AgentFleet

Native Claude Code subagents

A fleet of specialist agents that runs on any provider

AgentFleet installs 19 native Claude Code subagents ("lanes"), each named for a role — plan, implement, review, test — never a model. Point it at your Claude subscription, an Anthropic API key, or any Anthropic-compatible gateway, and every lane gets the best-fit model automatically.

One command installs and updates; a wizard configures everything.

v2.20.0 · November 7, 2026 · What's new

macOS

curl -fsSL https://agentfleet.wecansync.com/install.sh | sh

Linux

curl -fsSL https://agentfleet.wecansync.com/install.sh | sh

Windows

irm https://agentfleet.wecansync.com/install.ps1 | iex

Read it first: install.sh · install.ps1.

Read the docs View on GitHub 
$ curl -fsSL https://agentfleet.wecansync.com/install.sh | sh
Verifying SHA-256… ok
Installing under ~/.claude and ~/.local/bin — no root required.

How should Claude Code reach its models?
1. Claude subscription (default)
2. Anthropic API key
3. Custom gateway

Why AgentFleet

A team of specialists, not one all-powerful agent

Most standard AI tools act as a single, all-powerful agent that can easily make accidental changes and break your software. AgentFleet innovates by replacing this risky setup with a highly structured team of specialized AI assistants.

Planners read and strategize, while coders and testers handle implementation. Reviewers run checks to verify findings and edit only when the brief explicitly requests FIX MODE and lists the files they may change. The main agent integrates the work and owns outward-facing actions; the included lane-guard mod blocks covered destructive and publishing commands in fleet subagents.

Planning stays read-only

Planning, research, exploration, and triage lanes get no Bash, Edit, or Write tools by default. Diagnosis can run checks but never edits; reviewers verify findings in CHECK mode by default.

8 lanes make the changes

Implementation, tests, UI, and docs lanes edit code. No lane has the tool to spawn nested agents, and writable lanes are instructed never to commit, push, or deploy: the main agent integrates and owns every outward-facing action.

Features

Built to survive a provider change

Every design decision in AgentFleet exists so that switching providers, or switching models within one provider, never means rewriting your workflow.

Auto-compact by percentage

One percentage for every model: default 70% of usable context, configurable from 20–95% or off. Restart once after changing the percentage; model switches retain it without recalculation (gateway cap changes still need a restart). Configure auto-compact.

Reviewers that verify and fix

Review lanes run tests, linters, and builds to verify findings. They edit only in explicit FIX MODE with a file scope. Restrict their tools with reviewer-access preferences. See reviewer modes.

Lane guard

The default lane-guard mod blocks covered git mutations, GitHub writes, publishing, deployment, migrations, destructive SQL, and dangerous deletions in fleet subagents. It never blocks the main conversation. See guard limits.

Outage failover

A gateway 5xx failure marks a lane failing for 15 minutes. Routing avoids it and tells the main agent which alternative to use. Local lane reports record changed files and forbidden commands. See outage reports.

Project suggestions

/fleet-suggest scans locally for useful plugins, language servers, MCP-server plugins, mods, hooks, and project skills. It installs only what you pick. The band shows one new suggestion at a time and clears when you open /fleet-suggest. Learning counts command/tool names and doc sites, never prompt text; it is on for new installs and preserves existing settings. Explore suggestions.

Any provider

Works with your Claude subscription, an Anthropic API key, or any Anthropic-compatible gateway whose models are discovered from GET /v1/models.

Role-named lanes

Lane names describe a role — fleet-implement, fleet-review — never a model, so names and workflows survive provider changes.

Capability tiers

Every discovered model gets a tier — cheap, fast, balanced, deep — and each lane is ranked against live models by tier, with automatic fallbacks.

One-command switching

agentfleet use <profile> switches providers without reinstalling. Lanes keep their names; only the models behind them change.

Safe installs

Backups before every apply, one-command rollback, idempotent re-runs, and no overwriting of files it doesn't own.

Privacy-safe routing log

The routing hook logs the lane, reason, timestamp, and character count — never prompt text — to a mode-0600 file that purges itself.

Status band and optional mods

The included fleet-status mod draws a band above the prompt only when something needs action. Optional JEV compaction shows token savings, verbatim messages kept, time, or a fallback reason above the prompt; /jev shows session stats. It sends conversation text to api.typesafe.ai. The token-weather context forecast is on by default. See mods, requirements, and privacy.

One-line install, automatic updates

The same command installs or updates a checksum-verified release; a wizard configures unanswered options in a terminal. Same-major releases also update in the background without prompts. No root or admin rights, ever.

See all features

How it works

From one command to a ranked fleet

Install

One command downloads a checksum-verified release and runs the installer entirely under your home directory.

The wizard picks a provider

Choose a provider on first install. After terminal installs and updates, configure unanswered feature questions; agentfleet configure revisits them all.

Lanes are ranked by tier

At startup, every live model is matched against each lane's required capability tier, with the next-best models kept as fallbacks.

Claude Code uses lanes for large work

By default Claude works directly and routes large tasks to the matching lane by role. agentfleet config work-style direct|balanced|fleet changes how readily it delegates. Restart after a provider switch and the same lane names still apply.

Diagram: 19 lanes are matched to capability tiers, which are matched to live provider models 19 lanes fleet-plan fleet-implement fleet-review fleet-tests …and 15 more Tiers cheap fast balanced deep Live provider models Claude subscription Anthropic API key Custom gateway (GET /v1/models)

The fleet

19 lanes: 12 task lanes, 7 variants

Planners and research lanes are read-only by default. Reviewers run checks and fix only in explicit FIX MODE; diagnosis runs checks but never edits. No lane can spawn nested agents. The main agent owns integration and outward-facing actions, with lane-guard blocking covered forbidden commands in fleet subagents.

Task lanes

fleet-planRead-only

Planning and architecture

fleet-implementWritable

General implementation

fleet-reviewCheck / Fix

Code review

fleet-testsWritable

Test writing

fleet-uiWritable

UI/frontend work

fleet-docsWritable

Documentation

fleet-security-reviewCheck / Fix

Security review

fleet-diagnose-staticRun checks

Static diagnosis

fleet-explore-narrowRead-only

Bounded exploration

fleet-triage-staticRead-only

Issue triage

fleet-research-codebaseRead-only

Repository research (no web tools)

fleet-research-webRead-only

Web research (WebSearch, WebFetch)

Variant lanes

Named for a capability, never a model, so names survive provider changes.

fleet-plan-altRead-only

Alternate planning model

fleet-implement-altWritable

Alternate implementation model

fleet-implement-fastWritable

Low-latency implementation

fleet-implement-deepWritable

Reasoning-capable implementation

fleet-implement-cheapWritable

Budget or free-tier implementation

fleet-review-altCheck / Fix

Independent review model

fleet-review-deepCheck / Fix

Deep review: security, migration, concurrency, data-loss

Providers & profiles

Switch providers without reinstalling

Profiles save the current setup before switching, so switching back restores it exactly. Restart Claude Code after switching.

agentfleet profiles          # list saved profiles (* = active)
agentfleet use native        # switch to your Claude login
agentfleet use <name>        # switch to a saved gateway profile
agentfleet save <name>       # name the current setup

Gateway tokens are stored at ~/.claude/agentfleet/secrets/<name>.token (mode 0600). Switching to native removes gateway environment overrides and maps lanes back to the opus/sonnet/haiku aliases.

Safety & privacy

Nothing changes without a backup

Backups before every apply

All managed files are copied to a mode-0700 snapshot under ~/.claude/backups/claude-agents-config/ before any write.

Rollback

agentfleet rollback restores the newest backup (agentfleet rollback --backup DIR restores a named one). The backup is kept until you remove it.

Idempotent & non-destructive

Repeated installs produce the same configuration. Uninstall removes only files and hook entries this bundle owns.

No accidental overwrites

Existing files without the bundle's ownership marker are never overwritten without --force-owned.

Credential safety

The scanner checks for credential-shaped literals. The gateway token is the only secret normally needed.

Privacy-safe routing log

Only a no-match entry keeps a redacted, max-160-character excerpt. Text is purged after 14 days, entries after 90.

Requirements

What you need first

FAQ

Frequently asked questions

Does it work with my Claude subscription?

Yes. It's the default provider: the wizard's first choice uses Claude Code's own login, and lanes map to the opus, sonnet, and haiku aliases.

What does "any provider" actually mean?

An Anthropic API key, or any Anthropic-compatible gateway whose models are discovered from GET /v1/models. Catalogs report context size but not price or speed, which is why explicit tier labels exist.

Can I pin a lane to a specific model?

Yes: claude-fleet-setup --prefer fleet-implement=my-model-id. Clear a pin with claude-fleet-setup --prefer fleet-implement=. A lane with an explicit preferred list respects your order.

How do I roll back or uninstall?

agentfleet rollback restores the newest backup. agentfleet uninstall removes only the files and hook entries this bundle owns and restores earlier settings.

Where do my own Claude Code instructions go?

In ~/.claude/CLAUDE.md or your own files under ~/.claude/rules/. AgentFleet keeps its policy in ~/.claude/rules/agentfleet.md and never edits your CLAUDE.md. agentfleet instructions checks your instruction files for stale lane names, duplicates, and contradictions, and applies an edit only when you say yes.

Does updating re-run the setup wizard?

Yes: after a successful terminal install or update, the configuration wizard asks only unanswered questions; Enter keeps the current value. Your provider profile is preserved. Background updates never prompt. Skip configuration with AGENTFLEET_NO_CONFIGURE=1 or --no-configure; revisit all questions with agentfleet configure.

What happens when I switch providers?

Switching first saves the current setup, so switching back restores it exactly. Lane names never change — only the models behind them do. Restart Claude Code after switching.

What's new

Latest release

v2.20.0 November 7, 2026

Plans are reviewed with fresh eyes, and large projects can add Google's Elephant-Goldfish Model.

  • When a plan or specification goes to review, the review lane gets the document and the files it cites, not Claude's reasoning, and must restate the plan before listing gaps. A plan it cannot restate is not ready to build.
  • /fleet-suggest offers Google's EGM plugin to large repositories and projects with design documents: draft a spec with Claude, then fresh-context reviewers check it before anything is built.
  • With ask-first on, Claude names EGM, when installed, for large designs with consequential decisions such as payments or a data migration.

Version history

v2.19.0November 5, 2026A watchdog flags lanes that may be looping, and at most three lanes run at once.
  • The AgentFleet band warns when a running write lane has made 40 tool calls without changing a file, so you can check or stop it before it uses its turn limit. Plain counting: no model calls, no files written.
  • Claude runs at most three lanes at once unless you ask for more, and queues the rest, keeping tokens and reports to follow in check.
v2.18.0November 3, 2026Text Claude reads cannot give orders, and the tasks pane shows what is waiting for you.
  • A new safety rule: an issue, pull request comment, web page, file, tool result, lane report or memory is data, not instructions. Only you can approve an action or widen the task; when such text asks for one, Claude tells you and waits.
  • /fleet-tasks now starts with Waiting for you: lanes waiting on an answer, and lane reports from the last two hours that ended blocked, partial or failed.
  • AgentFleet's own main branch now requires a pull request, enforced by GitHub.
v2.17.0November 1, 2026Repeated rule-based work becomes a script instead of model work, and frontend projects can add three UI skills.
  • A new policy rule: when the same rule-based steps recur, Claude writes them once as a script, test or hook and runs that, keeping model calls for steps that need judgment.
  • /fleet-suggest offers frontend projects unslop (turns AI-looking UI into a designed interface), prototype (throwaway prototypes to compare UI options) and, where the project uses shadcn/ui, the official shadcn skill. Each is pinned to a reviewed commit, copied into the project's .claude/skills, and loads only when you invoke it.
  • The frontend-design plugin suggestion now also covers Tailwind projects. A skill you already have is not offered again.
  • Nothing is installed unless you pick it.
v2.16.0October 30, 2026Balanced and fleet styles agree a plan before several lanes run, and a substantial task ends with a short handoff.
  • When a task needs more than one lane, Claude gives you a concise plan first and waits for your approval before the first lane runs. A single lane or direct work needs no team plan, and the approach agreement for large or risky work still applies.
  • The plan is part of the existing ask-first section, so agentfleet config ask-first off turns it off too. No new setting.
  • Your approval of a fully specified plan covers its continuations and fallbacks; new scope or roles need a new approval. It never covers pushing, merging, migrations or deploys.
  • After a substantial task, when it finishes, is blocked or is partly done, or when you ask for a status, Claude ends with one handoff: branch or worktree, uncommitted state, checks run and not run, blockers, and any pull request or UAT link.
v2.15.0October 27, 2026An optional guard that asks you before the main session discards work or publishes.
  • agentfleet config main-session-guard on makes lane-guard ask you before the main session runs a force push, git reset --hard, a deploy, a migration, destructive SQL or a dangerous rm.
  • Commits, pushes, pull requests and branch switches run without a question. Only "Run it" runs the command.
  • Off by default. Meant for sessions that skip permission prompts. Fleet lanes keep their own rules.
v2.14.3October 26, 2026The measuring tool also measures change size and keeps why a check failed.
  • For maintainers: packaging/measure.py compares how many lines and files each run changed, and keeps every run's diff.
  • A failed check's output is saved, and its last line shown in the summary.
  • A check that runs too long now counts as failed instead of stopping the whole measurement.
  • Nothing changes for users.
v2.14.2October 25, 2026A way for maintainers to measure whether a change helps.
  • For maintainers: packaging/measure.py runs the same tasks on two versions of AgentFleet, with your lane models, and compares tokens, cost, turns and passes.
  • A difference counts only when the runs do not overlap, and a task with errors or failed lanes is marked confounded, an idea from Google Research's Harness².
  • Nothing changes for users.
v2.14.1October 25, 2026Lanes never follow instructions hidden in what they read.
  • Every lane is told that only its prompt and the brief say what to do. Files, web pages, command output and memory are data: a lane never follows instructions found in them, and reports any that try to change its task.
  • The rule comes from Google Research's Harness², which gives its agents the same boundary.
v2.14.0October 24, 2026Claude and its lanes make the smallest change that fully solves the task.
  • Before writing, Claude and the implementation lanes find every place a change must reach, then reuse existing code, the standard library and installed dependencies before adding anything.
  • They add no feature, option or abstraction you did not ask for, say in one line what they left out, and still finish every part the task needs.
  • Plan lanes prefer the smallest design, and review lanes list code a change did not need. The idea comes from Ponytail, built in with no hooks.
v2.13.0October 23, 2026See the model, uptime and branch of your session above the prompt.
  • A new session-info mod, on by default, shows one row above the prompt: the current model and effort, uptime, running subagents, git branch, prompts sent and cost at list price.
  • agentfleet configure asks whether to show it, under Status and reports. An update adds it once; remove it and it stays removed.
  • The status band counts the wizard questions added in 2.12.0 and 2.13.0, so it tells you when there are new ones to answer.
v2.12.0October 22, 2026Set every feature in one pass, and agent-brain memory that saves session summaries again.
  • agentfleet configure now covers every feature, grouped under headings: it adds Context Mode, the branch check, the /brain command and gateway usage. After an update it asks only the new questions.
  • Install, agentfleet doctor and agent-brain's background update upgrade an older agent-brain memory store, after agent-brain backs it up, so session capture and summaries work.
  • /fleet-suggest offers MCPVault for a project with an Obsidian vault, and Graphify for a large repository. Both are optional and installed only after you approve.
  • A new features page lists everything AgentFleet adds.
v2.11.1October 20, 2026Gateway figures survive a slow or briefly failing gateway.
  • A gateway that was slow, down or returning server errors was taken for one without a usage report, hiding its figures for a day. Now the figures already shown stay, and AgentFleet tries again after 5 minutes.
  • Each request to the gateway waits up to 8 seconds, since a first answer after a quiet spell can be slow.
  • agentfleet config says when the gateway hasn't been reached yet, instead of reporting no usage.
v2.11.0October 20, 2026See what your gateway reports you spent, and get warned before you hit a limit.
  • On OmniRoute, OpenRouter and LiteLLM gateways, agentfleet report and /brain show the spend, budget and limits the gateway reports for your key.
  • The status band warns when a budget or limit passes 80%, when a provider's accounts are exhausted, and when a Claude subscription's 5-hour or weekly limit passes 80%.
  • Prices a gateway publishes in its model list are used for estimated cost without typing them, and lane runs are priced by the model you configured.
v2.10.2October 20, 2026Hooks that keep working after a Node.js upgrade.
  • Hooks now use the Node.js path on your PATH, such as Homebrew's /opt/homebrew/bin/node, instead of the versioned folder behind it, which an upgrade deletes.
  • agentfleet doctor reports a hook whose Node.js or Python no longer exists, and how to fix it.
  • agentfleet config context-mode on|off succeeds when Context Mode is already in that state, and checks the Node.js Context Mode actually runs before enabling it.
v2.10.1October 20, 2026Free agent-brain memory, on this machine.
  • Setup asks once whether to turn on agent-brain memory: Claude saves decisions and project context on this machine and reads them back in later sessions. It's free and needs no account.
  • agentfleet config agent-brain-memory on|off changes it later.
  • Lanes get the memory search tool only when memory is on.
v2.10.0October 19, 2026See how each lane and model performs, and what it costs.
  • Lanes start their report with an outcome: done, partial, blocked or failed. agentfleet report shows each lane's and model's runs, outcomes, cut-off runs, errors and tokens, plus the main conversation's tokens.
  • Optional model prices: agentfleet config prices MODEL=INPUT,OUTPUT, in dollars per million tokens, or a prices table in a team preset. agentfleet report and /brain then show an estimated cost.
  • Claude continues a lane that reports partial or blocked instead of starting a new one, and stops a lane before redoing its task.
v2.9.0October 18, 2026Optional Context Mode pilot.
  • Enable the upstream plugin with agentfleet config context-mode on. It is off by default; no setup question is added.
  • lane-guard permits Context Mode search and stats in lanes, but blocks its other tools. The main conversation is unaffected.
  • Measure total tokens, time and correctness before keeping it enabled. No measured session savings are claimed.
v2.8.3October 16, 2026A cd no longer splits a session's project.
  • Usage learning and suggestions stay with the folder Claude Code started in.
  • The daily update installs agent-brain's session-handle fix.
v2.8.2October 15, 2026Updates keep the effort you chose for a lane.
  • An effort set with claude-fleet-setup --effort or a team preset was reset to the default by every update. It now stays.
v2.8.1October 15, 2026agent-brain works on Windows.
  • When the daily update installs a newer agent-brain, AgentFleet lets it rewrite its Claude Code hooks, so Windows machines get agent-brain 0.5.5's fixed hooks without any command.
  • A check script for Windows machines: irm https://agentfleet.wecansync.com/windows-check.ps1 -OutFile windows-check.ps1.
v2.8.0October 15, 2026Lanes cost fewer tokens.
  • Claude reviews lane work itself; a review lane runs only for security, data-loss or migration changes, or when you ask.
  • Fixes go back to the lane that wrote the code, which keeps what it has read, and a failed lane is re-dispatched only once.
  • Lanes get a warning near their turn limit, so they report instead of being cut off.
  • When the gateway itself is failing, Claude works directly instead of retrying lanes.
  • Teams can pin each lane's models and effort in a preset, and reserve an expensive model for the lanes that name it.
v2.7.2October 15, 2026Updates now reach your installed mods.
  • Claude Code kept running each mod as it was first installed, so mod fixes never arrived. An update now updates each installed mod whose version changed and says which; restart Claude Code to load them. Disabled mods stay disabled.
v2.7.1October 15, 2026/brain shows your memories, in a pane you can scroll.
  • /brain opens a pane with usage at the top and the project's memories below, critical ones first. Arrow keys move, Enter shows a memory in full.
  • agent-brain 0.5.2 lists memories 50 at a time, so a long list no longer fills Claude's context.
v2.7.0October 15, 2026A /brain command for agent-brain, and no more cloud question at the start of a session.
  • The new brain mod adds /brain: token usage for this session, today and this week, and the project's latest memories when agent-brain memory is active.
  • agent-brain 0.5.1 no longer asks at the start of a session whether to link the project to agent-brain cloud. agent-brain config link-offer on brings the question back.
v2.6.2October 14, 2026Installing over an existing gateway setup now fixes it instead of stopping.
  • Mismatched auto-compact settings are set to one value instead of stopping the install.
  • Model settings the gateway does not serve are replaced with models it does, and the installer says which.
  • Every gateway install fetches the gateway's models at startup, and /fleet-setup fetches them itself when it checks your settings.
v2.6.1October 14, 2026The setup wizard installs and connects the free agent-brain only.
  • The wizard no longer asks you to sign in to agent-brain or turn on its memory; it installs the free tool, which records the tokens and cost of each session, and connects it to Claude Code.
v2.6.0October 14, 2026Claude checks the git branch with you before changing files, and the setup wizard installs agent-brain and keeps it up to date.
  • Before changing files in a git repository, Claude shows the current branch, uncommitted changes and whether it is behind its remote, then asks whether to continue, switch, or start a new branch, with its recommendation. No more work landing on last task's branch. Turn it off with agentfleet config branch-check off.
  • The wizard offers to install agent-brain, which records the tokens and cost of each session. The download is checked against the release manifest before it is installed.
  • Automatic updates now keep agent-brain current too, once a day, never to an older version.
  • agent-brain writes its own Claude Code hooks; AgentFleet no longer keeps a second copy.
v2.5.0October 13, 2026Claude checks with you before big or unclear tasks, lanes keep working when a model runs out, and everything uses fewer tokens.
  • When a lane's model hits its quota or fails, the lane moves to the next of its fallback models instead of stopping. You can pin your own chain, such as Haiku, then Luna, then Gemini.
  • Claude decides whether to split work across lanes after reading the code, not from how long your prompt is, and tells you when it splits.
  • Lanes no longer spend turns saving memory, and keep their reports short.
  • Before a large, unclear or risky task, Claude reads the project's docs, investigates, and agrees the approach with you: Spec Kit or a plan, worktrees for several tasks, and the structure of a new project. Small, clear tasks are still done straight away.
  • The token-weather context forecast is on by default.
v2.4.3October 13, 2026Cleaner learned suggestions in /fleet-suggest.
  • A learned skill is suggested once, for the command you run most, instead of once per similar command.
  • No more skill suggestions for general-purpose tools such as git, head or tail.
  • python3 -m pytest suggests a pytest skill, even when options come before -m.
v2.4.2October 13, 2026/fleet-suggest now helps set up how Claude works in each project.
  • For a GitHub repository, it drafts a "Git workflow" section for the project's CLAUDE.md: a branch and pull request per feature, the test command to run first, and no commits, pushes or deploys unless asked. It is written only after you approve.
  • For a new or small project, it offers Spec Kit, which writes a spec, a plan and a task list before code.
  • agentfleet config shows agent-brain: not installed when the agent-brain CLI is missing.
v2.4.1October 13, 2026Fixes the instruction check added in 2.4.0.
  • agentfleet instructions, and the instruction check offered by agentfleet configure, no longer stop with a Python error before checking anything.
v2.4.0October 13, 2026Claude works directly by default and uses lanes for large work. AgentFleet's policy no longer lives in your CLAUDE.md.
  • Work style: agentfleet config work-style direct|balanced|fleet. The default, balanced, suggests a lane only for large prompts; direct uses lanes only when you ask; fleet is the 2.3 behavior. Restart Claude Code after changing it.
  • AgentFleet's rules moved to ~/.claude/rules/agentfleet.md and shrank from about 1,400 to about 350 words. Your ~/.claude/CLAUDE.md is no longer touched, so a first install over a personal one works. Upgrades restore an original that an earlier install replaced.
  • agentfleet instructions [--project DIR] [--check] finds stale lane names, duplicates, and contradictions in your instruction files. Each edit needs your approval and gets a backup. It sends those files' text to your model provider; --check only reports.
  • Lighter lanes: effort is high or medium instead of xhigh, except plan and the -deep lanes. Lanes now stop and report what is done and what remains instead of running out of turns silently.
  • agentfleet config agent-brain off removes the agent-brain memory step after every reply and lane run. The wizard asks about work style, agent-brain, and the instruction check; fleet-status 1.3.0 shows your work style.
v2.3.1October 12, 2026Tested on real Windows: installs, updates, rollbacks and the command wrappers now work there end to end.
  • agentfleet update, rollback and uninstall no longer fail with "The batch file cannot be found" when they replace the agentfleet.cmd wrapper they run from.
  • Wrappers keep non-ASCII paths such as C:\Users\José intact and restore the console's code page afterwards.
  • A crashed run's lock no longer blocks the model catalog, agent generator or suggestions on Windows, and a live lock is never taken over.
  • Rollback restores over read-only files, and claude-fleet-setup creates agent files for new lanes on Windows.
  • On every platform, rerunning or updating a 2.3.0 install keeps learning on; only installs older than 2.3.0 start with it off.
v2.3.0October 12, 2026One command installs and updates, a wizard configures every feature, and team presets share settings without secrets.
  • agentfleet configure revisits all questions; terminal installs and updates ask only unanswered ones. Background updates never prompt.
  • Export with agentfleet configure --export team.json; teammates use AGENTFLEET_PRESET=team.json with the install command or with AGENTFLEET_PRESET=team.json agentfleet update. Each person supplies their own TypeSafe key. Team setup.
  • One auto-compact percentage for every model: default 70%, range 20–95 or off. Restart once after a percentage change; model switches retain it without recalculation, with a restart only for gateway cap changes. Low percentages now work on smaller models.
  • Optional JEV shows before/after tokens, reduction, verbatim messages kept, elapsed time, or a fallback reason. /jev shows session totals and the last five compactions. Conversation text goes to api.typesafe.ai.
  • The suggestion band shows one new item at a time, clears on /fleet-suggest, and remembers Hide. Learning is on for new installs; existing settings are preserved. It never records prompt text.
  • fleet-status 1.2.1 adds options and JEV lines; fast-jev-compaction is 0.3.0-agentfleet.3. agentfleet config shows every feature.
v2.2.0October 11, 2026Model-aware auto-compact, reviewers that verify and fix, guarded lanes, outage failover, and local project suggestions.
  • Reviewers run checks in CHECK mode by default and edit only in explicit FIX MODE with listed files. agentfleet config reviewer-access full|run|read-only controls their tools.
  • The included lane-guard mod blocks covered destructive and publishing commands in fleet subagents, never the main conversation.
  • Gateway 5xx failures mark lanes failing for 15 minutes. Routing suggests an alternative; local lane reports track changed files and forbidden commands.
  • /fleet-suggest scans locally and installs only what you pick, from a 57-entry catalog of plugins, mods, hooks, project skills, and /init.
  • Project-local lane overrides, a usage report, plugin and MCP health checks, and hardened Windows shims.
v2.1.0October 9, 2026Claude Code mods: a status band when something needs action, with optional context forecasts and Jev compaction in a local plugin catalog.
  • fleet-status is installed and on by default. It shows restart notices, failed automatic updates, lanes without a model, and model changes waiting for review; /fleet-status prints the full status.
  • The agentfleet catalog includes two opt-in mods: Anthropic's token-weather and tamaratran's fast-jev-compaction. Jev compaction needs a paid TypeSafe API key and sends conversation text and tool inputs to TypeSafe.
  • Mods need Claude Code 2.1.287 or later and show in the terminal and Desktop's Code tab. fleet-status reads only four AgentFleet files, never settings or tokens, and sends nothing anywhere.
  • Opt out with --no-mods or AGENTFLEET_MODS=0; --mods opts back in. Updates refresh mod files without re-registering disabled or uninstalled plugins. Read the mods guide.
v2.0.6September 25, 2026Windows fixes: the one-command install works on the PowerShell built into Windows, and updates repair hooks an older install left behind.
  • irm https://agentfleet.wecansync.com/install.ps1 | iex works on Windows PowerShell 5.1. It had stopped on the Microsoft Store python3 placeholder and wrongly reported Node.js 18+ as missing. agentfleet update and automatic updates on Windows are fixed too.
  • The Windows installer finds a Python installed with uv python install even when it is not on PATH.
  • agentfleet, claude-agents-doctor, and claude-fleet-setup on Windows run with the Python the installer used, instead of failing with "Python was not found".
  • Updates refresh hooks that differ only in quoting or comment style, and never leave a hook pointing at a script they just removed.
v2.0.5September 24, 2026Any number of provider profiles, each with its own settings.
  • New agentfleet add NAME --gateway-url URL and agentfleet remove NAME add and remove gateway profiles without reinstalling.
  • Each profile keeps its own provider policy, model pins, exclusions, and tier labels, instead of sharing one policy across every profile.
  • Gateway tokens are handled more safely: no redirect leaks, a digest instead of plaintext in the install journal, and no live token sent to a new URL during a dry run.
  • agentfleet update (including automatic updates) now succeeds after agentfleet use native.
  • Uninstall keeps the API-key mask for a gateway you switched to and restores a pre-install gateway exactly.
v2.0.4September 23, 2026Model-to-lane interview in /fleet-setup, automatic updates.
  • /fleet-setup suggests a complete fleet from each model's description, tier, and context size, or walks you through it model by model.
  • claude-fleet-setup --exclude/--include MODEL keeps a model out of every lane while leaving it in the /model picker.
  • Automatic updates: a same-major-version release installs itself in the background at most once a day. Turn it off with agentfleet auto-update off.
  • A local .claude/ directory no longer breaks installs made from a git clone.
v2.0.3September 23, 2026Redesigned website with full documentation, plus a missing CLI command.
  • Redesigned landing page and a full documentation site at /docs/.
  • agentfleet rollback [--backup DIR], which the README already documented but which didn't exist yet.
  • Docs now correctly say that writable lanes are only instructed, not technically prevented, from committing, pushing, or deploying.
v2.0.2September 23, 2026Fixed agentfleet update, which had failed for everyone.
  • agentfleet update works again: Cloudflare had rejected the default Python user agent before the request reached the server.
  • An update now removes preference values 2.0.1 stopped shipping, but only if you never changed them yourself.
v2.0.1September 23, 2026Dropped shipped lane pins, neutral defaults, a doctor fix.
  • Updating from 1.0.0 no longer carries over the old shipped model pins as if you had chosen them yourself.
  • CLAUDE.md and settings no longer include the maintainer's personal preferences; put your own in ~/.claude/rules/*.md.
  • agentfleet doctor now covers all 19 lanes and flags any custom lane marked read-only.
v2.0.0September 23, 2026The project becomes AgentFleet: any provider, role-named lanes, profiles.
  • Any provider: a Claude subscription, an Anthropic API key, or any gateway that serves GET /v1/models.
  • 19 role-named lanes (fleet-plan, fleet-implement, fleet-review, ...) replace the 30 lanes that used to be named after models.
  • Tier-based model assignment with automatic fallbacks, a first-run wizard, and provider profiles (agentfleet use native|<name>).
  • A one-command install that verifies a SHA-256 checksum.
v1.0.0September 18, 2026The portable Claude agents configuration, before the AgentFleet name.
  • Initial release: a portable Claude Code agents configuration installed entirely under ~/.claude and ~/.local/bin.
  • Provider-aware fleet repair: model discovery, advisory fallback models, and reconciliation as the gateway's model list changes.
  • A /fleet-setup skill, automatic context-window sync on model switch, and Windows UTF-8 output fixes.

Full changelog on GitHub

Install AgentFleet in one command

No root, no admin rights, checksum-verified releases.

Get started View on GitHub