feat(mcp): server management commands, self/atmos-pro presets, and config UX fixes#2720
Open
Erik Osterman (Cloud Posse) (osterman) wants to merge 45 commits into
Open
feat(mcp): server management commands, self/atmos-pro presets, and config UX fixes#2720Erik Osterman (Cloud Posse) (osterman) wants to merge 45 commits into
Erik Osterman (Cloud Posse) (osterman) wants to merge 45 commits into
Conversation
Extracts the standalone `atmos mcp install` command and its supporting MCP client/installer library changes from the stalled Atmos Pro install branch (osterman/pro-install-cmd), leaving behind the Pro-specific `--mcp` wiring on `atmos pro install` so this can ship independently. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Fills the missing Docusaurus page for the atmos mcp install command, matching the existing mcp/export and mcp/list command doc conventions. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds a top-level MCP overview page (following the /pro and /ai landing page precedent) documenting Atmos as an MCP client -- connecting to external MCP servers and installing them into AI coding assistants -- and cross-linking to the existing MCP server (Atmos-as-server) docs. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Restructures the /ai page's MCP section into Client/Server tabs that both link to the dedicated /mcp page, and expands /mcp itself from a client-only page into the actual unified hub the tabs point to (adds an "As a Server" section alongside the existing client content). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Captures the design for signing requests to AWS's hosted HTTP MCP servers natively via Atmos Auth identities, so the mcp-proxy-for-aws third-party proxy is no longer required. Proposed, not yet implemented -- open questions flagged for a follow-up design pass. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Verifier binaries downloaded during toolchain installs can carry the com.apple.quarantine/com.apple.provenance xattrs, causing Gatekeeper to block execution even though the binary was verified during download. Strip the xattrs and ad-hoc re-sign on darwin so the verifier can run; add an opt-out via verifier_trust: disabled for environments that need one. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A leaf command (no subcommands) whose Args validator rejects extra/ missing positional arguments (e.g. `atmos config unset mcp.enabled true` against ExactArgs(1)) previously reported the misleading "Unknown command mcp.enabled" -- implying an unrecognized subcommand, when the real problem is too many arguments. Cobra's Find() only stops descending when no child command name matches, so for a leaf command a failed Args validator can only mean the arguments are wrong, never an unrecognized subcommand; surface Cobra's own validator message instead. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…relative paths atmos config set previously always wrote strings unless --type was passed explicitly, so `atmos config set mcp.enabled true` silently wrote the string "true" instead of a boolean. Infer the type from the Atmos config schema by reflecting on yaml/mapstructure struct tags when the dot-path matches a known field, falling back to string for unmodeled paths (e.g. vars.*). Also echo the value that was written and the deleted/formatted file, and show paths relative to the current directory instead of absolute. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Round out MCP client management with atmos mcp add/remove (declare servers in atmos.yaml) and atmos mcp install/uninstall (push/pull to AI client config files), plus built-in self and atmos-pro presets so `atmos mcp add` works with zero arguments. Adds an interactive prompt to enable mcp.enabled when adding the self preset, and a nudge suggesting atmos-pro when Atmos Pro is configured but not yet added. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Erik Osterman (Cloud Posse) (osterman)
requested a review
from a team
as a code owner
Contributor
|
Tip Atmos Pro
No affected stacks workflow was detected for this pull request. 
|
Erik Osterman (Cloud Posse) (osterman)
added
the
minor
Dependency Review✅ No vulnerabilities or license issues found.Scanned Files
|
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Contributor
|
Warning Review limit reached
Next review available in: 11 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (62)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Resource Changes Found for
|
Codecov Report❌ Patch coverage is ❌ Your patch check has failed because the patch coverage (81.10%) is below the target coverage (85.00%). You can increase the patch coverage or adjust the target coverage. Additional details and impacted files@@ Coverage Diff @@
## main #2720 +/- ##
==========================================
+ Coverage 81.61% 81.66% +0.04%
==========================================
Files 1639 1710 +71
Lines 154739 161092 +6353
==========================================
+ Hits 126294 131548 +5254
- Misses 21527 22350 +823
- Partials 6918 7194 +276
Flags with carried forward coverage won't be shown. Click here to find out more.
🚀 New features to boost your workflow:
|
…CP client commands
atmos stack config {get,set,delete,format,list} and atmos vendor config
{get,set,delete,format,list} reuse the DisplayPath/type-inference machinery
from atmos config set. pkg/yaml/edit.go gains the shared read/write/delete
helpers both command families and their AI-tool counterparts build on.
Also continues the atmos mcp add/install/uninstall command work: registry
lookups, install/uninstall symmetry, and config validation fixes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…om ai.tools.enabled
atmos ai chat failed with an Anthropic API-key error even when a CLI provider
(claude-code/codex-cli/copilot-cli/gemini-cli) was already installed and
authenticated. ai.GetProvider now probes PATH for one before falling back to
anthropic, so ai.enabled: true alone is enough to start using an existing
subscription -- explicit default_provider/--provider still always wins.
atmos mcp start also required ai.tools.enabled on top of its own mcp.enabled
opt-in, which was pure friction: MCP's job is exposing tools to external
clients, unrelated to atmos ai chat's own tool-use loop. Tools now always
register for MCP once mcp.enabled: true.
Renamed ai.tools.{allowed_tools,restricted_tools,blocked_tools} to
{allowed,restricted,blocked} (no back-compat shim) and gave `allowed` a
second job: when non-empty it now also gates which tools are registered at
all, not just whether they skip the confirmation prompt.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Expands the Atmos AI/MCP tool surface with atmos_config_*, atmos_stack_config_*, atmos_vendor_config_*, atmos_auth_whoami, atmos_auth_list, atmos_list_commands, atmos_command_help, atmos_describe_dependents, atmos_describe_workflows, atmos_list_workflows, atmos_list_components, atmos_list_values, atmos_toolchain_list, atmos_secret_list, and atmos_validate_component -- mirroring the corresponding CLI commands so an AI assistant or MCP client can introspect and manage config, stacks, vendoring, auth, and workflows without shelling out. Also adds registration-time allow-list filtering (isToolAllowed) to RegisterTools/registerAll: when ai.tools.allowed is non-empty, only matching tools are registered at all, not just exempted from confirmation. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds a welcome flow for new/empty Atmos projects (chat_welcome.go) instead of just reporting "no stacks found" -- the executor's system prompt gets the same treatment so atmos ai exec offers to scaffold a first stack/component too. Also refines session/provider handling and turn-step rendering in the chat TUI. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Wraps the --failure-mode/--log-order validation errors as static sentinels (err113) and extracts the "plan"/"destroy" subcommand string literals to named constants (revive) -- unrelated to this branch's other work, but blocking on push since the pre-push hook lints the full repo. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
Warning This PR exceeds the recommended limit of 10,000 lines.Large PRs are difficult to review and may be rejected due to their size. Please verify that this PR does not address multiple issues. |
|
💥 This pull request now has conflicts. Could you fix it Erik Osterman (Cloud Posse) (@osterman)? 🙏 |
`atmos ai skill install --all` didn't exist and there was no way to install every bundled skill in one shot short of looping `atmos ai skill install <name>` per skill, or fetching the whole cloudposse/atmos repo over the network. Rather than bolt on a separate --all flag, this mirrors atmos mcp install/uninstall's existing convention: omitting the positional argument means "act on everything" (mcp install/uninstall already do this for configured servers). `atmos ai skill install` with no <source> now installs every skill in the embedded catalog (single confirmation, not one per skill, skipping already-installed skills unless --force). `atmos ai skill uninstall` with no <name> mirrors this to remove every installed skill. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
… comma list `atmos ai skill install`/`uninstall` (and multi-skill package install) print a "Discovered N skills: a, b, c, ..." summary. With dozens of bundled skills this comma-joined string got routed through the same markdown word-wrap used for regular prose messages, which doesn't treat hyphens as safe break points and produced garbled output like "atmos-\naws-ecr," mid-identifier. Adds ui.FormatColumns(), a small ls-style column-major grid layout (evenly spaced, sized to the terminal width, falling back to 80 columns when no real width is detected), and a shared printItemColumns() helper that writes the header via the normal icon/markdown path but the grid itself as plain text so it isn't reflowed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Resolves a conflict in cmd/root.go between main's refactor of initCobraConfig/SetUsageFunc/SetHelpFunc into named helper functions and this branch's leaf-command arg-count error handling, by applying the leaf-command special case inside main's new rootUsageFunc.
…teractive
atmos mcp install/uninstall already had --scope project|user and --global,
but silently defaulted to project whenever neither was passed -- it never
asked. resolveInstallScope now returns (string, error) and, when running
interactively (real TTY, not --yes, not CI) with neither flag explicitly set,
shows a single-choice scope picker before the existing client picker,
mirroring promptForMCPClients's form/theme/cancel-handling. Also fixes
resolveInstallScope not checking cmd.Flags().Changed("global") -- it
previously read the bool value directly, unable to distinguish "not passed"
from "explicitly passed as false".
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…/--global Fixes a real incident: running atmos ai skill install/uninstall from this repo's own root wrote through (install) and deleted (uninstall) its intentional .claude/skills/<name> symlinks, which point into agent-skills/skills/<name> for contributor auto-discovery (per docs/prd/atmos-agent-skills.md). distributeToClients/removeClientCopies now check isSymlink() on the target before writing or deleting, skipping that client with a warning instead -- matching the existing refuse-to-delete pattern at pkg/terraform/clean/deleter.go and the skip-symlink-on-copy pattern at internal/exec/copy_glob.go. The same unguarded os.RemoveAll existed on the canonical --force reinstall path (prepareInstallPath); it now returns the existing ErrRefuseDeleteSymbolicLink sentinel instead of silently unlinking whatever's there. Also adds --scope project|user and --global to atmos ai skill install/uninstall, mirroring atmos mcp install/uninstall, since Claude Code, Gemini CLI, and GitHub Copilot all document both a project-local and a personal/user-level skill directory. Distribution now resolves scope-aware paths (~/.claude/skills, ~/.copilot/skills for the vscode client -- distinct from its project-level .github/skills, ~/.gemini/skills), and when neither flag is given interactively, prompts to choose (mirrors the just-added MCP scope prompt); a non-TTY session, --yes/--force, or CI still defaults to project with no prompt. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Separates the error block from whatever was printed before it (e.g. output from the command that failed), instead of running directly into it with no visual break. Regenerated the affected CLI snapshot goldens to match.
…cope prompt atmos ai skill install (installing the whole bundled catalog, or --force against an already-fully-installed catalog) printed 2 lines per skill (per-client distribution + per-skill success) -- ~94 lines for 47 skills. distributeToClients now returns which clients it actually distributed to instead of printing per call; batch installers (InstallAllBundled, installMultiSkillPackage) resolve and announce the client list once up front instead of once per skill, and no longer print a redundant per-skill success line -- only failures are worth a line, the batch summary already covers success. Also stopped printing the "switch skills with Ctrl+A" usage hint when zero skills were actually installed (e.g. everything was already installed and skipped). Also fixes atmos ai skill install --force silently skipping the new scope prompt: resolveSkillScope read "yes"/"force" directly from viper, ignoring which command it belonged to -- but install's --force means "reinstall", not "skip prompts" (that's --yes), while uninstall has no separate --yes so --force is its only skip signal. resolveSkillScope now takes an explicit skipPrompt bool, mirroring resolveSkillClients's existing pattern, with each caller passing its own correct flag. Also adds backtick code-formatting to skill/client names, flags, and paths in the affected toast messages, since they render as styled inline code rather than plain text. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
… and what changed
atmos ai skill install is always run from a plain shell, never from inside
atmos ai chat, so "Usage: Run atmos ai chat and switch skills with Ctrl+A"
never made sense there -- removed entirely (single-skill and batch install
paths).
The batch summary ("N skills installed successfully") also didn't say where
skills landed, and collapsed "freshly installed", "--force reinstalled an
existing skill", and "skipped because already installed" into a single
opaque count. installOneBundledSkill/installOneSkillFromPackage now report a
three-way outcome (installed/updated/skipped) instead of a bool, and the
batch summary reads accordingly:
47 skills installed successfully
Location: ~/.atmos/skills
47 skills updated successfully (all were --force reinstalls)
Location: ~/.atmos/skills
47 skills installed successfully (46 new, 1 updated)
Location: ~/.atmos/skills
0 skills installed
47 already installed (use --force to reinstall) (no Location: nothing landed)
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…xed width resolveFormatterWidth() computed the real detected terminal width but never returned it -- it always used the configured max_width verbatim whenever set, even when the actual terminal was narrower. With this repo's atmos.yaml setting max_width: 120, every error's padded content (including the title pill) was rendered at exactly 120 columns regardless of the real terminal width, so in any narrower terminal it wrapped onto extra visual lines, making the gap between the title and message look much bigger than intended. Now resolves to min(configured_max, detected_width), matching max_width's documented "maximum" semantics. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…e a spinner
atmos ai skill uninstall --force (no explicit --scope) silently defaulted
scope to "project" -- the registry never tracks which scope a skill was
actually distributed to. Combined with running --force from a directory that
happens to have its own client signal (e.g. this repo's own .claude/), this
guessed wrong and could target unrelated real files at the wrong-but-real
project path instead of the actual (user-scope) distributed copies -- which
is exactly how .claude/skills/atmos-asciicast (the one entry there that isn't
a symlink like its siblings) got deleted. Uninstall now checks both project
and user scope whenever scope isn't explicit, so the real distributed copy
is never silently missed or mistargeted; removeClientCopies no-ops for
whichever scope nothing was actually distributed to.
Checking both scopes means legitimately encountering symlinks unrelated to
the current session (e.g. this repo's own bundled-skill symlinks). Rather
than warning once per skill -- alarming and confusing when most of them are
for a scope the user never touched -- removeClientCopies now returns a count
and the caller reports one aggregate "N client-distributed copies left in
place (symbolic link)" line.
Batch install (InstallAllBundled/installMultiSkillPackage) now wraps the
per-skill loop in a progress spinner ("Distributing to: ...") that resolves
to the final tally, instead of printing two separate static lines. The
completion message always names the actual client + directory (e.g. "→
claude-code (~/.claude/skills)") rather than leaving that to the spinner's
progress text, since non-TTY output (CI, piped, tests) only ever shows the
completion message.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The macos-intel Build job failed on a transient DNS resolution failure reaching github.com while go mod download fetched a few indirect dependencies not indexed on the Go module proxy (gopsutil, wazero, etc.). The existing retry (3 attempts, constant 10s delay) only covered ~20-30s of runway, which wasn't enough to ride out the outage -- all 3 attempts failed within the same DNS blip, failing the whole build leg and cascading into skipped downstream jobs (e.g. k3s-required's matrix gate). Switch to exponential backoff (5 attempts, 5s initial delay, 60s cap), matching this repo's own documented convention for network-flaky steps (website/docs/workflows/workflows/workflow/steps/retry.mdx). This gives a DNS blip up to ~2 minutes of growing backoff to clear instead of ~20s. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…nk-line fix errors/formatter.go's leading blank line (ad4d1e4) needed 58 golden snapshots regenerated, but only ~20 were updated at the time -- the rest were skipped because this sandbox's terminal-width detection differs from the CI reference environment, and blindly regenerating them mixed in unrelated word-wrap changes. These 28 were the ones left behind, and they're exactly what failed CI (Acceptance Tests, linux + macos): every added line here is a blank line before an error/workflow-error header, hand-verified against the original wrapping to exclude the sandbox's width-detection noise. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…y leading newline Two independent regressions from the leading-blank-line fix in errors.Format(), both only reproducible with color enabled (this sandbox runs plain by default, so they passed locally but failed in CI): - cmd/ai/skill install/uninstall tests asserted substrings like "~/.atmos/skills" or "0 skills installed" directly against captured UI output. Under color, styled runs are wrapped in separate ANSI spans (e.g. "~" and "/.atmos/skills" end up in adjacent-but-distinct spans), splitting the literal substring and failing a naive Contains check. Strip ANSI codes before asserting, matching the atmosansi.Strip convention already used elsewhere in these test files. Reproduced and verified fixed locally with CLICOLOR_FORCE=1. - pkg/ai/analyze.Context.RunAnalysis appends errUtils.Format()'s raw result (now leading with a blank line, meant to visually separate the error from prior terminal output) into the captured Stderr sent to the AI provider. That leading blank line is a display-only concern and doesn't belong in the AI-facing capture, so trim it there specifically -- the actual terminal write on the line above is untouched and still gets the blank line. Also drops resolveSkillClients' now-constant title parameter (unparam): uninstall moved to its own resolveUninstallClients in an earlier commit, so install.go is its only remaining caller and the title never varies. Two other CI failures in this run (a `registry.opentofu.org` module lookup timeout in a real `tofu init` during an acceptance test, Linux-only) are unrelated transient network flakiness with no existing retry mechanism to tune -- left alone rather than guessing at a fix I can't verify. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…ic MCP client paths in tests
redactHomePath built the redacted display path via raw string prefix+trim
("~" + strings.TrimPrefix(path, homeDir)), which loses the separator
between "~" and the remainder whenever homeDir itself carries a trailing
separator (observed on Windows CI: homedir.Dir() returned one, producing
"~.atmos\skills" instead of "~\.atmos\skills" and failing two cmd/ai/skill
install tests' Contains checks). Rebuilt it on filepath.Rel, which handles
trailing-separator and normalization differences correctly regardless of
platform, and always joins with exactly one separator.
Also fixes two unrelated, pre-existing Windows-only failures in
pkg/mcp/install, exposed by this same CI run:
- TestResolveTarget_NewProjectAndUserScopeClients asserted a single
cross-platform suffix for Zed's and Goose's user-scope config path, but
zedUserPath/gooseUserPath correctly use each tool's real, documented
Windows locations (%APPDATA%\Zed\settings.json and
%APPDATA%\Block\goose\config\config.yaml) which differ in casing and
directory structure from the Unix path. Added Windows-specific expected
suffixes rather than changing the (correct) implementation.
- TestInstallJSONTarget_ClaudeDesktopGlobalScope: claudeDesktopUserPath
intentionally prefers the real %APPDATA% env var over the injected
homeDir on Windows, to match where Claude Desktop itself actually reads
config from -- but the test never overrode %APPDATA%, so on a real
Windows runner it was reading/writing the CI runner's actual user
profile instead of the test's isolated temp dir, and cross-test
pollution there was masking new file creation. Set %APPDATA% to a
subdirectory of the test's temp home (a no-op on non-Windows).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…'t eat Windows separators redactHomePath already builds a correct ~-prefixed path (fixed in 8c187c7), but ui.Successf/ui.Infof render their whole message through the project's CommonMark-based Markdown formatter, which treats a backslash before punctuation (e.g. the "\." before a dot-directory like ".atmos") as an escape and silently drops the backslash -- while a backslash before a letter (e.g. "\s") is left alone. On Windows this turned "~\.atmos\skills" into "~.atmos\skills", still failing TestInstallCmd_RunE_NoArgsInstallsEveryBundledSkill and TestInstallCmd_RunE_DistributingToShowsRealClientDirectory on the same commit that was supposed to fix them. CommonMark code spans are exempt from backslash-escape processing, so wrapping the redacted path in backticks (matching the existing convention for client names) fixes the display without touching redactHomePath itself, whose raw return value two existing unit tests assert on directly. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
💥 This pull request now has conflicts. Could you fix it Erik Osterman (Cloud Posse) (@osterman)? 🙏 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
what
atmos mcp installcommand (client-config installer for AI coding assistants) from the stalledosterman/pro-install-cmdbranch, leaving all Atmos Pro-specific wiring behind.atmos mcp add,atmos mcp remove, andatmos mcp uninstall, rounding out the existinglist/install/export/status/testset.add/removeonly editmcp.serversinatmos.yaml(the source of truth);install/uninstallpush/pull that declared config to/from AI client config files (Claude Code, Cursor, VS Code, Codex, Gemini).selfandatmos-propresets soatmos mcp add(no arguments) works out of the box, resolving to Atmos's own MCP server or the Atmos Pro MCP server respectively.selfpreset interactively offers to flipmcp.enabled: trueinatmos.yamlif it isn't already, since that preset requires it to function.atmos mcp list/statusnow nudge users to runatmos mcp add atmos-prowhen Atmos Pro is configured but its MCP server hasn't been added yet.atmos mcp installwith nothing configured now interactively offers to add+install theselfpreset instead of just printing a static hint./mcpdocs landing page (website/docs/mcp/mcp.mdx) covering both directions — Atmos as an MCP client and as an MCP server — cross-linked from/aivia tabs, matching the existing/proand/ailanding pages.atmos mcp install/uninstall's supported-client list from 5 to 15: adds Claude Desktop, Windsurf, Cline, Cline CLI, Zed, OpenCode, Goose, GitHub Copilot CLI, Antigravity, and MCPorter, alongside the existing Claude Code, Cursor, VS Code, Codex, and Gemini. Includes a new YAML config writer (for Goose) and bespoke per-client entry renderers where a client's schema differs from the commonmcpServersshape (Zed'scontext_servers, OpenCode'smcp).## Troubleshootingsection to theatmos mcp installdocs covering post-install behavior that varies by client (some need a restart, others need a manual settings toggle, some auto-reload).atmos ai skill installpreviously always wrote to~/.atmos/skills/, so VS Code / GitHub Copilot's.github/skills/couldn't discover installed skills.--path/ATMOS_AI_SKILL_PATHto override the install directory.atmos ai skill install <name>now auto-detects which AI clients (Claude Code, VS Code/Copilot, Gemini) are in use in the current project and copies the skill into each one's well-known skill directory, mirroringatmos mcp install's detect/--client/--all-clients/interactive-picker UX.atmos ai skill uninstallgained the same--client/--all-clientssymmetry to clean up those copies.atmos ai skill install/uninstallwith no<source>/<name>now acts on every skill (every bundled skill for install, every installed skill for uninstall) instead of requiring one invocation per skill — mirrorsatmos mcp install/uninstall's existing "no server names given = act on everything configured" convention rather than adding a separate--allflag.atmos config set/delete/formatUX polish:bool,int,float,string,yaml) from the Atmos config schema via reflection onyaml/mapstructurestruct tags, soatmos config set mcp.enabled truewrites a real boolean without needing--type=bool.Argsvalidator rejects the wrong number of arguments (e.g.atmos config unset mcp.enabled trueagainstExactArgs(1)) previously reported the misleadingUnknown command mcp.enabled— implying an unrecognized subcommand — when the real problem was too many arguments. Now surfaces Cobra's own accurate validator message.atmos toolchain installcan carry thecom.apple.quarantine/com.apple.provenancexattrs, causing Gatekeeper to block execution even though the binary was already checksum/signature-verified. Strips the xattrs and ad-hoc re-signs the binary on darwin, with averifier_trust: disabledopt-out.docs/prd/atmos-mcp-integrations.md(proposal only, not implemented in this PR) for natively signing requests to AWS-hosted MCP servers using Atmos Auth identities, without a third-partymcp-proxy-for-awsbridge.why
atmos mcp install's client-installer code was implemented on a branch entangled with unrelated, stalled Atmos Pro onboarding work; splitting it out lets the MCP work ship independently.atmos.yamland running client-specific commands manually —atmos mcp add self/atmos mcp add atmos-procollapse that to one command.atmos config set mcp.enabled truesilently writing the string"true"instead of a boolean is a footgun users would hit immediately while following the new MCP add/enable flow; schema-based type inference removes the need to know or pass--type=bool.atmos config unset mcp.enabled true), and the root cause turned out to be a small, generally-applicable flaw in how Cobra Args-validator failures were reported for leaf commands.--client/--all-clients/picker UX for skills keeps the two subsystems consistent for users learning either one.references
osterman/pro-install-cmdbranch (Atmos Pro-specific--mcpinstall flag intentionally left behind there).