight multi-agent coding orchestrators against Orchemax. Every cell is a fact from the peer’s own page, or Not documented.

Swipe sideways to read every column.

Capability Traycer Paperclip Gas Town Superset Multica claude-squad Conductor Ruflo Orchemax
Agent CLIs it runs Free plan “works with any coding agent”; names Claude Code, Codex, Cursor and OpenCode Claude Code, Codex, Cursor, Gemini CLI, Pi, Grok Build and Kimi Code, plus custom process and HTTP adapters Presets for claude, gemini, codex, kiro, cursor, auggie, amp, opencode, copilot, pi and omp Many CLIs listed as fully supported (Claude Code, Codex, Cursor Agent, Gemini CLI, OpenCode and more); any other CLI agent works without configuration Claude Code, Codex, Cursor, Copilot, Kimi, OpenCode and more, run by a daemon on your machine Claude Code, Codex, Gemini and other local agents including Aider Claude Code, Codex, Cursor and OpenCode, with your own subscriptions and keys Described as a meta-harness for Claude Code and Codex Presets for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Copilot, Aider and Pi, with auto-approve set per preset; a CLI it has never named still runs
Worktree isolation Not documented Optional isolated execution workspaces (git worktrees, operator branches) Each hook is a git worktree that holds an agent’s persistent work A git worktree with its own branch and terminals per task; the README says worktrees “do not sandbox processes or prevent merge conflicts” Not documented Each task gets its own isolated git workspace Local workspaces on your Mac and, on paid plans, a cloud microVM sandbox per agent; git worktrees are not named Not documented A worktree per worker task; the Chair promotes it only after the seal passes
Shared gates before merge Not documented Review and approval stages on tasks are enforced; checks on the code before a merge are not documented The Refinery runs verification gates on the merged stack in a bisecting merge queue and isolates the failing merge request The first-task walkthrough has you inspect the patch and run the tests yourself; an automatic merge gate is not documented Review gates: work lands in review, not in main, and you decide what ships; automatic checks on the code are not documented You review changes before applying them; an automatic gate is not documented Not documented A Team Gateway Checklist doc lists before-merge gates; what the gates run is not stated on the page Sealing runs your own rules from one manifest and the built-in guards on every worker task, and git hooks and the IDE hook run them too; a failed seal makes promote refuse the task; there is no merge queue
Rules and lessons at the first turn Not documented Skills are loaded on each run; sealed rules and lessons at the first turn are not documented Startup hooks inject mail and run gt prime for context recovery; sealed rules and lessons at the first turn are not documented Agents come with built-in skills; sealed rules and lessons at the first turn are not documented Skills are “a playbook every agent reuses”; sealed rules and lessons at the first turn are not documented Not documented Not documented Agents “learn from every task, remember across sessions”; sealed rules at the first turn are not documented A sealed core knowledge pack and the project’s lessons go into every agent’s brief at its first turn; lessons are consulted again before an action
Coordinator across machines Remote hosts and agent-to-agent communication on the Free plan; one coordinator with succession is not documented Remote access over a private-network bind such as Tailscale; a coordinator across machines is not documented Wasteland federates Gas Towns through DoltHub; for the tmux-backed roles, native Windows is a minimal CLI-only environment and WSL is advised Connect another machine to reach its workspaces from the desktop app, the CLI or your phone A daemon on each machine registers as a runtime, and agents are assigned to a runtime Not documented Cloud workspaces on paid plans, an iOS app and an API; coordination across your own machines is not documented Federation lets agents on different machines exchange work, with mTLS and ed25519 identity One coordinator lease per account, ordered succession when the holder goes away (orch coord); Windows, Linux and macOS Chairs measured in our own build
Presence and read receipts Not documented Heartbeats wake agents and state changes are recorded as activity; presence and read receipts are not documented Mailboxes, identities and a Witness that detects stuck agents; read receipts are not documented A sidebar tracks each agent with working indicators, completion chimes and dock badges; read receipts are not documented Not documented Not documented Shared workspaces show who is active; read receipts are not documented The console tracks workflows, agents and cost; presence and read receipts are not documented Bus messages carry delivered and read-at receipts (orch peer status); orch machines lists each machine online or offline with its OS and seats; a waiting seat is woken at once
Junk guard Not documented Not documented Not documented Not documented Not documented Not documented Not documented Not documented A guard names compiler and IDE leftovers (.o, .dll, .pyc, numbered backups) staged or present beside shared source; orch doctor --junk lists what each seat left behind and, with --fix, removes what ended sessions left after bundling unmerged branches
Cost read-out No token metering on Free; Traycer Credits bill the models run through Traycer; a read-out of your own agents’ spend is not documented Token and cost tracking by company, agent, project, goal, issue, provider and model, with budget hard stops scheduler.max_polecats caps concurrency, not spend; OTLP telemetry is exported; a spend read-out is not documented Not documented Usage analytics show what each run cost, per agent and per issue Not documented Not documented A cost-tracker plugin with token usage, budgets and cost alerts orch usage reads tokens by session, worker and model, plan-billed seats included; orch gateway usage --by adds a daily rollup per provider or seat
License and price Free $0 (bring your own agent, no token metering); Sync $16 per user per month billed yearly; Traycer Credits pay as you go MIT, self-hosted; Paperclip Cloud is a waitlist MIT Elastic License 2.0; the desktop app is “free forever”, mobile needs Superset Pro Source-available: Apache 2.0 plus extra conditions AGPL-3.0 Free $0; Pro $50 per month; Teams $60 per user per month; Enterprise custom MIT Free $0, Starter $10, Professional $25 and Team $35 per seat (3-seat minimum), each per month; model spend is bring-your-own

LLM gateways

Five gateways against the gateway inside Orchemax, on what Claude Code needs from one.

Swipe sideways to read every column.

Capability LiteLLM Portkey OpenRouter Helicone Cloudflare AI Gateway Orchemax
Claude Code through it Tutorial: Claude Code to the LiteLLM proxy through ANTHROPIC_BASE_URL, with usage tracking and cost controls Integration page: Claude Code through one ANTHROPIC_BASE_URL to Anthropic, Amazon Bedrock or Google Vertex AI Integration page: ANTHROPIC_BASE_URL set to OpenRouter, with its key as the auth token A Claude Code page logs your sessions through Helicone; it is marked maintained but no longer actively developed Integration page: Claude Code against the gateway’s Anthropic endpoint, which exposes /v1/messages POST /v1/messages on the local gateway: a provider that speaks Anthropic Messages gets the request bytes untouched, any other is translated over the OpenAI wire
Your own provider keys Your provider keys sit in the proxy config; a Claude Code page covers bring-your-own-key Claude Code names a provider slug such as @anthropic-prod, @bedrock-prod or @vertex-prod; providers are managed in the Model Catalog Bring your own key: $25,000 a month of list-price inference with no fee, then 5 percent (Standard and Business) Bring your own key is listed as supported in the gateway overview Store your Anthropic key as a provider key, or use Unified Billing credits (Cloudflare bills credits, Anthropic bills your key) Keys stay on your machine on Free, Starter and Professional; on Team they are parked on your account and each machine leases a copy for fifteen minutes
A plan login routed through it A page covers Claude Code signed in with a Claude Pro, Max, Team or Enterprise subscription, routed through the gateway The Anthropic Direct route “includes Max plan and OAuth support” Not documented Not documented Not documented Plan-billed seats keep their own login and send no model traffic through Orchemax; the gateway answers 403 to an Anthropic plan token
Failover and how visible it is Router retry and fallback logic across deployments; how visible a failover is, is not documented A fallback strategy across Anthropic, Bedrock and Vertex AI targets Requests fail over between multiple Anthropic providers “Automatic failover across providers” Model fallback when a request errors or times out (the overview page also lists request retry); the cf-aig-step response header names the step that answered Only to another Anthropic-compatible provider you configured, and never silent: an X-Orchemax-Failover header, a log line and the attempt ledger; ai.failover_same_vendor_only keeps a prompt with one vendor
Caching and savings A Claude Code page on prompt-cache routing Simple and semantic caching on the Production plan Prompt caching on Standard and above Caching listed among the gateway features Caching “for faster requests and cost savings” orch gateway usage shows crush savings and cache hits apart, as bytes kept off the wire
Spend view Virtual keys with per-key, team and user budgets, cost tracking and an admin UI in the open-source proxy; the Enterprise page also lists spend by key, user, team and org, and tag-based spend tracking Spend per team, project or developer, with hard budget limits Activity logs and export on every plan; budgets and spend controls on Standard and above Usage, costs and performance across providers in one dashboard A cost metric estimated from the tokens sent and received, with custom costs that override public prices orch usage and orch gateway usage --by read tokens, failovers and billed USD per provider or seat on your machine; the account server rolls usage up by provider and seat role
Hosting Open-source proxy you host; Enterprise runs in your own infrastructure or air-gapped An open-source version you host yourself; Enterprise adds private cloud deployment and VPC hosting; Developer and Production record logs with set retention Requests go to openrouter.ai, with eu.openrouter.ai and us.openrouter.ai for in-region routing on Business and Enterprise; a self-hosted option is not listed Gateway paid with credits at 0 percent markup; open source, with self-hosting guides for Docker and Kubernetes A Cloudflare service, available on all plans Runs on your machine inside the orch CLI; Team keys live on the Orchemax server
Price Open source; Enterprise is custom-priced (SSO, SCIM, audit logs) Developer free (10k recorded logs a month); Production $49 a month; Enterprise custom; an open-source version you host yourself Inference at provider list price; platform fee 5.5 percent on Standard or 8 percent on Business Hobby free (10,000 requests); Pro $79 a month; Team $799 a month; Enterprise by contact Core features free on all plans (analytics, caching, rate limiting); log storage has plan limits Free $0, Starter $10, Professional $25 and Team $35 per seat (3-seat minimum), each per month; model spend is bring-your-own

Usage and cost

Three tools that read what coding agents spend, against what Orchemax reads.

Swipe sideways to read every column.

Capability Vantage LangWatch ccusage Orchemax
Sources it reads Billing data from cloud, AI (Anthropic, OpenAI, Cursor, OpenRouter and others) and SaaS integrations Claude Code, Codex, Gemini CLI and opencode, through npx langwatch claude Local usage logs of Claude Code, Codex, OpenCode, Amp, Gemini CLI, Copilot CLI and more The agents it runs: their transcripts, plus every request through its own gateway
Cost per session Not documented Per-session cost, with cache reads and writes as separate token classes A session report with estimated cost orch usage by session, worker and model
View against a plan Cost Reports, Budgets and LLM Token Allocation are listed features; a view by coding-agent plan is not documented A “theoretical vs billed” view prices your sessions at API list prices next to what a Max plan costs you Claude Code 5-hour billing blocks (ccusage blocks) Plan-billed seats are included in orch usage; daily token caps (budget.workspace_tokens_per_day, seat_tokens_per_day) deny a dispatch once reached
Savings Cost recommendations and Autopilot for AWS Savings Plans; savings from a gateway are not documented Cache health per session; a savings figure is not documented Cache creation and cache read tokens are tracked separately; a savings figure is not documented Crush savings and cache hits are shown apart, as bytes kept off the wire
Where the data lives Self-service plans with sign-up and sign-in on vantage.sh; a self-hosted option is not listed Sessions land in your LangWatch project Local: reads the logs on your machine without uploading your data The runtime is local; a linked account receives counters and events, never code
Price Starter free up to $2,500 of tracked cloud spend and 3 users; Pro $30 a month; Business $200 a month; Enterprise custom Free for individual use; the post covers no other plan MIT license; no paid plan is listed Free $0, Starter $10, Professional $25 and Team $35 per seat (3-seat minimum), each per month; model spend is bring-your-own

Three lanes explained

Where a coding agent’s model traffic goes decides what Orchemax can see and measure. There are three lanes, and a project can use all three.

The agent’s own account

The agent keeps its plan login (Claude Pro or Max, Codex, Cursor and similar). Orchemax adds the rules, gates and bus around it but carries no model traffic, and its gateway refuses an Anthropic plan token.

Your own providers

The agent uses an API key of yours straight against the provider. Orchemax is not in the path, and the provider bills you. The gateway is optional.

The Orchemax gateway

The agent authenticates with an Orchemax virtual key and its calls go through the gateway, with your provider keys behind it. You get one connection, key rotation, crush, visible failover and savings that are measured.

Where they beat us

The rows where a peer is ahead today.

  • Merge throughput. Gas Town’s Refinery is a bisecting merge queue that runs verification gates and isolates the failing merge request. Orchemax has no merge queue.
  • Hosted breadth. OpenRouter fronts 500+ models across 80+ providers from a hosted dashboard. Orchemax routes the providers you configure and hosts no catalog.
  • Enterprise identity and audit. LiteLLM Enterprise and Portkey Enterprise list SSO and role-based access, and LiteLLM adds SCIM and audit logs. Orchemax has no SSO or SCIM.
  • Plan logins through a gateway. LiteLLM and Portkey document routing a Claude plan login through the gateway. Orchemax keeps plan seats on their own login and refuses an Anthropic plan token.
  • Apps on desktop and phone. Superset and Conductor ship desktop apps with an iOS companion, and Traycer a mobile app. Orchemax is a CLI plus a web console.
  • Spend detail. Paperclip tracks cost by goal, issue, provider and model with budget hard stops, Multica shows what each run cost per agent and issue, and ccusage reads many agents’ local logs for free. Orchemax caps tokens per day per workspace or seat at dispatch.

Choose X if…

For orchestrators and gateways. The usage and cost tools sit beside either.

Orchestrators

  • Gas Town if merge throughput across many agents is the bottleneck.
  • Paperclip if budgets, org charts and approvals above the agents matter most.
  • Traycer if planning, tickets and agent-to-agent chat in a desktop and mobile app are the job.
  • Superset or Conductor if a polished desktop app with a workspace per task is the job.
  • Multica if you want to assign issues to agents like teammates and see what each run cost.
  • claude-squad if parallel sessions in isolated git workspaces, in a terminal, is all you need.
  • Ruflo if swarms, shared memory and federation around Claude Code and Codex is the job.
  • Orchemax if the CLI you already opened should stay the CLI you open, with gates, a sealed knowledge pack and a coordinator underneath.
  • Herdr if surviving disconnects and finding the blocked pane is the job, then add Orchemax under the CLIs.

Gateways

  • LiteLLM if you want an open-source gateway you host and own.
  • Portkey if you want a gateway with observability and guardrails, hosted or self-hosted.
  • OpenRouter if one hosted endpoint over hundreds of models is the job.
  • Helicone if open-source observability with a gateway is the job.
  • Cloudflare AI Gateway if you already run on Cloudflare and want a gateway from its dashboard.
  • Orchemax if the gateway should sit beside the gates, the lessons and the coordinator, with savings measured.

Works alongside your code scanners; Orchemax does not replace them.

How the cells were made

  • Every cell was read on the peer’s own page on 2026-10-21. The page it cites is in the HTML next to the cell and in the list below.
  • Not documented means we did not find it on the pages listed. It does not mean the peer lacks it.
  • Orchemax cells describe what the 0.9.0 build ships and cite our own docs.
  • Prices and plans change. Re-read the peer’s page before you decide.

Pages read

Start free — no credit card Compare plans Read the docs