cd /news/ai-agents/show-hn-claude-code-hits-its-5-hour-… · home topics ai-agents article
[ARTICLE · art-134637] src=github.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

Show HN: Claude Code hits its 5-hour limit, Codex picks up in the same terminal

LegCli released Leg, a terminal tool that watches coding-agent usage limits and hands off work to the next agent in the same terminal when a limit is hit, starting with Claude Code's five-hour cap and moving to Codex. Leg tracks usage per agent and account via Claude Code's usage endpoint and StopFailure hook, Codex's read-only app-server rate-limit read, agy's log, and Grok's billing proxy endpoint, and refreshes a context-handoff-bundle published on PyPI so the session resumes without retyping. The default ladder runs claude/fable, claude/opus, claude/sonnet, then each remaining installed agent, and Leg strips API-key environment variables such as ANTHROPIC_API_KEY and OPENAI_API_KEY to enforce subscription logins.

read35 min views2 publishedSep 19, 2026
Show HN: Claude Code hits its 5-hour limit, Codex picks up in the same terminal
Image: Michielbdejong (auto-discovered)

Type leg claude, leg codex, leg agy or leg grok instead of the bare command. You get the same interactive agent; Leg opens a board next to it, watches the usage limit, keeps a handoff bundle current, and when the limit hits it starts the next agent in the same terminal from that bundle.

Claude hits the five-hour wall. The terminal reads handing off to codex, and codex carries on there. Nothing is retyped. (the full 53-second run)

You keep using your coding agents exactly as you do today, in any terminal, from your own config directory: Leg adds its hooks in a separate per-session settings file and never edits yours (unless you turn on the portable harness, which writes only marked, backed-up, Leg-owned files and regions). leg claude --model opus is claude --model opus with four things running alongside it:

  1. A board. Opened once in your browser, reused after that. Every Leg session in every terminal is a row on it, and the row's register reads status, repo on branch, uncommitted and unpushed counts, agent and model, and how long it has been quiet:waiting on you leg on main dirty 3 ahead 2 claude/fable . Under that come the prompt, the one thing worth knowing, the files it is touching and the usage bucket that will stop it. Capacity is one strip at the top rather than a region, with the login panels and their gauges behindCapacity and models . Two sessions editing the same file in one repo are flagged on both rows, and a second session in a checkout that already has one gets its own worktree and aLand button instead of writing over the first.
  2. Usage tracking per agent and account, from what each CLI already exposes: Claude Code's usage endpoint and itsStopFailure hook, Codex's read-only app-server rate-limit read, agy's log, and Grok's billing proxy endpoint.
  3. A context handoff bundle (context-handoff-bundle ) refreshed as the session goes, so the work is always ready to hand off.
  4. The handoff itself. Near the limit you get a warning. At the limit Leg saves the bundle, stops the agent, and starts the next option in the same terminal from that bundle. The options are a ladder of rungs, and a rung is an agent, a login and a model: the default isclaude/fable ,claude/opus ,claude/sonnet , then each remaining installed agent. A Fable limit walls Fable and not the login, so the first move is usually to another model on the same subscription, and for claude that move isclaude --resume <id> --model opus , which keeps the conversation instead of replaying the bundle. Only when every claude rung is out does it move to another CLI. Settings holds the ladder for new terminals,leg ladder edits it from a shell, andDetails on a row edits the copy that terminal is running. Nothing is retyped. When every rung is out, Leg tells you which resets first and when, waits for that reset with a countdown, and starts that agent from the bundle.

Subscription logins only: Leg strips ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_BASE_URL, ANTHROPIC_CUSTOM_HEADERS, OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_API_BASE, GEMINI_API_KEY, GOOGLE_API_KEY, GOOGLE_GEMINI_BASE_URL, GOOGLE_GENAI_USE_VERTEXAI, GOOGLE_GENAI_USE_ENTERPRISE, GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_LOCATION, GOOGLE_APPLICATION_CREDENTIALS, GROK_API_KEY, XAI_API_KEY, CLAUDECODE, CLAUDE_CODE_*, CLAUDE_EFFORT, and CLAUDE_PLUGIN_DATA before any agent starts. It then sets CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=0 for a detached Claude print session. Leg never edits ~/.claude/settings.json or any other settings file of yours; its hooks ride in a separate per-session --settings file. The one thing it does write outside ~/.leg by default is the folder-trust answer, below; the optional portable harness is the other, and only after leg harness enable. leg uninstall removes only ~/.leg.

Each agent CLI asks once, the first time it runs in a directory, whether you trust that folder, and Claude Code asks a second question when a CLAUDE.md above the repo imports a file from outside it. A handoff fires when the limit hits, which is usually when nobody is watching, so an agent that stopped on that prompt would sit there until morning with the bundle already written.

Before starting an agent, Leg records the same answer you would have given, for the repository you already chose by typing leg claude in it:

agent file what is written
claude ~/.claude.json (or$CLAUDE_CONFIG_DIR/.claude.json ) projects["<repo>"].hasTrustDialogAccepted: true
claude the same entry, only when such an import exists hasClaudeMdExternalIncludesApproved ,hasClaudeMdExternalIncludesWarningShown
codex ~/.codex/config.toml [projects."<repo>"] trust_level = "trusted"
agy ~/.gemini/antigravity-cli/settings.json trustedWorkspaces: ["<repo>", "<worktree>"]

For Claude Code this is the documented remedy: its permissions guide says to set projects["<path>"].hasTrustDialogAccepted to true in ~/.claude.json, where <path> is the repository root.

Leg never creates one of those files: if it is not there, that CLI has not run as you yet and its own first-run flow is next, with you at the keyboard. It never rewrites a file to say what it already says, and it never removes what is already in one. When it approves an external CLAUDE.md import it prints the full path of every file it approved, to the terminal and to the session timeline on the board, so the approval is on the record rather than invisible.

Set LEG_TRUST=never to switch all of it off and answer the prompts yourself.

Prerequisites: Node 22 or newer, git, Python 3 with pip, and at least one logged-in agent CLI (claude, codex, agy or grok).

npm install -g @ucsandman/legcli
pip install -U context-handoff-bundle
cd <any repo>
leg claude

npm install -g leg-agents is the same release: it pins this version of @ucsandman/legcli and exposes the same leg binary.

That is the whole setup. The first leg <agent> starts the board on http://127.0.0.1:4747 and opens it; later sessions reuse it. Anything after the agent name passes straight through (leg codex -m gpt-5.3-codex-spark, leg claude --resume). The agent's own prompt and permission flags pass through unchanged, and your settings file is never edited: Leg's hooks ride in a separate per-session --settings file.

Leg is commercial, source-available software, and the source is on GitHub: every .mjs file that runs is in the package you just installed, at $(npm root -g)/legcli/src, and the license lets you read it and modify your own copy. There is nothing compiled, minified or bundled to see through.

Nothing is guessed from screen scraping. Each tap was read from the CLI's source or documentation and then checked on a real machine (2026-09-11, Claude Code 2.1.268, codex-cli 0.153.4, agy 1.2.0); the rightmost column says which.

agent usage percentages the wall (limit hit) how Leg attaches status
claude GET api.anthropic.com/api/oauth/usage with the login Claude Code stored, the same data as/usage and the built-in status line (five_hour ,seven_day ,utilization ,resets_at ); polled every 60 s StopFailure hook witherror: rate_limit (docs ) one extra settings file per session via --settings , carrying only Leg's own hooks;autoContinueAtUsageLimit is set tofalse because Leg owns the handoff observed live
codex read-only account/rateLimits/read through the app-server, polled every 60 s by the board and active attach; windows are identified by duration (300 minutes = 5h, 10080 = 7d) task_complete.error.codex_error_info: usage_limit_exceeded , message "You've hit your usage limit … try again at …" (codex-rs/protocol/src/error.rs ) no model turn and no hook are injected; the board reads the CLI backend and records only returned windows verified by source and regression tests
agy none exposed (agy's own status line fetches a quota summary that is written nowhere) RESOURCE_EXHAUSTED , "it resets in …", "out of quota" in the log --log-file per session;~/.gemini/antigravity-cli/history.jsonl gives the prompts and conversation id observed live (a real RESOURCE_EXHAUSTED with its reset was read from the log on 2026-09-11)
grok GET cli-chat-proxy.grok.com/v1/billing?format=credits andGET cli-chat-proxy.grok.com/v1/user?include=subscription with OAuth token from~/.grok/auth.json (creditUsagePercent ,currentPeriod weekly reset); polled every 60 s -32003 rate limit error, "You've hit the rate limit for your plan. Try again later.", "Rate limited (429)",StopFailureKind::RateLimit (xai-org/grok-build ) --debug-file grok.log per session;~/.grok/sessions/<encoded-cwd>/prompt_history.jsonl gives prompts and session id verified by source and proxy endpoint

Why not Claude Code's status line JSON (rate_limits.five_hour.used_percentage): on 2.1.268 the custom statusLine Leg passes through --settings did not run, so the endpoint poll is the source. Leg still writes a statusLine entry that records the same fields, so the moment a build honours it the poll becomes a fallback.

For Codex, Leg does not infer availability from a lower percentage: only an explicit available answer from the backend clears an earlier wall. The board labels each bar as <n>% used and marks an old reading as stale rather than presenting it as current.

  1. Warning. At 85 % of any window (LEG_WARN_PCT ) the session card turns amber, the event log names the next option, and the terminal bell rings once.
  2. Limit. claude: theStopFailure hook fires withrate_limit . codex: the rollout reportsusage_limit_exceeded . agy: the log saysRESOURCE_EXHAUSTED . The account is marked walled until the reset the CLI reported (or the soonest known window reset).
  3. Bundle. Leg writes structured notes (task, the last messages from the transcript,git diff --stat , dirty files, files edited this session, recent commits, why it stopped) and runscontext-handoff-bundle save --repo-local with one slug per session, updated in place (--update <slug> ) at every checkpoint, about every two minutes while the session is active, and at every warning, limit and hand-off. If the session maintained.leg/SYNTHESIS-<session-id>.md , Leg inlines it into the resume file as a## Synthesis section ahead of the raw dump.
  4. Switch. The agent process is stopped and the terminal is restored. Two claude destinations keep the conversation itself instead of the bundle: a weaker model on the same login (claude --resume <id> --model <alias> ), and another claude login you added withleg accounts add , which sees the same transcript through theprojects junction the account carries (claude --resume <id> under that login'sCLAUDE_CONFIG_DIR ). A weekly or Fable wall on one login then continues on the other with the conversation it already had, and the timeline sayskept the conversation . Every other rung starts in the same terminal with a short pointer prompt: read.leg/RESUME-<session-id>.md (thecontext-handoff-bundle load output, the## Synthesis section if present, and the reason for the switch), checkgit status andgit diff , continue, do not ask the human to restate the task. The pointer prompt directs the next agent to read Synthesis first, treat ruled-out approaches as settled, and start from the top-ranked next step. The same text is copied to.leg/RESUME.md , the file people open by habit, and both are stamped with the commit and the live terminals they describe.claude "<prompt>" ,codex "<prompt>" andagy -i "<prompt>" all open the normal interactive session with that first turn.
  5. The ladder. A rung is an agent, a login and a model, and Leg walks from rung 1 every time. Other models on the same login come first, then other accounts of the same agent, then every other agent in the saved order, each tried once. It is a priority list, not a rotation: put agy at the bottom and agy is the last rung from a Claude terminal and from a Codex terminal alike. The board prints the exact sequence with the rung running now marked, plus the first rung eligible from current install and limit state. UseChange the ladder in a terminal's expansion to change that terminal,leg ladder to do it from a shell, or Settings to set the default copied by new terminals. A rung whose CLI is missing, whose wall has not reset, or that shares the window that is already out, is skipped with its reason.
  6. All out. The terminal prints each option with its reset time, soonest first, then stays open with a countdown to the first reset and starts that agent from the bundle when it arrives. The row sayswaiting for <agent> at <time> . Ctrl-C (or End on the row) quits with exit 3 instead.

You can force a handoff any time: the Hand off now button on the row, or leg sessions handoff <id>. Verified on this machine: leg claude opened the real Claude Code TUI with Leg's hooks firing into the session log, the usage poll recorded 36 % of the 5h window and 74 % of the 7d window, the warning fired at 96 % of the 7d window and named codex as the next option, and a limit saved the bundle, stopped claude and started codex in the same terminal with the pointer prompt. A real StopFailure arrived on 2026-09-11 and is kept at fixtures/live/claude/limit-rate_limit.json; to drive the path on demand, leg sessions simulate-limit <id> sends the same StopFailure rate_limit payload Claude Code would send through Leg's hook: verified end to end on a haiku session, the hook set the limit, the runner saved the bundle, stopped claude and started codex, which read .leg/RESUME.md on its first turn. The simulated wall clears after two minutes and is never kept as evidence. The first real StopFailure was saved that way, with secrets scrubbed, at fixtures/live/claude/limit-rate_limit.json, and the claude docs row flipped to observed-live (node scripts/live-limits.mjs). The same capture is wired for codex usage_limit_exceeded and agy RESOURCE_EXHAUSTED; no payload for either has been kept yet.

Terminals started by this version can change their ladder while they run: Settings holds the default for new terminals, and Details on a row edits the copy that terminal is running. An older terminal stays on the order it started with; its row says a restart is needed and can save the desired default for the next launch. A normal agent exit ends the terminal. It does not trigger a handoff.

With the portable harness enabled (leg harness enable, off by default), the hand-off also prepares the destination's environment before it starts: the source agent's global rules, identity, hooks, skills, subagents, slash commands, MCP servers and permissions, rendered into the destination's own files as far as it can represent them. The terminal then reads, for instance, codex harness partial · 8/8 components, 3 dropped · 1 file(s) written, and the terminal's details list what was dropped and why. The source is fingerprinted so an unchanged environment costs a few stat calls; a policy (warn, sync, strict) says whether a hand-off may write, and strict refuses a destination it cannot make safe. Credentials never move and which login runs stays the account layer's decision.

Two agents in one working tree write over each other's files. So when you start leg codex in a checkout where leg claude is already live, the new session gets its own git worktree, <repo>/.leg-worktrees/<session-id> on branch leg/<session-id>, cut from the branch the checkout has out, and the terminal prints one line saying where it is. The agent starts there; the card, the usage tracking and the handoff work the same. --no-worktree shares the checkout on purpose (Leg takes the flag out; the agent never sees it).

The card of a session with its own worktree has a Land button. It sends the branch through the merge queue: whatever the agent left uncommitted is committed on the branch, the branch is rebased onto its base, the repo's test command runs (package.json test, pytest, or none with a warning), and the base is fast-forwarded, never merged. When a step fails nothing lands and the card says why: rebase-conflict with the files, tests-red with the end of the output, dirty-trunk when the checkout has local changes the landing would overwrite. Local changes it would not touch are left alone. The landed-on-trunk list says which terminal landed each commit. Remove safely prunes a finished session only when its worktree is clean and its branch is already on the base. Remove record is a separate visible button with a confirmation: it removes only Leg's saved session record and deliberately keeps the worktree, branch, unmerged commits, and dirty files.

Verified live on 2026-09-11 with three haiku sessions in a throwaway repo. The first stayed in the checkout; the second and third each got a worktree and appended a line to README.md. Land on the second, clicked on the real board, ran the repo's tests and fast-forwarded main; Land on the third bounced with rebase-conflict on README.md and kept its commit on its branch; the landed-on-trunk list named the second terminal (the screenshot above).

Off until you run it. leg share on binds the board to your Tailscale address (or --bind lan, or an address you name) and gives every human their own name and token; until then the board stays on 127.0.0.1 and there is no token at all.

leg share on              your own link, printed once
leg share add sam         sam's link, printed once (a guest)
leg share add dana --role operator    dana runs cards, not this machine
leg share                 who is on the board (never a token again)
leg share rotate sam      sam's old link stops working
leg share off             back to 127.0.0.1; every link stops working

There are three roles. owner is everything: the machine's settings, the harness, every terminal, the cards, the history index and the audit trail. operator is the pipeline board and their own terminals: they add, run, approve, reassign and kill cards, and they never see this machine's settings, its home path, its repository paths, its conversation index or the audit. guest is the terminals lane, read-only and redacted, with Request handoff as their only button.

A token is kept as a sha256 hash, so a lost link is re-issued, never re-read. The board takes the token out of the address bar and keeps it in the browser. Your own browser on this machine needs no token.

What another human sees is the Terminals region, read-only. Each panel says whose terminal it is. On a panel that is not theirs there is no prompt, no file name, no path, no bundle and no event log; what stays is the agent and session tail, the status word, the sentence read-only: wes owns this terminal, repo@branch, the worktree line, the elapsed clock, and one button, Request handoff. The instrument head prints not shared in place of every percentage. A request lands on the owner's panel as sam asked to take this terminal at 11:04 PM with Approve sam and Dismiss sam. The background side of the board (cards, logs, the floor) stays the owner's alone. A terminal belongs to the human who started it: LEG_PERSON=sam leg claude on the same machine is sam's card, not yours.

The security pass that goes with it: every /api route needs a token, the event stream included; twenty wrong tokens from one address and that address waits a minute; one identity gets 600 requests a minute; a guest gets 403 on everything that is not theirs; and the tests send a bad and a missing token to every route. Verified live on 2026-09-11: two terminals on one machine, one wes's and one sam's; sam's board showed wes's card with the prompt hidden and only Request handoff, and sam's request reached wes's board (~/.leg/board.log: "hand-off requested … by sam").

Off unless you hand it a certificate. Leg does not issue one: a self-signed pair teaches everyone on the board to click through a warning, which is worse than plaintext on a network that is already private. On Tailscale the pair is one command:

tailscale cert <machine>.<tailnet>.ts.net
leg share on --tls-cert <machine>.<tailnet>.ts.net.crt --tls-key <machine>.<tailnet>.ts.net.key

LEG_TLS_CERT and LEG_TLS_KEY do the same without writing the paths into share.json, and they win over it. With a pair configured the shared address serves https and every printed link says https://; the companion listener on 127.0.0.1, which exists so this machine's own browser needs no token, stays plain http, because the certificate is for the shared name and loopback traffic never leaves the machine. A pair that is half-configured, missing, unreadable or empty stops the board with exit 3 rather than quietly serving plaintext. Renew the pair, then leg down && leg up to pick it up.

Settings → Audit trail is one list across every terminal and every card, newest first: hand-offs, landings, approvals, reassignments, kills, each with the person or agent that did it, the repo and the time. Filter by person or by kind. It reads what the ledger already recorded, so nothing new is stored, and it prints how much it read beside the answer (14 terminals and 3 cards, 812 events read) so an empty trail cannot be mistaken for a quiet week. Owner only: the trail names repositories and people.

leg digest (default window 8 hours; --since 2d, --since 30m, or an ISO time; --json for the record; GET /api/digest?since= on the board, owner only) is the trail read the other way round: grouped by repository, what needs you first. The first line is the volume it was read from (3 terminals, 2 cards, 1 landing, 412 events read (66 sessions and 9 cards on disk)), then needs you: a live terminal waiting on a question, a card parked for a human, a card that failed, a terminal that was lost, in that order and each with how long ago. Then one block per repository: every terminal that moved in the window with its login and model, state, turns, files, commits ahead and the events worth a line (a wall, a hand-off and whether it kept the conversation, an all-out wait, the end), every card with its last event, every landing with who pressed Land, and last the walls standing right now with their reset times. Nothing new is recorded; a window with nothing in it says so with its counts.

leg <agent> opens it; leg open reopens it; leg down stops it.

  • Verdict and capacity strip : the largest sentence on the page, then one strip with a token per login carrying the binding bucket, a short track, the percentage and one state word or clock:claude 63% 6:01 PM ,codex back Sat 10:11 PM ,agy no figure ,grok no reading .Capacity and models opens a drawer holding the full login panels: the 5h and 7d rails, the reset and how long that is, where and when Leg read the number,at the wall with itsback <day time> , and on the claude panel a rail of model chips (fable 63% ,opus 12% , a walled one sayingout until 9:14 PM ). The sticky instrument head is still what/floor puts at the top.
  • Terminals : one full-width row per session, the ones that need an answer first. The register reads the status word,repo on branch ,dirty <n> andahead <n> ,agent/model ,quiet <n>m , andown worktree, from main when the session cut its own worktree; then the first prompt as a button, exactly one sentence (the highest-ranked thing true about the terminal), analso: disclosure naming the rest, the files as comma-separated text, the binding bucket as63% of the fable week , and an elapsed clock with the session tail. Two live sessions in one repo touching the same file printcodex (codex-99ab) is changing src/server.mjs in another checkout; whoever lands second rebases on both rows. A Claude row parked at a permission or idle prompt rises to the top, sayswaiting on you in place of its status word, prints the question verbatim, and puts a count in the browser tab ((2) Leg , with a dot on the favicon). After a Land the sentence islanded on <base>, <7-char sha>, <n> files, +<added>/-<removed> , orLand was attempted at <time> onto <base> and bounced: <first line of the reason>. The branch still holds every commit; nothing was lost.
  • Landed on main : one flat list across every repo the board can see, newest first, each row with the short sha, the subject, arepo@branch chip, and when plus who. A commit a Land put there sayslanded by <agent> (<id tail>) .
  • Buttons , in a fixed order that never reflows: Land, Hand off now, Details, End, under the clock at the right of the row. Land is drawn only when it can run; its reason is said once above the panel. Once a session has ended, Remove and Remove record take End's place. Details opens an expansion in flow under the panel.Hand off now takes the first open rung of the ladder. To name the destination instead, open Details and useHand off now to , which lists every rung with its model, whether it keeps the conversation, and the reason a greyed one cannot be picked (at its usage limit, back 11:40 ,shares the window that is out, buys nothing ,not installed on this machine ).leg sessions handoff <id> --to claude/default/opus is the same choice from a terminal. If the rung you picked walls between the click and the hand-off, the work still continues down the ladder and the terminal says which one took it instead.
  • Below it, Conversations : every coding-agent conversation on this machine as a count that opens, the ones Leg started and the ones Claude Code, Codex, Grok, Antigravity and Copilot keep in their own stores. Filter by agent, search, repository, or only what Leg started; a row opens its last messages in place, with theleg history continue command to copy where the agent can resume by id. The same drawer lists every checkout Leg can see. ThenSettings . Background tasks are not down here: a live card is a row in theBackground panel directly under Terminals (see below), and only the finished ones collapse into a ledger line beside these counts.

The board reads ~/.leg/sessions/*/session.json over server-sent events; a session whose runner process is gone is marked lost, never shown as live.

Claude Code, Codex, Grok, Antigravity and Copilot each keep their history in their own place. leg history is one list over all of them, whether or not Leg started the conversation, and leg worktrees is one list over every checkout: git's, Leg's own, and the ones those conversations ran in.

leg history                          newest first, every agent; --provider, --repo, --search, --json
leg history show claude:0fc5         where it ran, its last messages, whether Leg can continue it
leg history continue claude:0fc5     leg claude --resume <id> in that folder, supervised like any session
leg worktrees                        path, repo, branch, exists, uncommitted, owner, conversations, stale

Nothing moves: each agent's store stays where it was, Leg writes only its own ~/.leg/history/index.json, reads transcripts from their head and tail, and opens messages only when you open a conversation. A session Leg started and the same conversation in the agent's store are one row, marked leg; the rest are external. On a shared board the whole group is the owner's. The support matrix (which agents list, show messages, continue) and every file read are in docs/history.md.

Optional. leg accounts add claude work creates ~/.leg/accounts/claude/work, junctions your hooks, skills, agents, commands, plugins, rules, scripts, output-styles, tools and projects into it, copies settings.json, CLAUDE.md and the status-line scripts (refreshed from your real ~/.claude before every launch), and prints one line to paste:

$env:CLAUDE_CONFIG_DIR='C:\Users\you\.leg\accounts\claude\work'; claude auth login

projects is Claude Code's conversation store, so the second login sees the same conversations and the same auto-memory as the first, and a hand-off from one login to the other keeps the conversation: the terminal moves to work with claude --resume <id> and no bundle prompt. That is the whole point of a second 20x login when the Fable or the weekly window on the first one is out. An account made by an older Leg gets the junction the next time it starts. Claude Code writes that directory; Leg only reads it.

Same for codex (CODEX_HOME; config.toml, AGENTS.md, skills, prompts, rules, plugins, agents, hooks, memories shared; a codex hand-off still takes the bundle, because codex resume under a second CODEX_HOME has not been observed). agy 1.2.0 has no config-directory override, so it stays one account. Only the login lives in the account directory; leg accounts rm removes the junctions and the directory and never touches your real home.

The terms, as published (effective dates below):

  • Anthropic Consumer Terms (effective 2025-10-08): "You may not share your Account login information, Anthropic API key, or Account credentials with anyone else" and you "must not … bypass any of our systems or protective measures."
  • Anthropic Usage Policy (effective 2025-09-15): do not "Coordinate malicious activity across multiple accounts to avoid detection or circumvent product guardrails" or "Utilize automation in account creation."
  • OpenAI Terms of Use (effective 2026-01-01): "You may not share your account credentials or make your account available to anyone else" and you may not "circumvent any rate limits or restrictions or bypass any protective measures."

Owning two paid subscriptions is not named as prohibited by either. Rotating to a second account of the same vendor because the first one is rate-limited sits close to OpenAI's "circumvent any rate limits" wording and Anthropic's "circumvent product guardrails". Leg's default chain switches vendors (claude → codex → agy), which is plainly fine. Same-vendor rotation only happens after you run leg accounts add; that is your call.

  • Never edited :~/.claude/settings.json ,~/.claude.json ,~/.codex/config.toml , agy's files, your repo's settings. Claude Code gets hooks through a per-session--settings file under~/.leg ; codex and agy get nothing injected.
  • Read, never written : each agent's own history (~/.claude/projects ,~/.codex/sessions ,~/.grok/sessions ,~/.gemini/antigravity-cli ,~/.copilot/session-state ) forleg history ; the index it builds lives under~/.leg/history/ , and no SQLite file is ever opened.
  • Written only after leg harness enable (the portable harness , off by default): the destination client's global rules file (~/.codex/AGENTS.md ,~/.gemini/GEMINI.md ,~/.claude/leg-rules.md plus one@ line inCLAUDE.md ), markedleg harness regions inside the client's own config files, per-skill directory links, and one file per subagent and slash command. Every file carriesGENERATED by Leg harness , every overwrite is backed up under~/.leg/harness/backups , a hand-edited file is skipped and named, the source client is never written, and no credential ever moves (an MCP key becomes${NAME} ).leg harness disable stops it and removes nothing.
  • Written in your repo :.leg/ (session notes,RESUME.md and oneRESUME-<session-id>.md per hand-off),.context-handoffs/ (the bundles) and.leg-worktrees/ (a second session's worktree), all added to.git/info/exclude , plus theleg/<session-id> branch of a session with its own worktree. Landing fast-forwards your branch; nothing is ever pushed.
  • Stripped from every agent's environment :ANTHROPIC_API_KEY ,ANTHROPIC_AUTH_TOKEN ,ANTHROPIC_BASE_URL ,ANTHROPIC_CUSTOM_HEADERS ,OPENAI_API_KEY ,OPENAI_BASE_URL ,OPENAI_API_BASE ,GEMINI_API_KEY ,GOOGLE_API_KEY ,GOOGLE_GEMINI_BASE_URL ,GOOGLE_GENAI_USE_VERTEXAI ,GOOGLE_GENAI_USE_ENTERPRISE ,GOOGLE_CLOUD_PROJECT ,GOOGLE_CLOUD_LOCATION ,GOOGLE_APPLICATION_CREDENTIALS ,CLAUDECODE ,CLAUDE_CODE_* ,CLAUDE_EFFORT , andCLAUDE_PLUGIN_DATA .CLAUDE_CODE_* does not includeCLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS : Leg sets that one to0 for a detached Claude print session.
  • Read but never written or printed : Claude Code's stored login, sent only toapi.anthropic.com for the usage numbers. The ledger scrubs bearer tokens and key shapes from every line regardless.
  • leg uninstall --yes : removes~/.leg (sessions, usage, extra account directories with their junctions, v0.1 cards, the board pidfile) and nothing else; thennpm rm -g @ucsandman/legcli .
leg claude|codex|agy|grok [agent args…] the interactive agent, board alongside, handoff on limit
      [--no-worktree]                  share the checkout with a live session instead of a worktree
leg sessions ls [--json]             every session and its usage
leg sessions show|events <id>
leg sessions handoff|end <id>        same as the board buttons
leg sessions handoff <id> --to <agent>[/<account>]   hand off to a destination you name
leg sessions rm <id>                 forget an ended session
leg sessions simulate-limit <id>     the real limit path without a real wall (claude, agy, grok)
leg history [ls] [--provider p] [--repo r] [--search q] [--managed|--external] [--live] [--all] [--json]
                                     every conversation on this machine, Leg's own and the agents' own (read only)
leg history show <id> [--messages n] [--json] | continue <id> [agent args…] | refresh [--full] | providers
leg worktrees [--repo <path>] [--no-dirty] [--json]   every checkout: git's, Leg's, the conversations' (read only)
leg digest [--since 8h|2d|<iso>] [--json]   what happened while you were away: needs-you first, then every
                                     terminal, card, landing and wall in the window, by repository (read only)
leg accounts ls                      logins and their 5h/7d usage
leg accounts add <claude|codex|grok> <name> | rm <agent> <name> | terms
leg harness status|inspect|check|explain|history [--json]   the portable harness, read-only
leg harness enable [--source claude|codex] [--policy warn|sync|strict] [--yes]
leg harness sync [--to codex,agy] [--force] [--dry-run] | diff <client> | doctor
leg harness capture [claude|codex] | source <client> | policy <mode> | disable
leg license                          the license on this machine, or where to buy one
leg license activate <key> | deactivate | refresh   (refresh renews a Team key)
leg adapter list|show <n>|check <n>  every adapter, built-in and custom
leg adapter template [--name n]      a starter spec to fill in
leg adapter add <file.json> | rm <n> any CLI as a card agent, from JSON
leg share                            who is on the board (off by default; Team plan)
leg share on [--bind tailscale|lan|<addr>] [--port N] [--tls-cert <f> --tls-key <f>] | off
leg share add|rotate|rm <name> [--role owner|operator|guest]   one link per human, printed once
leg open | down | status             the board
leg uninstall [--yes]

Environment, all optional: LEG_HOME (default ~/.leg), LEG_PORT (4747), LEG_ACCOUNT (start on a named login), LEG_WARN_PCT (85), LEG_NO_HANDOFF=1 (warn and record, never switch), LEG_NO_OPEN=1 (do not open the browser), LEG_USAGE_POLL_MS (60000), LEG_CLAUDE_ARGS / LEG_CODEX_ARGS / LEG_AGY_ARGS / LEG_GROK_ARGS (extra args for a leg Leg starts after a hand-off, e.g. -m gpt-5.3-codex-spark), LEG_CLAUDE_BIN, LEG_CODEX_BIN, LEG_AGY_BIN, LEG_GROK_BIN, LEG_CHB_BIN, LEG_PERSON (whose terminal this is when the board is shared), LEG_TLS_CERT / LEG_TLS_KEY (serve the shared board over https; they win over share.json), LEG_RATE_MAX (600 requests a minute per human) and LEG_RATE_MAX_FAILURES (20 wrong tokens per address).

Version 0.1 was the other way round: you dropped a task card on the board and Leg ran the agents headless in a git worktree, one per card, with a fallback chain, path leases, a scheduler and a merge queue. All of that still runs, and a card is now a terminal you are not sitting at: same register, same one sentence, same ladder, same bundle. Live cards are rows in the Background panel directly under Terminals; finished ones fall into one ledger line that opens.

Starting one is a single field. Run in the background: takes the task, and the sentence under it is inferred with its nouns as buttons: the repo of the terminal you were last in, the saved ladder, and the build workflow, which stops with its changes in the card's worktree and does not merge them. More settings is the full form, with the Build, test, and merge and Factory workflows that include an automatic land station, plus fallback agents, permissions, approval gates, turn caps, leases, trunk, merge method, tests, title, and scripted test and demo adapters. End, and keep going as a card on a terminal's End row makes a card out of the work in front of you; Take over on a card gives you the command that turns it back into a terminal.

  • leg up boots the board with the scheduler and merge queue and streams redacted logs;leg card add --repo <path> --task "<t>" --chain claude,codex --queue creates a card; presetsbuild ,build-land ,factory ; station kinds agent, test, land, human.
  • Adapters spawn the CLIs headless as argv, never through a shell, with their own permission modes and never a bypass flag: claude -p --output-format json --permission-mode <m> ,codex exec --json -s <m> -C <worktree> ,agy -p --output-format json --mode <m> --add-dir <worktree> ,grok --prompt-file <f> --output-format json --permission-mode <m> --cwd <worktree> ;fake ,fake-claude ,fake-codex ,fake-agy for tests and demos.
  • Any other CLI is a card adapter too, from a JSON spec and no code:leg adapter template --name muse > muse.json , fill in the command and its flags,leg adapter add muse.json , then--chain muse,claude .leg adapter check muse prints the exact command line a leg would run before one does. Seedocs/adapters.md . A custom adapter runs cards; it is not an interactiveleg <agent> terminal, because that needs a usage tap and a wall signal, which only the four above expose.
  • A leg that ends on a limit signal, a stall, a crash or exit 0 without .leg/DONE hands off with a bundle to the next adapter in the same worktree; aland station rebases, tests and fast-forwards trunk or bounces the card with the failure in the bundle.
  • Optional mirrors, off unless set in .env : OpenClaw Workboard (LEG_SYNC_WORKBOARD=1 ) and DashClaw (LEG_SYNC_DASHCLAW=1 ).

The full v0.1 story, with the fake-limit demo and the real claude→codex run, is in docs/concepts.md, docs/DEMO.md, docs/real-run.md and docs/board-guide.md.

Leg binds 127.0.0.1. leg share on is the supported way to listen anywhere else: it binds your Tailscale or LAN address and every human gets their own token (see More than one human). Without share, setting LEG_BIND to a non-loopback address needs LEG_TOKEN too, or the server refuses to start (exit 3), and requests then need Authorization: Bearer <token>. Tokenless owner access also requires a loopback hostname (127.0.0.1, localhost, or [::1]), which prevents a DNS-rebound hostname from inheriting local access. Either way there is no TLS.

  • The board did not open :leg open , or visithttp://127.0.0.1:4747 .~/.leg/board.log has the server's output.
  • claude's card shows "usage unknown" : Claude Code has no stored claude.ai login in that config directory (runclaude auth login ), the stored token expired (startclaude once, it refreshes), or the usage endpoint answered with something Leg does not recognise. The card says which. The wall is still caught through the hook; only the percentages are missing.
  • codex usage is unavailable or stale : the read-only Codex app-server quota request failed or has not completed in the last five minutes. The board retries every minute; an active Codex session also keeps its rollout tap as a fallback for percentages and the wall signal.
  • agy's card has no percentage : expected, agy exposes none. Leg sees the wall when agy hits it.
  • A session shows lost : the terminal that ran leg <agent> is gone (closed, crashed, machine slept through a kill). Remove it from the board.
  • Nested session :leg claude typed inside a Claude Code shell works; the parent'sCLAUDECODE markers are stripped so the child starts.
  • npm install dies with edgesOut (clone only): the global npm is older than Node; runnpx --yes npm@latest install once.

More in docs/faq.md.

guide read it when
Getting started you want leg claude running in five minutes
Concepts sessions, accounts, bundles, and the v0.1 cards, stations, chains and leases
Board guide every word, number and button on the board explained
Configuration environment variables and options
Portable harness carrying rules, hooks, skills, agents, commands and MCP servers to the agent a hand-off lands on: what moves, what does not, policies, ownership, secrets
History leg history andleg worktrees : every conversation and checkout on this machine across agents, the support matrix, what is read and what is written
Adapters what each CLI exposes and how Leg attaches to it
CLI contracts exact argv per CLI and the limit-signal table with sources
FAQ a question the others did not answer
Demo andreal run the v0.1 handoff, fake and real
Vocabulary statuses, outcomes and event types
Roadmap v2 where this is going
Reuse anddeviations what was ported and every place the plan changed
Website the public page: static HTML in site/ , preview withpython -m http.server 4780 --directory site , deployed to Vercel from that directory; PRODUCT.md and DESIGN.md at the root carry its brief and tokens

Issues and pull requests are welcome. Read CONTRIBUTING.md for the dev setup and the rules (zero runtime deps, argv spawns only, no bypass flags, a privacy check on every commit). Security reports go through SECURITY.md.

npm install
npm test          # node --test + privacy check
npm run lint

Any real agent session started only to test Leg runs on the cheapest model (leg claude --model haiku); the live checks in test/ never start one.

Maintainer releases use npm trusted publishing with no NPM_TOKEN. Bump the package, lockfile, site metadata and release notes, then push main. npm version (and npm run sync-alias) writes packages/leg-agents to the same version and pins @ucsandman/legcli to it; npm test fails if they drift. CI waits for the Ubuntu and Windows test matrix, validates both @ucsandman/legcli and leg-agents against npm, and publishes each only when that version is missing and newer than the stable latest. Existing versions skip cleanly; older, prerelease, lockstep, and registry-error cases fail the job. The npm trusted publisher is bound to ucsandman/legcli and .github/workflows/ci.yml (bind leg-agents the same way). Publication uses --provenance=false.

Parts of the runner, ledger and git snapshot were ported from a private repository that was MIT licensed (see NOTICE and docs/REUSE.md), with chat identifiers, machine paths and personal names removed. The test suite runs a privacy check on every commit, and the fixtures store home paths as ~.

Leg is commercial software under the Leg License Agreement. It ships as readable JavaScript so you can see what it does on your machine, and you may modify it for your own use, but not redistribute it or work around the license check. Versions 0.2.0 and 0.3.0 were published under MIT and remain available. The version in this source tree is 0.15.0; see npm for published versions and CHANGELOG.md for release notes.

The first 14 days are a free trial: install it, type leg claude, and everything in both plans is on, with no key and no card. The clock starts on your first session and leg license status shows what is left. After that it needs a license: Personal, $79 once, one human on any number of machines, every release for 12 months and the version you have keeps working after that; Team, $12 per seat per month, Personal plus leg share for more than one human on the board. Buy at the site, then leg license activate <key>. There is also a 30-day money-back guarantee after buying, no reason required. A key is a signed token checked offline with the public key in src/license.mjs; only a Team key renewal talks to the site. The bare agent CLIs are never affected by any of this; only what Leg adds is licensed.

── more in #ai-agents 4 stories · sorted by recency
── more on @legcli 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-claude-code-…] indexed:0 read:35min 2026-09-19 ·