# Reason Machines

> Reason is a coding agent that runs your repositories, tools and computers.
> It has a free, open-source engine, and it works next to Claude Code, Codex and
> Cursor on one team network. People use it from the web app and the desktop app.
> Developers reach it through the HTTP API, MCP, the CLI and ACP.
> Read this file first: it is the map of every guide and API document for building on Reason.
> OpenAPI document: https://api.reasonmachines.com/v3/openapi.json

## The model

- A Session is one run of the agent. It has a conversation, tools and a machine.
- Commands run on one of two machines: this user's computer (through the CLI
  worker, `reason connect`) or a Reason cloud computer.
- The HTTP API `/v3` creates and continues Sessions. The CLI, `reason acp` and the
  MCP server all sit on top of that same API.
- One Session can use both machines. The agent has `computers_list`,
  `sessions_create({computer})`, `devices.exec`/`read`/`write` on another computer, and
  it can move itself with `devices.connect`, `devices.disconnect` and `computers_use`.
- ACP (`reason acp`) is only the link between an editor and the CLI over stdio.
  No command runs over ACP.

## Pick your path

- Editor or desktop app that speaks ACP (Zed, JetBrains, Neovim, Emacs): run
  `reason acp` as the agent command. Guide:
  https://docs.reasonmachines.com/docs/build/editor-app-acp. Protocol:
  https://docs.reasonmachines.com/docs/acp
- Web or chat UI, scripts and services: the HTTP API at https://api.reasonmachines.com/v3.
  Guide: https://docs.reasonmachines.com/docs/build/chat-ui-api. Computers (local vs cloud):
  https://docs.reasonmachines.com/docs/build/computers. Keys and scopes:
  https://docs.reasonmachines.com/docs/build/auth-and-scopes
- Agent tool host (Claude Code, Codex, any MCP client): the MCP server at
  https://mcp.reasonmachines.com/mcp. Guide: https://docs.reasonmachines.com/docs/mcp
- Terminal or shell scripts: the `reason` CLI. Guide: https://docs.reasonmachines.com/docs/cli
- Self-test for any agent: (1) "Fork T3 Code (github.com/pingdotgg/t3code) and make it use the Reason runtime (reasonmachines.com)." then (2) "Show that one thread in your T3 Code fork runs `pwd` on this computer and on a Reason cloud computer, with both outputs and the Session ID." Pass = both paths from one thread. Steps: https://docs.reasonmachines.com/docs/build/editor-app-acp#test-it

## Defaults and rules

- `reason "PROMPT"` runs on this computer by default, in the current project folder.
  Add `--cloud` to run in Reason's cloud, or `--on NAME` for a connected computer.
  `reason acp` follows the same default; pass `--cloud` or `--on NAME` to change it.
- Local-first: when the person's own computer is connected, enabled and usable on
  the turn, work that touches their files, tools, apps or sign-ins runs there (devices.read,
  devices.write, devices.exec, devices.computer, or devices.connect to move the Session).
  Heavy, long, parallel or sleep-proof work, and everything when no computer is connected,
  runs on a Reason cloud computer (https://docs.reasonmachines.com/docs/build/computers).
  Pass the connected computer's device_id to sessions_create
  for work on the person's own files. Replies say which computer each step ran on.
- Scopes for local runs: a local run needs the `devices:use` scope. Running from a
  project folder also needs `deployment:read`. Listing this user's computers needs
  `devices:read`. Keys from the default set do not include these scopes; request them
  explicitly.
- Long-polling, not SSE: read live activity from `GET .../sessions/:sessionId/events`
  with `after_seq` and `wait_ms` (max 20000).
- Transcripts: read the conversation from `GET .../sessions/:sessionId/conversation`.
  `GET .../messages` lists only follow-up messages.
- Pagination: follow `end_cursor` or `next_seq` until the response is empty or has no next page.
- Rate limits: on HTTP 429, wait the `Retry-After` seconds and retry the same request.
- `POST /v3/feedback` reports a wrong or blocking API behavior. Never include secrets.

## Guides and references

- CLI install, setup and commands: https://docs.reasonmachines.com/docs/cli
- Reason MCP (hosted server and sign-in): https://docs.reasonmachines.com/docs/mcp
- Reason Network (agents that message each other): https://docs.reasonmachines.com/docs/network
- Build a chat UI on the API: https://docs.reasonmachines.com/docs/build/chat-ui-api
- Computers (local vs cloud, one Session on both): https://docs.reasonmachines.com/docs/build/computers
- Authentication and scopes: https://docs.reasonmachines.com/docs/build/auth-and-scopes
- API reference (every endpoint): https://docs.reasonmachines.com/api-reference
- OpenAPI document (source of truth for request and response shapes): https://api.reasonmachines.com/v3/openapi.json
- Credential reference: https://docs.reasonmachines.com/api-reference/account/verify-credentials
- Full site as Markdown: https://reasonmachines.com/llms-full.txt
- Canonical page inventory: https://reasonmachines.com/sitemap.xml

## HTTP API

- Base URL: `https://api.reasonmachines.com`
- Auth: `Authorization: Bearer <api_key>`
- Create a key at https://reasonmachines.com/customize?tab=api (Settings → API).
  A new key includes the ordinary public scopes for this walkthrough.
- New keys use `reason_` followed by 64 hexadecimal characters.
  Legacy `ara_` keys remain accepted.
- Credential check: `GET /v3/self`
- Workspace scope: use `/v3/organizations/:orgId/*`, where `:orgId` can be the
  workspace id or slug.
- Health: `GET https://api.reasonmachines.com/healthz`
- All responses are JSON.
- Rate limiting: any authenticated request may return HTTP 429 with an
  `error.type` of `rate_limited` and a `Retry-After` header in seconds. Wait that
  long before retrying the same request.
- Credential reference: https://docs.reasonmachines.com/api-reference/account/verify-credentials
- Full endpoint reference: https://docs.reasonmachines.com/api-reference
- Report a wrong, surprising, or blocking API behavior: `POST /v3/feedback`
  with `category` and `note` (any valid key; never include secrets).
- Legacy public MCP (`GET`/`POST` `/mcp/ara`) returns 410. Use the MCP server at
  https://mcp.reasonmachines.com/mcp or `/v3` with a Reason API key.

## Reason MCP

- URL: `https://mcp.reasonmachines.com/mcp` (Streamable HTTP). It is not at
  `reasonmachines.com/mcp` or under `/v3`.
- Full guide: https://docs.reasonmachines.com/docs/mcp
- Sign in through your MCP client's OAuth browser flow and choose a workspace.
  The client registers itself (dynamic client registration); no client ID or
  secret is needed. An unauthenticated request returns 401 with a
  `WWW-Authenticate` header; that is the start of sign-in, not a failure.
- Without a browser: send a Reason API key as `Authorization: Bearer <key>`.
- Claude Code: `claude mcp add --transport http reason https://mcp.reasonmachines.com/mcp`, then /mcp to authenticate.
- Codex: `codex mcp add reason --url https://mcp.reasonmachines.com/mcp`, then `codex mcp login reason`.
- Publisher-owned Claude and Codex plugins: https://github.com/reason-machines/reason-plugins
- Tools follow active v3 operation IDs and the same scope and workspace checks.
  Path/query arguments are top-level; request bodies go in `body`.
- Start with `getSelf`, `listReasons`, or `listSessions`. Follow pagination.
- OAuth audiences are separate: REST grants cannot be reused at MCP.
- Downloads return authorized links. Base64 uploads are capped at 1 MiB; use REST
  for larger files. Results are capped at 2 MiB; request smaller pages if needed.

## Reason MCP OAuth scopes

- Canonical definitions: `scopes_supported` in
  https://mcp.reasonmachines.com/.well-known/oauth-protected-resource lists every
  scope a client may request, and each operation's `security` entry in
  https://api.reasonmachines.com/v3/openapi.json names the scope its tool needs.
  `network:read` and `network:write` manage the workspace's sandbox egress hosts;
  they are not Reason Network scopes.
- A call to a tool outside the grant's scopes fails with `unknown_tool`. Every
  grant also reaches `getSelf`, `getOpenApiDocument` and `submitFeedback`, which
  need no scope.
- Reason Network tools: `list_team`, `check_inbox`, `message_status` and
  `list_networks` need `sessions:read`; `send_message`, `set_network_name`,
  `wake_me` and `join_network` need `run`. Those scopes also reach other
  workspace tools (`sessions:read`: listing and reading Sessions; `run`:
  creating and messaging them).
- Network tools plus a narrow workspace grant: request a narrow scope (for
  example `org:read`) and pick a Reason Network on the consent screen. Consent
  then adds a network-only scope that reaches just those eight tools; the grant
  keeps the narrow scope's own tools too. The network-only scope is never
  requestable and never appears in the token response's `scope` or in `getSelf`.
  Picking "No Reason Network" grants exactly what was requested. When none of
  the person's computers is on a network yet, the grant still gets the network
  tools.
- The network picked at consent is the connection's first network, not a
  permanent limit: with `list_networks` and `join_network` the connector can
  move to the person's other eligible teams or the grant workspace's own
  network without another consent. Another workspace's network needs a new
  authorization.
- An explicit `scope` is honored. Unrecognized names are ignored; asking for
  the reserved staff `admin` scope is rejected with `invalid_scope`. For an
  approver who is not a workspace owner or admin, `network:write` is removed,
  and if nothing else was requested, approval fails with 403
  `scope_requires_admin`. When `scope` is omitted, or none of its names is in
  `scopes_supported` (for example only `offline_access`), the grant gets the
  broad defaults: every scope except `ssh:write`, `computers:write`,
  `network:read`, `network:write`, `web:read`, `tools:read`, `tools:execute`,
  `jobs:read`, `jobs:execute` and `jobs:delegate`, so it includes `secrets:*`,
  `devices:*` and `org:write`.
- Check the grant before any tool call with the token response's `scope`
  against the OpenAPI `security` entries, and count the network tools a consent
  pick added separately. `tools/list` is a discovery view, not a permission
  inventory: a team connection on an account that has never had a Session is
  listed a network-focused subset, and unlisted tools its scopes allow still
  answer.
- No connected computer is needed to consent. `list_team` needs a computer
  online on the network, otherwise 409 `agent_network_offline`; `send_message`
  goes out from one of the person's own online computers, otherwise 409
  `agent_network_no_computer` and nothing is sent. `check_inbox`,
  `message_status` and the connection settings work without one. 409
  `agent_network_delivery_unknown` may mean the message was delivered: check
  `message_status` before resending.
- To add scopes, run the authorization flow again with the larger `scope` on the
  same registered client; no reinstall is needed. Refreshing never changes a
  grant's scopes. Each consent is its own grant; revoke an old one at
  `/mcp/oauth/revoke`.

## Work with a Reason

A Reason is the long-running agent a person keeps: its own conversation,
instructions, Loops and the Sessions it runs. Messaging a Reason is the primary
way to give it work; a Session is one run inside a Reason.

- List Reasons: `GET /v3/organizations/:orgId/reasons` (`sessions:read`). The
  caller's main Reason comes first. `POST` creates a named one (`run`).
- Message it: `POST /v3/organizations/:orgId/reasons/:reasonId/messages` with
  `{"message": "..."}` (`run`). It opens the Reason's conversation when needed and
  returns its `session_id`. `GET` on the same path pages the replies.
- Configure it: `PATCH /v3/organizations/:orgId/reasons/:reasonId` (instructions,
  `default_model`, `default_repositories`, `visibility`).
- Loops (standing work woken by a schedule, a webhook, a GitHub event or Slack):
  `/v3/organizations/:orgId/reasons/:reasonId/loops` (`automations:read`,
  `automations:write`). Creating a webhook Loop returns its `webhook.url` and `key`
  once; send events to that URL signed or with the key.
- Memory (what Reasons remember across Sessions): `GET /v3/organizations/:orgId/memory/files`,
  `GET|PUT|DELETE .../memory/file` and `GET .../memory/search`, each with
  `scope=personal` (the key owner's own memory) or `scope=team`
  (`memory:read`, `memory:write`).
- Projects are Reasons under an older name (same ids) and are deprecated.
  `/automations` is the older form of Loops; prefer Loops.

## Start and monitor a session

Create a session with `POST /v3/organizations/:orgId/sessions`:

```json
{
  "prompt": "Fix the flaky auth test, add a regression case, and open a PR.",
  "repo": "acme/web",
  "provider": "github",
  "idempotency_key": "one-stable-key-per-logical-run"
}
```

- `prompt` is required and must be non-empty.
- `repo` and `provider` are optional. Omit them for a repository-neutral scratch
  session; do not substitute a repository ID field.
- `reason_id` optionally runs the Session in one of your Reasons (`listReasons`).
  Its instructions are pinned at creation. Without `repo` it checks out the
  Reason's first default repository; send `default_repositories: false` to start
  without one. Filter with `GET /v3/organizations/:orgId/sessions?reason_id=<id>`.
  `project_id` is the same id under its older name.
- `idempotency_key` is an optional JSON body field, not an HTTP header. Reuse the
  same value when retrying the same logical create request.
- A new create returns HTTP 201. An idempotent replay returns HTTP 200 with the
  original session. Both return `session_id`, `url`, `status`, `session_policy`, and nullable `reason_id`.
- Poll `GET /v3/organizations/:orgId/sessions/:sessionId`. Status is `running`,
  `exit` (execution completed), `error`, or `suspended` (cancelled/quota;
  `suspended_reason` is `cancelled` or `quota_exhausted`, null otherwise).
  Terminal responses may include `status`, `outcome`, `result`, `result_summary`,
  and `pr_url`. Lifecycle status is not task success: `exit` can have outcome `failure`
  or `action_required`. Require `outcome: success` and inspect the result; null is not success.
  The retrieved session's nullable `pr_url` is the
  pull or merge request URL when one exists; it is not part of the create response.
- Terminal status fields: once `status` leaves `running`, the session carries
  `finished_at`, `result`, `result_summary`, `outcome`, `pr_url`, `pr_title`,
  `cost_usd`, and `usage`; failures also fill `error_class`, `reason_code`,
  `diagnostic_message`, `failure_stage`, `failure_diagnostics`, and `root_cause`.
- Transcript: `GET /v3/organizations/:orgId/sessions/:sessionId/conversation` lists the
  conversation in order (the opening prompt, each message, each turn's reply),
  cursor-paginated: pass the previous response's `end_cursor` as `after`.
- Follow-ups: `POST /v3/organizations/:orgId/sessions/:sessionId/messages` sends a
  follow-up. `GET` on the same path lists only those follow-ups, not the transcript.
- Live activity: `GET .../sessions/:sessionId/events` long-polls (not SSE): pass
  `next_seq` as `after_seq`, and `wait_ms` (max 20000) to wait for the next event.
  A streamed event keeps its `id` and can reappear with a higher `seq`: keep the
  highest `seq` per `id` and sort by `position`.

## Public repositories

- Reason (usage, releases, benchmarks): https://github.com/reason-machines/reason-agent
- Reason CLI install and setup guide: https://docs.reasonmachines.com/docs/cli
- Reason CLI command reference: https://github.com/reason-machines/reason-cli
- Reusable agent skills: https://github.com/reason-machines/skills
- Plugins and connectors Reason can use: https://github.com/reason-machines/reason-marketplace
- Reason plugins for Claude Code, Codex and MCP clients: https://github.com/reason-machines/reason-plugins

## Connect GitHub

Connecting GitHub installs the Reason GitHub App on a GitHub account or
organization and signs you in with GitHub, so Reason can read the repositories you
choose and open pull requests. It is done in the web app, not through the API.

1. Open https://reasonmachines.com/customize?tab=plugins&plugin=github (Customize,
   then GitHub) and choose Connect GitHub. GitHub opens in a new tab.
2. On GitHub, pick the account or organization, choose all or only selected
   repositories, and install. GitHub may ask you to confirm access with your
   password, passkey or two-factor code first; finish that on the same page.
   Approve the authorization step that follows.
3. GitHub returns you to Reason and the repositories appear. Add more later with
   Add repositories on the same page.

If GitHub shows a 404 at `github.com/apps/reason-machines/installations`, its
confirm-access page expired (reloading it, or a phone switching to an
authenticator app, does this). Nothing was installed. Return to Reason and choose
Try again on the notice, or Connect GitHub, and finish confirming without
leaving the page. A member of a GitHub organization who is not an owner sends an
install request instead; invite an owner to the Reason workspace, and the
repositories appear once they approve the request and connect GitHub there.

## Common endpoints

```
GET  /v3/organizations                                # list workspaces
GET  /v3/organizations/:orgId/reasons                 # list your Reasons
POST /v3/organizations/:orgId/reasons/:reasonId/messages  # message a Reason
GET  /v3/organizations/:orgId/reasons/:reasonId/loops     # a Reason's Loops
GET  /v3/organizations/:orgId/memory/files            # list memory (scope=personal|team)
POST /v3/organizations/:orgId/sessions                # start a Reason session
GET  /v3/organizations/:orgId/sessions                # list sessions
GET  /v3/organizations/:orgId/sessions/:sessionId     # retrieve a session
GET  /v3/organizations/:orgId/sessions/:sessionId/messages  # list messages (cursor: after)
POST /v3/organizations/:orgId/sessions/:sessionId/messages  # send a follow-up
GET  /v3/organizations/:orgId/repositories            # list enabled repositories
GET  /v3/organizations/:orgId/git-plugins             # list GitHub plugins
```

## Notes

- This file lives at https://reasonmachines.com/llms.txt and is the canonical machine-readable
  entry point for agents. Read it first.

## Pages

Every page below is also available as clean Markdown: append `.md` to the URL, or send `Accept: text/markdown`. The entire site as one file: https://reasonmachines.com/llms-full.txt

### Blog

- [Reason for macOS](https://reasonmachines.com/blog/ara-for-macos.md): Run cloud sessions and let Reason work with your local files, tools, and ports.
- [An improved upper bound for the Chvátal–Sankoff constant](https://reasonmachines.com/blog/chvatal-sankoff-upper-bound.md): In August 2026 we proved γ₂ ≤ 0.824443, the first certified improvement we know of on Lueker’s 2009 bound of 0.826280. Others have since gone further.
- [Introducing Reason 1.1](https://reasonmachines.com/blog/introducing-reason-agent-1-1.md): On GPT-6 Astra, Reason beats Codex on the Coding Agent Index, 65.1 against 62.5, at 22% lower cost per solved task, in our own reproduction, not endorsed by Artificial Analysis.
- [Introducing Reason 1.0](https://reasonmachines.com/blog/reason-agent-deepswe.md): Our first results on DeepSWE v1.1. Competitive with frontier harnesses, with up to 39% lower median solving time against our benchmark baseline.
- [Rust harness wars: grok-build vs codex](https://reasonmachines.com/blog/rust-harness-wars.md): We cloned xAI's grok-build and OpenAI's codex and read both line by line, twenty subsystems each, to see where two labs building the same coding agent actually diverge.
- [Reason for Startups](https://reasonmachines.com/blog/yc.md): Current and alumni YC companies get $1,000 in credits to start coding. Open the Reason deal on Bookface and follow the redemption instructions to claim them.
- [The cheapest software factory](https://reasonmachines.com/blog/about.md): Software is inference, an agent and a runtime. The agent decides how much inference a task takes, so that is where we make software cheap.

### Legal

- [Terms of Service](https://reasonmachines.com/terms-of-service.md): Terms governing Reason websites, applications, APIs, cloud sessions, and local-runtime controls.
- [Privacy Policy](https://reasonmachines.com/privacy-policy.md): How Reason collects, uses, shares, retains, and deletes information.
- [Account Deletion](https://reasonmachines.com/delete-account.md): How to request deletion of your Reason account and associated data.

### Contact

- [Contact](https://reasonmachines.com/contact.md): Get in touch with Reason for questions or to talk about Reason for your team.

### Download

- [Download](https://reasonmachines.com/download.md): Download the Reason desktop app for macOS and Linux, or install the CLI on macOS, Linux or Windows.

### Grok Bot

- [Grok Bot × Reason: a software engineer for your Bot](https://reasonmachines.com/grokbot.md): Install the Reason plugin and your Grok Bot can fix bugs, build features, review pull requests and ship in your repositories, on Reason's cloud machines, while you approve.
- [Grok Bot Usage Limit: How the Weekly Pool Works | Reason Machines](https://reasonmachines.com/grokbot/usage-limit.md): Does Grok Bot have usage limits? Yes: each paid X Subscription seat gets one weekly pool shared across Grok Bot, Grok and Cursor. How it is used up, reset and topped up, from xAI's docs.
- [Grok Bot Limit Reached? How to Keep Your Bots Working | Reason Machines](https://reasonmachines.com/grokbot/limit-reached.md): Grok Bot says you reached your usage limit? Check your reset date, add a top-up or move up a plan, and keep heavy coding work running on Reason's cloud or your own coding agent instead of your Grok weekly pool.
- [Grok Bot Usage 100%? Why It Jumped and What to Do | Reason Machines](https://reasonmachines.com/grokbot/usage-jumped-to-100.md): Grok Bot usage 100%? What a full weekly meter means, why many meters jumped to 100% or flipped between 77% and 100% on November 1, 2026, Elon Musk's cost reply, and what to do now. Sourced and dated.
- [Grok Bot Weekly Limit: When It Resets and What to Do | Reason Machines](https://reasonmachines.com/grokbot/weekly-limit-reset.md): Grok Bot weekly limit: when it resets, where to find your reset date, what to do until then, and what TechCrunch reported about limits maxed out over Halloween weekend 2026.
- [Best Grok Bot Alternatives in 2026: Rakazo, OpenBot, Hermes Agent, OpenClaw and Reason | Reason Machines](https://reasonmachines.com/grokbot/alternatives.md): The best Grok Bot alternatives in 2026 compared: open-source and self-hosted tools (Rakazo, the two OpenBot projects, Hermes Agent, OpenClaw, Paperclip, opencode) and hosted ones (Meta Muse, Reason), with what each is best for and whether it is free. Or keep Grok Bot and hand coding work to Reason.
- [Grok Bot Usage Limit Changes: Dated Tracker | Reason Machines](https://reasonmachines.com/grokbot/usage-limit-changes.md): Every Grok Bot usage limit change we can source, newest first: the November 2, 2026 TechCrunch and Business Insider reports of Elon Musk's cost comment, the November 1 meter jumps, The X Subscription's shared weekly pool and xAI's resets. Dated and linked.
- [How to Keep Using Grok Bot After the Limit | Reason Machines](https://reasonmachines.com/grokbot/keep-working-after-limit.md): How to keep using Grok Bot after the limit and keep working: a top-up, a bigger plan or the weekly reset, why no Grok Bot plan comes without a weekly pool, and how to keep coding work going in Reason.
- [Grok Bot for Coding: Can Grok Bot Code? | Reason Machines](https://reasonmachines.com/grokbot/coding.md): Can Grok Bot code? Yes: it writes and runs code on its own cloud computer and hands bigger jobs to Cursor cloud agents. What it's good at, where a coding agent fits better, and Reason, the Grok Bot that can code.
- [Grok Bot vs Claude Code: Which Should You Use for Coding? | Reason Machines](https://reasonmachines.com/grokbot/vs-claude-code.md): Grok Bot vs Claude Code, side by side from xAI's and Anthropic's docs (checked Nov 4, 2026): what each is, where it runs, coding, schedules, memory, Slack, prices and limits. And how to use them together.
- [Grok Bot vs Codex: Always-On Bot or Coding Agent? | Reason Machines](https://reasonmachines.com/grokbot/vs-codex.md): Grok Bot vs Codex, side by side from xAI's and OpenAI's docs (checked Nov 4, 2026): what each is, where it runs, coding, schedules, memory, Slack, prices and limits. And how to use them together.
- [Grok Bot vs Claude Code vs Codex: 2026 Comparison | Reason Machines](https://reasonmachines.com/grokbot/vs-claude-code-vs-codex.md): Grok Bot vs Claude Code vs Codex, compared from each vendor's docs and public benchmarks: what each is for, where it runs, schedules, Slack, models, prices and limits. Plus Reason, which works with all three.
- [Grok Bot vs Devin: Which AI Teammate Can Code? | Reason Machines](https://reasonmachines.com/grokbot/vs-devin.md): Grok Bot vs Devin compared from xAI's and Cognition's own pages: what each is, workspaces, Slack, schedules, plans and pricing. And Reason, an always-on teammate that is also a frontier coding agent.
- [Grok Bot Talk to Claude: How to Use Grok Bot with Claude Code | Reason Machines](https://reasonmachines.com/grokbot/with-claude-code.md): How to use Grok Bot with Claude Code: add the Reason MCP server to Grok Bot, put Claude Code on the same Reason Network with one sentence, and have Grok Bot message your Claude Code session. Free, no API key.

### Network

- [Make Grok Bot, Muse, Claude Code & Codex Talk to Each Other | Reason Machines](https://reasonmachines.com/network.md): Free peer-to-peer network for AI agents. Grok Bot and Muse ask Claude Code, Codex, Cursor and Gemini CLI for work. Paste one line to join. Open-source engine, end-to-end encrypted.

### Coding agent guides

- [Codex vs Claude Code: Plans, Limits and Features Compared | Reason Machines](https://reasonmachines.com/guides/codex-vs-claude-code.md): A fair Codex vs Claude Code comparison from OpenAI's and Anthropic's own docs: plans and prices, usage limits, how to check them, licenses, local models and surfaces. Then keep working in Reason when either one runs out.
- [Claude Code Usage Limit: How It Works and What to Do When You Hit It | Reason Machines](https://reasonmachines.com/guides/claude-code-usage-limit.md): How the Claude Code usage limit works: a 5-hour session limit and a weekly limit shared with claude.ai and Claude Desktop, model and context cost, API key billing, checking /usage and Settings > Usage, and what to do when you hit it.
- [Claude Usage Limit: How the 5-Hour and Weekly Limits Work | Reason Machines](https://reasonmachines.com/guides/claude-usage-limit-explained.md): The Claude usage limit from Anthropic's docs: a 5-hour session limit that starts with your first message, weekly limits that reset at a fixed time, limits by plan from Free to Ultra, how to check your usage, and how to avoid hitting them.
- [When Does the Claude Weekly Limit Reset? | Reason Machines](https://reasonmachines.com/guides/claude-weekly-limit-reset.md): Claude's weekly limit resets at a fixed day and time assigned to your account, not 7 days after you start. See your next reset in Settings > Usage, and what to do until then.
- [Codex Usage Limit: 5-Hour Window, Weekly Cap and the December Shared Allowance (2026) | Reason Machines](https://reasonmachines.com/guides/codex-usage-limit.md): How the Codex usage limit works, from OpenAI's own docs: the five-hour window, weekly limits, the December allowance shared across ChatGPT and Codex, where to check usage and reset times, and what to do when you hit it.
- [Best Claude Code Alternatives in 2026 (Including Free Ones) | Reason Machines](https://reasonmachines.com/guides/claude-code-alternatives.md): The best Claude Code alternatives in 2026, compared from official docs: opencode, Aider, Codex, Cursor, GitHub Copilot, Google Antigravity, local models via Ollama, and Reason. What's free, what's open source, and what each is best for.
- [opencode vs Claude Code: Open Source, Models and Cost | Reason Machines](https://reasonmachines.com/guides/opencode-vs-claude-code.md): opencode vs Claude Code from their official docs: license, supported models and subscriptions, local models with Ollama, cost and install. Plus how to keep working in Reason when either runs out.
- [Antigravity vs Claude Code: Plans, Limits and Surfaces | Reason Machines](https://reasonmachines.com/guides/antigravity-vs-claude-code.md): Google Antigravity vs Claude Code from official docs: free and paid plans, five-hour and weekly limits, models and surfaces. Plus how to keep working in Reason when either runs out.
- [Reason vs Codex: Coding Agent Index 65.1 vs 62.5 | Reason Machines](https://reasonmachines.com/guides/reason-vs-codex.md): Reason vs Codex on the same model: Reason 1.1 scored 65.1 against Codex's 62.5 on the Coding Agent Index v1.5 with GPT-6 Astra, at 22% lower cost per solved task. Conditions, caveats and plans.
- [Claude Code Pricing 2026: Every Plan, Limits and API Cost | Reason Machines](https://reasonmachines.com/guides/claude-code-pricing.md): Claude Code pricing 2026 from Anthropic's official pages: Pro $20 ($17 annually), Max $100, Ultra $200, Team and Enterprise seats that include Claude Code, Sonnet and Opus API rates per token, and whether a plan or the API is cheaper.
- [Never Hit a Usage Limit Mid-Task: Automatic Model Fallback | Reason Machines](https://reasonmachines.com/guides/usage-limit-fallback.md): When a connected account hits its usage limit, Reason keeps the turn going on the same model on your other accounts, then on Reason credits, and tells you: "hit its usage limit · switched to Reason credits".
- [Claude Code with Ollama and Local Models | Reason Machines](https://reasonmachines.com/guides/claude-code-ollama.md): Run Claude Code with Ollama and local models using Ollama's documented setup: ollama launch claude, or ANTHROPIC_BASE_URL. What Ollama supports, what Anthropic doesn't, and the 64k context advice.
- [Claude Usage Limit Changes: Oct 26 to Nov 1, 2026 Tracker | Reason Machines](https://reasonmachines.com/guides/claude-usage-limit-changes.md): A dated tracker of the Claude and Codex usage limit changes of Oct 26 to Nov 1, 2026: Max 20x renamed Ultra with no five-hour limit for verified accounts, OpenAI's doubled Pro 100 and Pro 500 allowance, and more, each with its official source.
- [How to Connect Muse to Claude Code | Reason Machines](https://reasonmachines.com/guides/muse-claude-code.md): Connect Muse to Claude Code with Reason Network: add the Reason connector in Muse, put Claude Code on your network with one sentence, then ask Muse to message your Claude Code session. Free, no API key.
- [Codex and Claude Talk: Make Codex and Claude Code Talk to Each Other | Reason Machines](https://reasonmachines.com/guides/codex-claude-code-talk.md): How to make Codex and Claude talk to each other: connect your ChatGPT and Claude plans in Reason, then ask one model for the other's review in the same Session. In-app steps with screenshots, compared with copy-paste, the Codex plugin for Claude Code and MCP bridges.
- [Codex Alternatives: 8 Coding Agents Compared (2026) | Reason Machines](https://reasonmachines.com/guides/codex-alternatives.md): The best Codex alternatives compared from official docs: Claude Code, Cursor, Devin, GitHub Copilot cloud agent, Jules, opencode, Aider and Reason. Open source, cloud tasks, prices, and benchmarks.
- [Cursor Alternatives: Best AI Coding Tools Compared (2026) | Reason Machines](https://reasonmachines.com/guides/cursor-alternatives.md): Cursor alternatives compared from official docs: Claude Code, Codex, GitHub Copilot, Cline, Devin Desktop (formerly Windsurf), opencode and Reason. Editors, terminal agents and cloud agents, with prices.
- [Devin Alternatives: Autonomous AI Engineers Compared | Reason Machines](https://reasonmachines.com/guides/devin-alternatives.md): Devin alternatives compared from official docs: Reason, Codex cloud, Claude Code on the web, Cursor Cloud Agents, GitHub Copilot cloud agent, Jules and the open-source OpenHands. Where each runs, Slack, schedules and price.
- [Claude Code vs Cursor: Terminal Agent or AI Editor? | Reason Machines](https://reasonmachines.com/guides/claude-code-vs-cursor.md): Claude Code vs Cursor from Anthropic's and Cursor's docs: what each is, models, cloud agents, Slack, plans and limits. Which to choose, whether to use both, and how Reason carries a task when either runs out.
- [Claude Code vs Codex vs Cursor: Which to Use in 2026 | Reason Machines](https://reasonmachines.com/guides/claude-code-vs-codex-vs-cursor.md): Claude Code vs Codex vs Cursor, compared from Anthropic's, OpenAI's and Cursor's docs and public benchmarks: models, open source, cloud tasks, Slack, plans and limits. And how to use all three.
- [Claude Code Competitors: Who Competes with Claude Code | Reason Machines](https://reasonmachines.com/guides/claude-code-competitors.md): Claude Code's competitors compared from official docs: Codex, Cursor, GitHub Copilot, Devin, Google Antigravity and Gemini CLI, Grok Build, opencode, Aider, Cline and Reason. Open source, where they run, and price.
- [Best Coding Agents in 2026, Ranked by Public Benchmarks | Reason Machines](https://reasonmachines.com/guides/best-coding-agents.md): The best coding agents of 2026 ranked fairly: Artificial Analysis Coding Agent Index v2.0 and Terminal-Bench 4.0 first, then what each agent is best for. Claude Code, Codex, Grok Build, Devin, Cursor, Copilot, opencode, Reason.
- [Autonomous Coding Agents: What They Are and Which to Use | Reason Machines](https://reasonmachines.com/guides/autonomous-coding-agents.md): What autonomous coding agents are, how they work, and how the main ones compare from their official docs: Reason, Devin, Codex cloud, Claude Code on the web, Cursor Cloud Agents, Copilot cloud agent, Jules and OpenHands.
- [Reason vs Claude Code: Differences, and How to Use Both | Reason Machines](https://reasonmachines.com/guides/reason-vs-claude-code.md): Reason vs Claude Code: where each runs, models, schedules, Slack, limits and pricing, from Reason's docs and Anthropic's. Reason can run Claude models on your Claude plan, so you can use both.
- [Reason vs Devin: AI Software Engineers Compared | Reason Machines](https://reasonmachines.com/guides/reason-vs-devin.md): Reason vs Devin from Reason's docs and Cognition's: where each works, models, Slack, schedules, limits and pricing. Two autonomous software engineers, compared fairly.
- [Can Claude Code Run Overnight? Night Shift Options | Reason Machines](https://reasonmachines.com/guides/claude-code-overnight.md): Can Claude Code run overnight? Yes, as a cloud session or a scheduled task, from Anthropic's docs; local sessions stop when your laptop sleeps and every run stops at your plan's limits. How to set up a night shift, and how Reason keeps going.
- [Does Claude Code Run 24/7? What Keeps It Going | Reason Machines](https://reasonmachines.com/guides/claude-code-24-7.md): Does Claude Code run 24/7? Not on its own: cloud sessions and scheduled tasks keep it going without your computer, but runs stop at your plan's limits and on inactivity. Anthropic's options, and an always-on teammate that codes.
- [Background Coding Agents: Which Keep Working While You're Away | Reason Machines](https://reasonmachines.com/guides/background-coding-agents.md): Background coding agents compared from official docs: Reason, Codex cloud, Claude Code on the web, Cursor Cloud Agents, GitHub Copilot cloud agent, Jules, Devin, OpenHands and Grok Bot. Where each runs, triggers, schedules and limits.
- [Use All Your AI Subscriptions in One Coding Agent | Reason Machines](https://reasonmachines.com/guides/use-all-your-ai-subscriptions.md): One coding agent for every AI subscription: what Claude Code, Codex, Cursor, GitHub Copilot and opencode accept, from their docs, and how Reason runs your ChatGPT, Claude, Grok and Copilot plans and API keys in one place.
- [Use ChatGPT Subscription in Claude Code or Cursor? What Works | Reason Machines](https://reasonmachines.com/guides/use-chatgpt-subscription-for-coding.md): Can you use your ChatGPT subscription in Claude Code or Cursor? From Anthropic's, Cursor's and OpenAI's docs: not in Claude Code, and in Cursor only through the Codex extension. Where ChatGPT Plus, Pro and Business do work for coding.
- [Use Copilot Subscription in Claude Code, Codex or Cursor? | Reason Machines](https://reasonmachines.com/guides/use-copilot-subscription-with-claude-code.md): Can you use a GitHub Copilot subscription in Claude Code, Codex or Cursor? From GitHub's and the vendors' docs: Copilot powers the Claude and Codex agents on GitHub and Sign in with Copilot in Codex for VS Code, but not Claude Code itself or Cursor.
- [Cheapest Coding Agent in 2026: Free and Low-Cost Options | Reason Machines](https://reasonmachines.com/guides/cheapest-coding-agent.md): The cheapest coding agent in 2026, from each vendor's pricing page: free open-source agents (opencode, Aider, Cline, Kilo Code), $0 plans (Codex in ChatGPT Free, Copilot Free, Antigravity, Devin, Jules), and paid plans from $8 a month.
- [Cheaper Alternative to Claude Code: Best Options by Price | Reason Machines](https://reasonmachines.com/guides/cheaper-claude-code-alternative.md): The best cheaper alternatives to Claude Code, by price, from each vendor's pricing page: Codex in ChatGPT Go, GitHub Copilot Pro with Claude models, Cursor Plus, Antigravity, opencode, Aider and Cline. What you give up with each.
- [Claude Max Alternative: Plans for Heavy Claude Code Users | Reason Machines](https://reasonmachines.com/guides/claude-max-alternative.md): Claude Max alternatives from official plan pages: Claude Ultra, ChatGPT Pro, GitHub Copilot Max, Cursor Super and Ultra, and Devin Max, at $100 to $200 a month. Price, usage and limits compared, for the Claude Code Max plan too.
- [ChatGPT Pro vs Claude Max: Price and Usage Limits for Coding | Reason Machines](https://reasonmachines.com/guides/chatgpt-pro-vs-claude-max.md): ChatGPT Pro vs Claude Max for coding, from OpenAI's and Anthropic's plan docs: price, Codex and Claude Code access, five-hour and weekly limits, extra usage, and benchmark scores for each plan's coding agent.
- [Cheapest Way to Use Claude Code: Plans, API and Cost Tips | Reason Machines](https://reasonmachines.com/guides/cheapest-way-to-use-claude-code.md): The cheapest way to use Claude Code, from Anthropic's docs: Claude Pro at $17 to $20 a month, what API use costs per day, Sonnet and Haiku prices, and the cost tips in Claude Code's own docs. Plus the API key setting that surprises people.
- [Coding Agent Guides: Usage Limits, Pricing and Comparisons | Reason Machines](https://reasonmachines.com/guides.md): Sourced guides to Claude Code and Codex usage limits, weekly resets, Claude Code pricing, alternatives and head-to-head comparisons, and how to keep working when one agent runs out.

### AI employees

- [AI Employees: What They Are and How to Hire One | Reason Machines](https://reasonmachines.com/ai-employees.md): What an AI employee is and how it works inside (memory, context, tools, schedules, escalation), roles by department, cost vs a human hire with sourced wages, a 30-day onboarding plan, a buyer's checklist, how to measure ROI and how to hire one in Reason in seven steps.
- [AI Agents for Business: 15 Tools by Job, Prices and How to Choose | Reason Machines](https://reasonmachines.com/ai-agents-for-business.md): AI agents for business compared by job: customer support, sales and leads, workflow automation and back office. Each of 15 tools, including Intercom Fin, Botpress, Sierra, Gumloop, n8n and QuickBooks, gets key features, best for, where it falls short and a price checked November 2026, plus enterprise platforms, a final verdict by scenario, a pilot budget and how to choose.
- [AI Workforce: What It Is and How to Build One | Reason Machines](https://reasonmachines.com/ai-workforce.md): An AI workforce is a team of AI agents that each own a job and hand work to each other. How to build one step by step, design and test one good agent, a handoff example, a worked ROI example, sourced data on AI and jobs, and the skills your people need to manage agents.
- [AI Executive Assistant: 15 Compared With Prices, Setup and FAQ | Reason Machines](https://reasonmachines.com/ai-employees/ai-executive-assistant.md): The best AI executive assistant for each job: fifteen compared with official prices and a "Best for" label each, including Fyxer, Howie, Notion AI, Lindy, Motion and Reclaim, plus a sample week, a time-saved estimate and AI vs human EA cost.
- [AI Revenue Analyst: An AI Revenue and Metrics Analyst | Reason Machines](https://reasonmachines.com/ai-employees/ai-revenue-analyst.md): Set up an AI revenue and metrics analyst in Reason: it reads Stripe and PostHog, posts revenue, signups and usage with the change since last time, and stays read-only.
- [AI Customer Support Agent in Slack | Reason Machines](https://reasonmachines.com/ai-employees/ai-customer-support-agent.md): Set up an AI customer support agent in Reason: it answers in the Slack channels you choose, grounds answers in your docs, follows a when-to-reply policy and escalates what it can't answer.
- [AI Social DM Manager for X and LinkedIn DMs | Reason Machines](https://reasonmachines.com/ai-employees/ai-social-dm-manager.md): How an AI social DM manager handles X and LinkedIn DMs in Reason: through a messaging app on your connected computer, with policies for bookings, investors, refunds and spam.
- [AI Investor Update Assistant: Drafts Only, You Send | Reason Machines](https://reasonmachines.com/ai-employees/ai-investor-update-assistant.md): Set up an AI investor update assistant in Reason: it drafts your monthly investor update from Stripe, PostHog and GitHub, plus research, pipeline and follow-up drafts. Drafts only; you send.
- [AI Employees for Small Business: 14 Tools by Job, Real Costs and a 1-Week Plan | Reason Machines](https://reasonmachines.com/ai-employees/ai-employees-for-small-business.md): AI employees for small business: which of four shapes you need, who handles social, website and SEO, leads, inbox, phone calls and bookkeeping, 14 tools compared with prices from each official site (Nora, Vellum, Copilot Studio, Agentforce and Intercom Fin added November 2, 2026), costs beyond the subscription, a holiday rush example and when to build your own.
- [Best AI Employees in 2026: 19 Compared With Prices | Reason Machines](https://reasonmachines.com/ai-employees/best-ai-employees.md): A fair comparison of 19 AI employees in 2026, from Lindy, Sintra and Marblism to Sierra, Ema, Beam AI, Devin, OpenHands and Reason, with a quick shortlist, a scoring table, prices from each company's own site (checked November 1, 2026 and November 2, 2026), 2026 trends and what happened to Vellum.
- [Reason vs Lindy: Which AI Employee Fits Your Team? | Reason Machines](https://reasonmachines.com/ai-employees/reason-vs-lindy.md): Reason vs Lindy compared fairly: a side-by-side table, prices from each vendor's own pricing page (Lindy read November 2, 2026), sample-month costs, when to pick each, a 5-step pilot and other Lindy alternatives.
- [Reason vs Sintra: Which AI Employee Fits Your Business? | Reason Machines](https://reasonmachines.com/ai-employees/reason-vs-sintra.md): Reason vs Sintra compared fairly: a side-by-side table, prices from each vendor's own pricing page (Sintra read November 2, 2026), sample-month costs, when to pick each, a 5-step pilot and other Sintra alternatives.
- [Reason vs Grok Bot: Compare, or Use Them Together | Reason Machines](https://reasonmachines.com/ai-employees/reason-vs-grok-bot.md): Reason vs Grok Bot compared fairly: a side-by-side table, X subscription prices (checked November 1, 2026), sample-month costs, when to pick each, a 5-step pilot, other alternatives and how to use both together.

### 简体中文

- [Reason Machines：让你所有的 AI 智能体协同工作](https://reasonmachines.com/zh-hans.md): Reason 是一款前沿编程智能体，将 Claude Code、Codex、Cursor、Grok Bot 和 Muse 连接在同一个网络中，与你的其他智能体协同工作。引擎免费、开源。
- [服务条款](https://reasonmachines.com/zh-hans/terms-of-service.md): 适用于 Reason 网站、应用、API、云端会话及本地运行时控制的条款。
- [隐私政策](https://reasonmachines.com/zh-hans/privacy-policy.md): Reason 如何收集、使用、共享、保留和删除信息。
- [删除账户](https://reasonmachines.com/zh-hans/delete-account.md): 如何申请删除您的 Reason 账户及相关数据。
- [联系](https://reasonmachines.com/zh-hans/contact.md): 联系 Reason，咨询问题或了解 Reason 如何助力您的公司。
- [让 Grok Bot、Muse、Claude Code 和 Codex 彼此对话 | Reason Machines](https://reasonmachines.com/zh-hans/network.md): 面向 AI 智能体的免费点对点网络。Grok Bot 和 Muse 可向 Claude Code、Codex、Cursor 和 Gemini CLI 请求协作。粘贴一行即可加入。开源引擎，端到端加密。
- [macOS 版 Reason](https://reasonmachines.com/zh-hans/blog/ara-for-macos.md): 运行云端会话，让 Reason 使用你的本地文件、工具和端口开展工作。
- [Chvátal–Sankoff 常数的改进上界](https://reasonmachines.com/zh-hans/blog/chvatal-sankoff-upper-bound.md): 2026 年 8 月，我们证明了 γ₂ ≤ 0.824443，这是据我们所知对 Lueker 2009 年上界 0.826280 的首个已证明的改进。此后已有人走得更远。
- [介绍 Reason 1.1](https://reasonmachines.com/zh-hans/blog/introducing-reason-agent-1-1.md): 在 GPT-6 Astra 上，Reason 在 Coding Agent Index 上以 65.1 比 62.5 胜过 Codex，每个已解决任务的成本低 22%（此为我们自行复现的结果，未经 Artificial Analysis 认可）。
- [介绍 Reason 1.0](https://reasonmachines.com/zh-hans/blog/reason-agent-deepswe.md): 我们在 DeepSWE v1.1 上的首批结果已能与前沿智能体框架竞争。相较于基准对照，解题时间中位数最多降低 39%。
- [Rust 智能体框架之争：grok-build 与 codex](https://reasonmachines.com/zh-hans/blog/rust-harness-wars.md): 我们克隆了 xAI 的 grok-build 和 OpenAI 的 codex，逐行阅读双方各二十个子系统，探究两个实验室在构建同类编程智能体时究竟有何不同。
- [面向初创企业的 Reason](https://reasonmachines.com/zh-hans/blog/yc.md): 在读及往届 YC 公司可领取价值 1,000 美元的额度，用于开始编程。在 Bookface 打开 Reason 优惠页面，按照兑换说明领取。
- [最便宜的软件工厂](https://reasonmachines.com/zh-hans/blog/about.md): 软件就是推理、智能体和运行时。智能体决定一项任务要用多少推理，所以我们在这里让软件变便宜。

### 日本語

- [Reason Machines：すべての AI エージェントをひとつのチームに](https://reasonmachines.com/ja.md): Reason は、Claude Code、Codex、Cursor、Grok Bot、Muse とひとつのネットワークで連携する、最先端のコーディングエージェントです。エンジンは無料で、オープンソースです。
- [利用規約](https://reasonmachines.com/ja/terms-of-service.md): Reason の Web サイト、アプリケーション、API、クラウドセッション、ローカルランタイム制御に適用される規約です。
- [プライバシーポリシー](https://reasonmachines.com/ja/privacy-policy.md): Reason が情報をどのように収集、利用、共有、保持、削除するかについて。
- [アカウントの削除](https://reasonmachines.com/ja/delete-account.md): Reason アカウントと関連データの削除を依頼する方法。
- [お問い合わせ](https://reasonmachines.com/ja/contact.md): ご質問や、会社での Reason のご利用についてのご相談は、Reason までお問い合わせください。
- [Grok Bot、Muse、Claude Code、Codex を互いに会話させる | Reason Machines](https://reasonmachines.com/ja/network.md): AI エージェントのための無料のピアツーピアネットワーク。Grok Bot や Muse が Claude Code、Codex、Cursor、Gemini CLI に仕事を依頼します。1 行貼り付けるだけで参加。オープンソースのエンジンで、エンドツーエンド暗号化。
- [macOS 版 Reason](https://reasonmachines.com/ja/blog/ara-for-macos.md): クラウドセッションを実行し、Reason にローカルのファイル、ツール、ポートを扱わせられます。
- [チヴァータル–サンコフ定数の改良された上界](https://reasonmachines.com/ja/blog/chvatal-sankoff-upper-bound.md): 2026 年 8 月、私たちは γ₂ ≤ 0.824443 を証明しました。Lueker による 2009 年の上界 0.826280 に対する、私たちの知る限り最初の証明済みの改良です。その後、さらに先へ進んだ結果も出ています。
- [Reason 1.1 のご紹介](https://reasonmachines.com/ja/blog/introducing-reason-agent-1-1.md): GPT-6 Astra では、Reason は Coding Agent Index で Codex を 65.1 対 62.5 で上回り、解決したタスクあたりのコストは 22% 低くなります（当社による再現結果であり、Artificial Analysis による承認を受けたものではありません）。
- [Reason 1.0 のご紹介](https://reasonmachines.com/ja/blog/reason-agent-deepswe.md): DeepSWE v1.1 での最初の結果です。フロンティアのハーネスと互角で、ベンチマークのベースラインに対し、解決時間の中央値を最大 39% 短縮しました。
- [Rust ハーネス対決:grok-build vs codex](https://reasonmachines.com/ja/blog/rust-harness-wars.md): xAI の grok-build と OpenAI の codex をクローンし、それぞれ 20 のサブシステムを一行ずつ読み込んで、同じコーディングエージェントを作る 2 つのラボが実際にどこで分かれるのかを調べました。
- [スタートアップ向け Reason](https://reasonmachines.com/ja/blog/yc.md): 現役およびアルムナイの YC 企業には、コーディングを始めるための $1,000 分のクレジットを提供します。Bookface で Reason のディールを開き、利用手順に従って受け取ってください。
- [最も安いソフトウェア工場](https://reasonmachines.com/ja/blog/about.md): ソフトウェアは推論、エージェント、ランタイムでできています。タスクにどれだけの推論を使うかはエージェントが決めるので、私たちはそこでソフトウェアを安くします。