Agent Intern

Run SEVERAL tasks IN PARALLEL across ALL backends in a single swarm. Each task is its own worker and names the backend to run on, so one swarm can mix Antigravity (Gemini), Codex, Copilot, Cursor, Grok, opencode and Muse workers — they run truly concurrently (capped at `max_concurrency`) and every answer comes back in one labelled block. A worker that fails is reported in place; the others still return. SECURITY: this launches N unsandboxed agents at once — N times the prompt-injection surface of a single call (see the module SECURITY note). Only use it with trusted prompts on trusted content.

Ask Antigravity (agy CLI, Gemini by default) a question in a NEW conversation. Uses your existing AI Pro authentication (silent-auth via Windows Credential Manager). Returns the model's final response as text. Good for fast tool-calling and short tasks; for heavier reasoning pick a bigger `model` or use the host model directly.

Continue the Antigravity conversation rooted at this workspace. Resumes the exact conversation id recorded for `workspace` (via agy's --conversation flag), not agy's global "most recent", so it stays correct even if agy was used elsewhere in between. On agy 1.1.8+ that id is the one agy itself reported for this bridge's last run in the workspace, so a follow-up resumes THIS thread even if you have since started a separate conversation in the same folder from Antigravity's own interface.

Generate an image with Antigravity (Gemini image model via agy CLI). Drives agy to produce a raster image on your existing AI Pro quota, saves it, and returns the absolute file path plus its real format and byte size. The host can then read the path to view the image. agy picks the image format itself (JPEG for photo-like images, PNG for flat graphics), so the returned path's extension is corrected to match the actual bytes (a requested out.png may come back as out.jpg). Runs a normal, unsandboxed agy session — same privileges/caveats as the other tools (see the module SECURITY note).

Generate several images IN PARALLEL with Antigravity (one worker per prompt). Like antigravity_image, but runs N image generations concurrently in isolated workers (capped at `max_concurrency`). Returns one block listing each image's final path/format/size (or its error). Extensions are corrected to the real bytes, exactly like antigravity_image. Same unsandboxed privileges/caveats as antigravity_swarm.

Report diagnostics for the agy bridge setup (spends no AI Pro quota). Reports the bridge's own version and whether a newer release is available (best-effort GitHub check; honors AGY_BRIDGE_NO_UPDATE_CHECK), then checks whether agy is on PATH (and its version/compat), how much AI Pro quota is left per model family (agy 1.1.11+ answers `/usage` in print mode for free — a family at 0% is reported as a problem, since every call against it will fail until its window resets), whether agy's state directories exist, whether the newest conversation transcript is readable, and whether the SQLite conversation store is present. Use this to debug empty or failed responses — or to see if the bridge itself is out of date, or if you are simply out of quota — before spending quota.

Ask OpenAI Codex (`codex exec`) a question or task in a NEW session. Uses your existing Codex login (ChatGPT or API key — see `codex login status`). Returns the agent's final message as text, read from codex's --output-last-message file (no stdout scraping). Codex is a capable coding agent, so this suits heavier reasoning and real code work, not just cheap tool-calling. Point `workspace` at a real project dir for context-aware answers.

Continue the Codex session rooted at this workspace (`codex exec resume`). Resumes the exact session id captured from the last codex_ask in this workspace, falling back to the newest on-disk session whose recorded cwd matches (so it still works after a server restart). The resumed session keeps its original sandbox and model — those are chosen when you start it with codex_ask.

Report diagnostics for the Codex bridge setup (spends no quota). Reports the bridge's own version and whether a newer release is available (best-effort GitHub check; honors AGY_BRIDGE_NO_UPDATE_CHECK) — the same update notice antigravity_status shows, so a Codex-only install still surfaces it — then checks whether codex is on PATH (and its version), whether you're logged in (`codex login status` — no model call, no quota), where codex stores its sessions, and how many workspace sessions are pinned this run. Use this to debug "codex not found" or auth errors before spending quota.

Ask the GitHub Copilot CLI (`copilot -p`) a question or task in a NEW session. Uses your existing Copilot login (OS credential store, or a COPILOT_GITHUB_TOKEN/GH_TOKEN/GITHUB_TOKEN env var — see `copilot_status`). Returns the agent's final message, read straight from stdout (the CLI's `-s` silent mode; no scraping). Copilot is a capable agentic coder — good for real code/repo work; point `workspace` at a project dir for context-aware answers.

Continue the Copilot session rooted at this workspace (resumes its `--session-id`). Resumes the exact session id the bridge set on the last copilot_ask in this workspace, falling back to the newest on-disk session whose recorded cwd matches (so it still works after a server restart). Unlike codex_continue, copilot re-applies permission flags on every call, so `sandbox` takes effect here too — e.g. analyze read-only with copilot_ask, then continue with "workspace-write" to apply the fix.

Report diagnostics for the Copilot bridge setup (spends no quota). Reports the bridge's own version and whether a newer release is available (best-effort GitHub check; honors AGY_BRIDGE_NO_UPDATE_CHECK) — the same update notice antigravity_status shows, so a Copilot-only install still surfaces it — then checks whether copilot is on PATH (and its version), an auth hint (copilot has no `login status` command, so this is best-effort — an env token is reported when set, otherwise login via the credential store is assumed and unverified), where copilot stores session state, and how many workspace sessions are pinned this run. Use this to debug "copilot not found" or auth errors before a call.

Ask the Cursor CLI (`cursor-agent -p`) a question or task in a NEW chat. Uses your existing Cursor login (OS credential store, or a CURSOR_API_KEY env var — see `cursor_status`). Returns the agent's final message, read straight from stdout (no scraping). Cursor is a capable agentic coder with a wide model menu (GPT / Claude / Grok / Composer); point `workspace` at a project dir for context-aware answers.

Continue the Cursor chat rooted at this workspace (resumes its chat id). Resumes the exact chat id the bridge minted on the last cursor_ask in this workspace, falling back to the newest on-disk chat whose recorded cwd matches (so it still works after a server restart). cursor applies permission flags per invocation, so `sandbox` takes effect here too — e.g. analyze read-only with cursor_ask, then continue with "workspace-write" to apply the fix.

Report diagnostics for the Cursor bridge setup (spends no quota). Reports the bridge's own version and whether a newer release is available (best-effort GitHub check; honors AGY_BRIDGE_NO_UPDATE_CHECK) — the same update notice antigravity_status shows, so a Cursor-only install still surfaces it — then checks whether cursor-agent is found (and its version), whether you're logged in (`cursor-agent status`), and where cursor stores its chats. Use this to debug "cursor not found" or auth errors before spending quota.

Ask Grok Build (`grok -p`) a question or task in a NEW session. EXPERIMENTAL. ⚠️ Community-verified only: this bridge has never completed an authenticated round-trip, because the author has no Grok subscription. Its flags are verified against grok 1.0.3, but the answer path is not. If it misbehaves, say so rather than working around it — and please report it. Needs a SuperGrok / X Premium+ login (`grok login`) or an XAI_API_KEY env var; run `grok_status` first to check. Returns the agent's final message, read from grok's `--output-format json` result. Point `workspace` at a project dir for context-aware answers.

Continue the Grok session rooted at this workspace. EXPERIMENTAL. Resumes the exact session id grok returned on the last grok_ask in this workspace (`-r <id>`), falling back to grok's own "most recent session for this cwd" (`-c`) when that in-memory pin is gone — so it still works after a server restart. grok applies permission flags per invocation, so `sandbox` takes effect here too: analyze read-only with grok_ask, then continue with "workspace-write" to apply the fix. Same experimental caveat and auth requirement as grok_ask.

Report diagnostics for the Grok Build bridge setup (spends no quota). Reports the bridge's own version and any newer release (the same update notice antigravity_status shows), then whether `grok` is found (and its version), whether you're authenticated, which models it offers, and where grok keeps its data. Auth and the model list both come from `grok models`, which answers even when logged out — so this is cheap and safe to call first. Use this to debug "grok not found" or auth errors before spending quota. This backend is EXPERIMENTAL and unverified end-to-end, so a green status here means the setup looks right, not that a live answer has ever been confirmed.

Ask Kimi Code (`kimi -p`) a question or task in a NEW session. EXPERIMENTAL. ⚠️ Community-verified only — built without a Kimi account, so no authenticated round-trip has ever run and the author cannot verify it. It won't answer until you authenticate: run `kimi login` (device-code) or put an API key in ~/.kimi-code/config.toml, then check `kimi_status`. Returns the agent's final message, read straight from stdout. Kimi Code is Moonshot's terminal coding agent (Kimi K2 family); point `workspace` at a project dir for context-aware answers. Kimi print mode has NO sandbox and auto-executes every tool call (like antigravity), so run it only with trusted prompts on trusted content.

Continue the Kimi session rooted at this workspace (`kimi -c`). EXPERIMENTAL. Resumes the previous Kimi session for this workspace via `-c/--continue` — Kimi scopes sessions per working directory, so there's no id to track. Errors if no prior kimi_ask ran in this workspace. Same experimental caveat and auth requirement as kimi_ask.

Report diagnostics for the Kimi bridge setup (spends no quota). EXPERIMENTAL. Reports the bridge's own version and any newer release (same update notice antigravity_status shows), then checks whether `kimi` is found (and its version), whether a provider is configured (`kimi provider list` — the auth proxy, since Kimi needs `kimi login` or an API key in config.toml), and where Kimi stores its data. This backend is unverified, so expect the auth row to say "no providers configured" until you log in.

Ask Meta's Muse Code (`muse exec`) a question or task in a NEW session. EXPERIMENTAL. ⚠️ The real model has never answered through this bridge — the author has no Muse plan. Muse's built-in offline `echo` provider verified everything else end to end (argv, event stream, answer, session resume), so a failure here is most likely auth or the model itself. If it misbehaves, say so plainly and please report it. Needs a Muse Code plan (`muse login`) or a META_API_KEY; run `muse_status` first. Returns the agent's final message (the `run_terminal` event of `muse exec --json`). Point `workspace` at a project dir for repo context.

Continue the Muse session rooted at this workspace. EXPERIMENTAL. Resumes the exact session the last muse_ask in this workspace created (the bridge names each session itself with `--session-id`). After a server restart it falls back to muse's own record of the workspace's most recent session (`muse export --last`); with no session there at all it errors rather than silently starting a fresh one. Muse applies safety flags per run, so `sandbox` takes effect here too. Same experimental caveat and auth needs as muse_ask.

Report diagnostics for the Muse Code bridge setup (spends no quota). Reports the bridge's own version and any newer release, then whether `muse` is found (and which binary the bridge runs), whether credentials are present (META_API_KEY or a `muse login`), any cached model catalog, the Windows OS sandbox state, and where muse keeps its data. Muse has no free auth probe, so a green auth row means credentials exist, not that they are still valid. This backend is EXPERIMENTAL: green here means the setup looks right, not that a real answer has ever been confirmed.

Ask opencode (`opencode run`) a question or task in a NEW session. The one backend here that needs NO subscription: opencode's own free hosted models (`opencode/*-free`, see `opencode models`) answer with zero credentials configured, so this works on a machine that has never logged in to anything. Add a key with `opencode auth login` for Claude/GPT-class models. Returns the agent's final message, reconstructed from opencode's `--format json` events. Point `workspace` at a project dir for context-aware answers. ⚠️ The free models are SLOW (queue-scheduled — a one-word answer has taken minutes), which is why timeout_s defaults to 300 here. Prefer a configured paid model for anything long, and don't mistake slowness for a hang.

Continue the opencode session rooted at this workspace. Resumes the exact session id opencode reported on the last opencode_ask in this workspace (`-s <id>`), falling back to opencode's own "most recent session for this directory" (`-c`) when that in-memory pin is gone — so it still works after a server restart. Session scoping is per directory (verified live: the same `-c` from a different directory starts a fresh session), so pass the same `workspace` you asked in. The permission policy applies per invocation, so `sandbox` takes effect here too: analyze read-only with opencode_ask, then continue with "workspace-write" to apply the fix.

Report diagnostics for the opencode bridge setup (spends no quota). Reports the bridge's own version and any newer release (same update notice antigravity_status shows), then checks whether `opencode` is found (and its version), how many provider credentials are configured, which model ids are available, and where opencode keeps its data. "0 credentials" is NOT a failure here — opencode's free `opencode/*` models still answer — so that row stays ok as long as models are listed.

Run a PREDEFINED swarm: a named panel of agents from different model families, each in its own role, all working on the same material in parallel. One call instead of building agent_swarm tasks by hand, and the same panel every time, so two runs (two applications, two drafts) can be compared.

List the predefined swarms preset_swarm can run, or show one in full. Lists each preset's kind, members (role and backend), rubric for a jury, and where it comes from. Built-ins can be replaced or extended with JSON files in ~/.agent-intern/swarms/ (every project) or <workspace>/.agent-intern/swarms/ (one project; its members run read-only and it cannot replace a built-in or user preset, because it arrives with whatever repo was cloned). Broken files are listed with the reason they were skipped.