Documentation

Reference

Agents and capabilities

What each coding agent can do through Helmlane, how to sign it in, and how any ACP agent or SDK plugin joins — generated from the capability matrix the apps themselves use.

How to read this

Every agent keeps its own conversations and its own strengths, so the controls differ from one to the next. Helmlane shows a control only when that agent supports it on that computer — it never renders a button that would fail when you tap it.

Supported. Proven and on.
Partly supported. On, with the caveat noted.
Awaiting verification. Off until it is proven.
Not supported. The note says what to do instead.

Some things work everywhere because Helmlane provides them itself: Handoff to another agent, the hosted Fork of a chat (wherever Follow up works), the terminal, tasks, git and previews.

At a glance

Capabilities of each built-in agent
Capability	Cursor	Codex	Claude Code	Claude Desktop	OpenCode	Gemini CLI	Antigravity CLI	Antigravity IDE	ZCode	Muse CLI	Amp	Aider
Read history	Supported: IDE composers + transcripts; cursor-agent chats listed from its own store	Supported: shared ~/.codex store: app, CLI and IDE threads	Supported: ~/.claude/projects JSONL DAG (excludes Desktop-marked sessions)	Supported: Desktop session index + Desktop-marked transcripts; cloud chats not exposed	Supported: shared OpenCode history	Supported: ~/.gemini/tmp JSONL	Supported: ~/.gemini SQLite+JSONL	Supported: IDE SQLite+JSONL	Supported: SQLite index; app-server only for already-active sessions	Supported: MSP session/list + view/page	Partly supported: threads Helmlane ran here, from its own recording of the stream	Partly supported: chats Helmlane started, from their own Aider history files
Start chats	Supported: in the running Cursor app (helmlane-bridge); refused when no window has the project open	Supported: CLI app-server (same binary, config and account as the app)	Supported	Partly supported: drives the app via the Helmlane Desktop Driver (Accessibility); asks on the Mac while in use; folder must be a sidebar project	Supported: server + project handoff when Desktop is installed	Supported	Supported: agy -p	Partly supported: IDE's own agent via the Helmlane bridge (IDE 2.5+)	Supported: session/create	Supported: MSP session/start	Supported: amp -x --stream-json --stream-json-input	Supported: aider --message-file on a Helmlane-owned history
Follow up	Supported: IDE chats in the app; cursor-agent chats on the real CLI; legacy agent-* on @cursor/sdk	Supported: app threads inside the running app (IPC); others on the CLI app-server	Supported: claude --resume on a resident session	Partly supported: drives the app on a sidebar session (not archived, title unique in its project group); asks on the Mac while in use	Supported: server + project handoff when Desktop is installed	Supported: gemini --resume	Supported: agy --conversation	Partly supported: Helmlane bridge (IDE 2.5+)	Supported: session/send	Supported: MSP session/resume + turn/start	Supported: amp threads continue <id> -x	Supported: --restore-chat-history
Live output	Supported: IDE: persisted-bubble tail; CLI: stream-json	Supported: app-server notifications or the app owner stream	Supported: stream-json	Partly supported: app-written transcript, per content block	Supported: shared-server SSE	Supported: stream-json	Supported: stream-json	Partly supported: native transcript + engine run state	Supported: NDJSON stream	Supported: MSP view items (text, reasoning, tools, subagents)	Partly supported: message-granular stream-json; no token deltas	Partly supported: reply-granular from the history file
Stop	Supported: composer.cancelChat / SIGINT	Supported: turn/interrupt, or through the app owner	Supported: control-channel interrupt; session survives	Partly supported: presses the app's Stop control	Supported	Supported: process SIGINT	Supported: process SIGINT/TERM	Partly supported: engine cancel via the bridge	Supported: session/stop	Supported: MSP turn/interrupt	Supported: process SIGINT	Supported: SIGINT, then SIGTERM
Steer mid-run	Partly supported: legacy @cursor/sdk chats only; the app and cursor-agent -p take no mid-run input	Supported: turn/steer, or through the app owner	Partly supported: interrupt and resend; no mid-turn fold in headless mode (measured across tool boundaries)	Not supported: no Desktop mid-turn surface; the app queues follow-ups	Not supported: not in the server API	Not supported	Not supported	Not supported	Not supported: ZCode 0.16.5 rejects session/send while a turn is active	Supported: MSP turn/steer with an expectedTurnId guard	Awaiting verification: steer: true input documented, unproven live	Not supported: one message per process
Approvals and questions	Awaiting verification: verify: IDE approvals via Cursor's Accept/Reject commands; questions shown, answered in Cursor	Supported: app-server approvals, or answered through the app owner	Supported: can_use_tool over the control channel (needs --permission-prompt-tool stdio)	Partly supported: permission/question cards answered with the card's buttons (live prompt not yet exercised)	Supported: server permissions and questions (question.asked → POST /question/:id/reply; OpenCode 1.18.26)	Not supported: approval modes are configuration	Not supported: agy cannot surface approvals headless; permission modes choose the posture	Partly supported: command/file/permission approvals + questions; deploy targets stay in the IDE	Supported: PermissionRequested / approvals	Supported: MSP approval/request + userInput/request	Not supported: no approval surface in execute mode; denials reported	Not supported: --yes-always; shell commands always declined
Queue	Supported	Supported	Supported: app-level	Partly supported: Claude Desktop's own queue	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: native session_input + app-level	Supported: app-level	Supported: app-level	Supported: app-level
Choose a model	Supported: IDE picker catalog; per-chat model on create and follow-up	Supported: model/effort/speed/mode; app threads otherwise keep app settings	Supported: Sonnet/Opus/Fable/Haiku aliases + 3P	Partly supported: model/effort/permission mode via the app's pickers	Supported: configured providers only	Supported	Supported: agy models slugs	Partly supported: the IDE's own model list	Supported: GLM series + custom providers	Supported: local model catalog + MSP model/list	Not supported: Amp routes models by agent mode, chosen in Amp	Supported: Aider default + pulled Ollama models (ollama_chat/)
Usage and cost	Supported	Supported: thread/tokenUsage/updated; not streamed for app-run turns	Supported: result tokens/cost + real context window	Partly supported: app usage meter when the session is on screen	Supported: native message telemetry	Supported: headless statistics	Supported: stream-json tokens	Not supported	Supported: SQLite model_usage + usage/stats	Supported: MSP session/tokenUsage + real context window	Partly supported: per-response tokens attributed to the agent mode; no price	Supported: exact per-response tokens from the analytics log; Aider price when known
Rename	Partly supported: IDE chats via composer.updateTitle; others Helmlane-only	Supported	Supported: app-level + native title ingestion	Partly supported: app session header; a title already used in the project group is refused	Supported	Supported: app-level	Partly supported: native title read/app-level	Partly supported: title mirrored into the IDE when running	Supported: app-level	Supported: MSP session/rename + app-level title	Supported: app-level	Supported: app-level
Archive	Partly supported: desktop mirroring only	Supported	Supported: app-level	Partly supported: archives through the app; unarchive stays in the app	Supported: app-level (native delete is not archive)	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level
Pin	Partly supported: desktop mirroring only	Supported	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level	Supported: app-level
Native fork	Not supported: Handoff available	Supported	Not supported: native identity requires the first prompt; Handoff available	Not supported: Handoff available	Supported	Not supported	Not supported: Handoff available	Not supported	Supported: session/fork	Supported: MSP session/fork with durable provenance	Not supported: Handoff available	Not supported: Handoff available
Approve for this chat	Not supported: not proven: Cursor prompts carry no session scope	Supported: exact read-only commands (the allowlist below) in their folder, approved by Helmlane each time (a plain accept, never acceptForSession); file changes, typing into a running command and network prompts never offered	Supported: exact read-only Bash commands (the allowlist below), in-folder edits (when Claude offers accept-edits), one WebFetch domain, WebSearch, one MCP tool, approved by Helmlane each time (a plain allow, never a Claude rule or settings file)	Not supported: prompts are answered in the app window	Not supported: always grants the whole project, not one chat	Not supported: no interactions	Not supported: no session-scoped answer	Not supported: no session-scoped answer	Not supported: not proven	Not supported: host session choices exist; no recorded fixture proves their scope yet	Not supported	Not supported
Attach images	Supported	Supported	Supported	Not yet	Supported	Not yet	Not yet	Supported	Not yet	Supported	Supported	Not yet
Sign in from phone	Not supported	Supported	Supported	Not supported	Not supported	Not supported	Not supported	Not supported	Not supported	Supported	Not supported	Not supported

Generated from the capability matrix, the same table the apps render from.

Cursor

IDE and CLI agent · cursor

Cursor's editor composers through a bridge in the running Cursor app, plus `cursor-agent` for CLI chats.

Sign in

On the computer, as the user the Helmlane agent runs as:

Sign-in cannot be finished from a phone. sign in from the Cursor app (cursor-agent login signs in the CLI, not the IDE)

Install Cursor

How Helmlane knows

Signed-in check: the IDE's own sign-in in its state store (cursorAuth/* tokens; a refresh token means the IDE renews its own access)
Shown as unknown when: the state store cannot be read
Engine: running Cursor app via helmlane-bridge + storage reader; cursor-agent for CLI chats; @cursor/sdk for legacy agent-* chats
Runs as: GUI-owned app runs + supervised cursor-agent -p

Available (13)

Read historyIDE composers + transcripts; cursor-agent chats listed from its own store
Start chatsin the running Cursor app (helmlane-bridge); refused when no window has the project open
Follow upIDE chats in the app; cursor-agent chats on the real CLI; legacy agent-* on @cursor/sdk
Live outputIDE: persisted-bubble tail; CLI: stream-json
Stopcomposer.cancelChat / SIGINT
Steer mid-run · partlylegacy @cursor/sdk chats only; the app and cursor-agent -p take no mid-run input
Queue
Choose a modelIDE picker catalog; per-chat model on create and follow-up
Usage and cost
Rename · partlyIDE chats via composer.updateTitle; others Helmlane-only
Archive · partlydesktop mirroring only
Pin · partlydesktop mirroring only
Attach images

Not available (3)

Approvals and questions · awaiting verificationverify: IDE approvals via Cursor's Accept/Reject commands; questions shown, answered in Cursor
Native forkHandoff available
Approve for this chatnot proven: Cursor prompts carry no session scope

Codex

CLI agent · codex

OpenAI Codex through its app-server, with native approvals, steering and forks.

Sign in

On the computer, as the user the Helmlane agent runs as:

Run codex login in a terminal on the computer, or start a device-code sign-in from your phone.

Finish an expired sign-in from your phone. codex login --device-auth — device code

Install Codex

How Helmlane knows

Signed-in check: app-server account/read (requiresOpenaiAuth with no account = signed out)
Shown as unknown when: app-server not started
Engine: codex app-server (JSON-RPC, app-bundled CLI) + Codex app IPC router for threads the app owns
Runs as: supervised child process + follower of the running Codex app

Available (16)

Read historyshared ~/.codex store: app, CLI and IDE threads
Start chatsCLI app-server (same binary, config and account as the app)
Follow upapp threads inside the running app (IPC); others on the CLI app-server
Live outputapp-server notifications or the app owner stream
Stopturn/interrupt, or through the app owner
Steer mid-runturn/steer, or through the app owner
Approvals and questionsapp-server approvals, or answered through the app owner
Queue
Choose a modelmodel/effort/speed/mode; app threads otherwise keep app settings
Usage and costthread/tokenUsage/updated; not streamed for app-run turns
Rename
Archive
Pin
Native fork
Approve for this chatexact read-only commands (the allowlist below) in their folder, approved by Helmlane each time (a plain accept, never acceptForSession); file changes, typing into a running command and network prompts never offered
Attach images

Not available (0)

Claude Code

CLI agent · claude

Anthropic's Claude Code, one resident process per conversation with a two-way control channel for approvals.

Sign in

On the computer, as the user the Helmlane agent runs as:

Run claude auth login in a terminal on the computer, or start the sign-in from your phone: open the link it shows, then enter the code the page gives you back in Helmlane.

Finish an expired sign-in from your phone. claude auth login — login URL; the page's code is relayed back from the phone (codeEntry)

Install Claude Code

How Helmlane knows

Signed-in check: claude auth status --json
Shown as unknown when: output unparseable
Engine: Claude CLI (claude -p stream-json + bidirectional control channel, CLAUDE_CODE_ENTRYPOINT=helmlane) + on-disk JSONL DAG reader
Runs as: one resident supervised process per conversation

Available (15)

Read history~/.claude/projects JSONL DAG (excludes Desktop-marked sessions)
Start chats
Follow upclaude --resume on a resident session
Live outputstream-json
Stopcontrol-channel interrupt; session survives
Steer mid-run · partlyinterrupt and resend; no mid-turn fold in headless mode (measured across tool boundaries)
Approvals and questionscan_use_tool over the control channel (needs --permission-prompt-tool stdio)
Queueapp-level
Choose a modelSonnet/Opus/Fable/Haiku aliases + 3P
Usage and costresult tokens/cost + real context window
Renameapp-level + native title ingestion
Archiveapp-level
Pinapp-level
Approve for this chatexact read-only Bash commands (the allowlist below), in-folder edits (when Claude offers accept-edits), one WebFetch domain, WebSearch, one MCP tool, approved by Helmlane each time (a plain allow, never a Claude rule or settings file)
Attach images

Not available (1)

Native forknative identity requires the first prompt; Handoff available

Claude Desktop

Desktop app · claude-desktop

The installed Claude Desktop app, driven through macOS Accessibility by the Helmlane Desktop Driver, with its own engine, account and settings. Cloud-only Claude chats are not visible.

Sign in

On the computer, as the user the Helmlane agent runs as:

Allow “Helmlane Desktop Driver” once in System Settings → Privacy & Security → Accessibility.

Sign-in cannot be finished from a phone. sign in from Claude Desktop (claude auth login signs in the CLI, not the app)

Install Claude Desktop

How Helmlane knows

Signed-in check: None available
Shown as unknown when: always: the app keeps its sign-in and the Desktop Driver exposes no query
Engine: the installed app (com.anthropic.claudefordesktop) driven through macOS Accessibility by the signed Helmlane Desktop Driver + Desktop's session index + Desktop-marked JSONL transcripts
Runs as: GUI-owned; driver app + transcript watcher

Available (12)

Read historyDesktop session index + Desktop-marked transcripts; cloud chats not exposed
Start chats · partlydrives the app via the Helmlane Desktop Driver (Accessibility); asks on the Mac while in use; folder must be a sidebar project
Follow up · partlydrives the app on a sidebar session (not archived, title unique in its project group); asks on the Mac while in use
Live output · partlyapp-written transcript, per content block
Stop · partlypresses the app's Stop control
Approvals and questions · partlypermission/question cards answered with the card's buttons (live prompt not yet exercised)
Queue · partlyClaude Desktop's own queue
Choose a model · partlymodel/effort/permission mode via the app's pickers
Usage and cost · partlyapp usage meter when the session is on screen
Rename · partlyapp session header; a title already used in the project group is refused
Archive · partlyarchives through the app; unarchive stays in the app
Pinapp-level

Not available (4)

Steer mid-runno Desktop mid-turn surface; the app queues follow-ups
Native forkHandoff available
Approve for this chatprompts are answered in the app window
Attach imagesNot delivered to this agent yet; the composer says so instead of dropping the image.

OpenCode

CLI and desktop app · opencode

OpenCode's own server API, with the model providers you configured in OpenCode.

Sign in

On the computer, as the user the Helmlane agent runs as:

Run opencode auth login on the computer and pick a model provider.

Sign-in cannot be finished from a phone. opencode auth login needs an interactive provider pick

Install OpenCode

How Helmlane knows

Signed-in check: GET /provider on its own server: the connected list (0 = nothing signed in, so expired; only Helmlane's injected Ollama = unknown)
Shown as unknown when: server unreachable (no state reported)
Engine: opencode serve HTTP API (published OpenAPI) + official JS/TS SDK, plus optional signed app discovery + opencode://open-project handoff
Runs as: supervised opencode serve process (+ GUI handoff when the Desktop app is installed)

Available (14)

Read historyshared OpenCode history
Start chatsserver + project handoff when Desktop is installed
Follow upserver + project handoff when Desktop is installed
Live outputshared-server SSE
Stop
Approvals and questionsserver permissions and questions (question.asked → POST /question/:id/reply; OpenCode 1.18.26)
Queueapp-level
Choose a modelconfigured providers only
Usage and costnative message telemetry
Rename
Archiveapp-level (native delete is not archive)
Pinapp-level
Native fork
Attach images

Not available (2)

Steer mid-runnot in the server API
Approve for this chatalways grants the whole project, not one chat

Gemini CLI

CLI agent · gemini

Google's Gemini CLI in non-interactive mode, resuming its own saved chats.

Sign in

On the computer, as the user the Helmlane agent runs as:

Run gemini in a terminal on the computer and finish the sign-in it offers, or provide an API key in the environment or ~/.gemini/.env.

Sign-in cannot be finished from a phone. no login subcommand: the paste-back flow (NO_BROWSER) only runs by starting a session, which would spend a request and write history

Install Gemini CLI

How Helmlane knows

Signed-in check: API key (env, ~/.gemini/.env), oauth_creds.json refresh token, gcloud ADC, settings selection (an API-key selection with no key, or no selection and no credential, = expired)
Shown as unknown when: personal-OAuth selection without the file (keychain), Vertex/Cloud Shell
Engine: gemini -p non-interactive mode (--resume, --output-format stream-json) + disk reader
Runs as: supervised CLI process (stream-json stdout)

Available (11)

Read history~/.gemini/tmp JSONL
Start chats
Follow upgemini --resume
Live outputstream-json
Stopprocess SIGINT
Queueapp-level
Choose a model
Usage and costheadless statistics
Renameapp-level
Archiveapp-level
Pinapp-level

Not available (5)

Steer mid-run
Approvals and questionsapproval modes are configuration
Native fork
Approve for this chatno interactions
Attach imagesNot delivered to this agent yet; the composer says so instead of dropping the image.

Antigravity CLI

CLI agent · antigravity-cli

The Antigravity CLI (`agy`) in headless mode, reading its own transcripts.

Sign in

On the computer, as the user the Helmlane agent runs as:

Sign in from Antigravity on the computer. The CLI has no status command, so Helmlane shows its sign-in state as unknown unless a keyless credential is in the environment.

Sign-in cannot be finished from a phone. sign in from Antigravity

How Helmlane knows

Signed-in check: None available
Shown as unknown when: always: agy has no status command, its token store is undocumented (possibly keychain), and GOOGLE_API_KEY/GOOGLE_APPLICATION_CREDENTIALS are not documented as honoured by agy 1.x (SPIKE.md §4 predates §7, which was never run live), so they are not counted
Engine: Antigravity CLI (agy) headless / Language Server + SQLite/JSONL transcripts; MCP for tools
Runs as: supervised CLI process (stream-json stdout)

Available (11)

Read history~/.gemini SQLite+JSONL
Start chatsagy -p
Follow upagy --conversation
Live outputstream-json
Stopprocess SIGINT/TERM
Queueapp-level
Choose a modelagy models slugs
Usage and coststream-json tokens
Rename · partlynative title read/app-level
Archiveapp-level
Pinapp-level

Not available (5)

Steer mid-run
Approvals and questionsagy cannot surface approvals headless; permission modes choose the posture
Native forkHandoff available
Approve for this chatno session-scoped answer
Attach imagesNot delivered to this agent yet; the composer says so instead of dropping the image.

Antigravity IDE

IDE · antigravity-ide

Conversations in the Antigravity IDE, observed from its transcripts. Stopping, approvals and model choice stay in the IDE.

Sign in

On the computer, as the user the Helmlane agent runs as:

Sign-in cannot be finished from a phone. sign in from the IDE

How Helmlane knows

Signed-in check: None available
Shown as unknown when: always: sign-in lives inside the IDE; neither the launcher nor the agent bridge exposes an auth query
Engine: helmlane-antigravity-bridge extension → the IDE's own language server (Connect RPC) + SQLite/JSONL transcripts
Runs as: GUI-owned process; bridge over an owner-only Unix socket

Available (12)

Read historyIDE SQLite+JSONL
Start chats · partlyIDE's own agent via the Helmlane bridge (IDE 2.5+)
Follow up · partlyHelmlane bridge (IDE 2.5+)
Live output · partlynative transcript + engine run state
Stop · partlyengine cancel via the bridge
Approvals and questions · partlycommand/file/permission approvals + questions; deploy targets stay in the IDE
Queueapp-level
Choose a model · partlythe IDE's own model list
Rename · partlytitle mirrored into the IDE when running
Archiveapp-level
Pinapp-level
Attach images

Not available (4)

Steer mid-run
Usage and cost
Native fork
Approve for this chatno session-scoped answer

ZCode

Coding agent · zcode

ZCode through its app-server protocol, with native approvals, queueing and forks.

Sign in

On the computer, as the user the Helmlane agent runs as:

Configure a model provider in the ZCode app on the computer.

Sign-in cannot be finished from a phone. configure a provider in the app

How Helmlane knows

Signed-in check: app-server listModels in a known workspace (models = authenticated, empty = no model provider configured = expired)
Shown as unknown when: no refresh requested, no workspace to ask in, app-server not ready
Engine: zcode app-server (JSON-RPC stdio protocol server) + SQLite reader (~/.zcode/cli/db/db.sqlite)
Runs as: supervised child process

Available (13)

Read historySQLite index; app-server only for already-active sessions
Start chatssession/create
Follow upsession/send
Live outputNDJSON stream
Stopsession/stop
Approvals and questionsPermissionRequested / approvals
Queuenative session_input + app-level
Choose a modelGLM series + custom providers
Usage and costSQLite model_usage + usage/stats
Renameapp-level
Archiveapp-level
Pinapp-level
Native forksession/fork

Not available (3)

Steer mid-runZCode 0.16.5 rejects session/send while a turn is active
Approve for this chatnot proven
Attach imagesNot delivered to this agent yet; the composer says so instead of dropping the image.

Muse CLI

CLI agent · muse-cli

Muse CLI through its session host protocol, with steering, approvals and forks.

Sign in

On the computer, as the user the Helmlane agent runs as:

Run muse login on the computer, or set META_API_KEY.

Finish an expired sign-in from your phone. muse login — device code

How Helmlane knows

Signed-in check: META_API_KEY, or the login record in muse/auth.json (token itself in the keychain)
Shown as unknown when: no store (older CLI)
Engine: muse serve (MSP JSON-RPC session host over stdio) + on-disk model catalog; session.jsonl reader kept for read-only degradation and terminal liveness
Runs as: supervised child process, one per sandbox posture

Available (15)

Read historyMSP session/list + view/page
Start chatsMSP session/start
Follow upMSP session/resume + turn/start
Live outputMSP view items (text, reasoning, tools, subagents)
StopMSP turn/interrupt
Steer mid-runMSP turn/steer with an expectedTurnId guard
Approvals and questionsMSP approval/request + userInput/request
Queueapp-level
Choose a modellocal model catalog + MSP model/list
Usage and costMSP session/tokenUsage + real context window
RenameMSP session/rename + app-level title
Archiveapp-level
Pinapp-level
Native forkMSP session/fork with durable provenance
Attach images

Not available (1)

Approve for this chathost session choices exist; no recorded fixture proves their scope yet

Amp

CLI agent · amp

Sourcegraph's Amp in execute mode, streaming JSON, with Helmlane keeping its own record of each run.

Sign in

On the computer, as the user the Helmlane agent runs as:

Create an access token in Amp's settings and set it as AMP_API_KEY for Helmlane on the computer, or run amp login there.

Sign-in cannot be finished from a phone. amp login is not documented as finishable elsewhere; sign in on the machine

Install Amp

How Helmlane knows

Signed-in check: AMP_API_KEY holding a Settings access token (sgamp_…), the documented non-interactive credential
Shown as unknown when: always otherwise: a saved amp login account is in an unpublished store, and Amp has no documented status command
Engine: amp --execute --stream-json --stream-json-input (Claude-Code-compatible JSONL) + Helmlane's own recording of each run's stream
Runs as: supervised CLI process per run

Available (11)

Read history · partlythreads Helmlane ran here, from its own recording of the stream
Start chatsamp -x --stream-json --stream-json-input
Follow upamp threads continue <id> -x
Live output · partlymessage-granular stream-json; no token deltas
Stopprocess SIGINT
Queueapp-level
Usage and cost · partlyper-response tokens attributed to the agent mode; no price
Renameapp-level
Archiveapp-level
Pinapp-level
Attach images

Not available (5)

Steer mid-run · awaiting verificationsteer: true input documented, unproven live
Approvals and questionsno approval surface in execute mode; denials reported
Choose a modelAmp routes models by agent mode, chosen in Amp
Native forkHandoff available
Approve for this chat

Aider

CLI agent · aider

Aider driven one message at a time over a chat history Helmlane owns, with your configured models or pulled Ollama models.

Sign in

On the computer, as the user the Helmlane agent runs as:

Nothing to sign in to: put model API keys in Aider's configuration on the computer, or pull a model with Ollama.

Sign-in cannot be finished from a phone. nothing to sign in to

Install Aider

How Helmlane knows

Signed-in check: — (no account; model keys live in Aider's configuration, local models need none)
Shown as unknown when: always
Engine: aider --message-file over a Helmlane-owned --chat-history-file, observed through that file and --analytics-log; pulled Ollama models as ollama_chat/<name>
Runs as: supervised CLI process per run

Available (11)

Read history · partlychats Helmlane started, from their own Aider history files
Start chatsaider --message-file on a Helmlane-owned history
Follow up--restore-chat-history
Live output · partlyreply-granular from the history file
StopSIGINT, then SIGTERM
Queueapp-level
Choose a modelAider default + pulled Ollama models (ollama_chat/)
Usage and costexact per-response tokens from the analytics log; Aider price when known
Renameapp-level
Archiveapp-level
Pinapp-level

Not available (5)

Steer mid-runone message per process
Approvals and questions--yes-always; shell commands always declined
Native forkHandoff available
Approve for this chat
Attach imagesNot delivered to this agent yet; the composer says so instead of dropping the image.

Any ACP agent

One adapter covers every agent that speaks theAgent Client Protocol. It appears in Helmlane as acp:<name>. Helmlane registers an agent from a reviewed launcher table when its binary is installed and its own--help names the ACP entry point:

Agent	Command	Registered
`goose`	`goose acp`	Automatically
`kimi`	`kimi acp`	Automatically
`qwen`	`qwen --experimental-acp` (or `--acp`)	Automatically
`gemini`	`gemini --experimental-acp` (or `--acp`)	Opt-in — the built-in Gemini CLI covers it
`opencode`	`opencode acp`	Opt-in — the built-in OpenCode covers it
`claude-code`	`claude-code-acp`	Opt-in — the built-in Claude Code covers it
`codex`	`codex-acp`	Opt-in — the built-in Codex covers it

Add your own agents, opt in to a table entry, or turn one off by writing~/.helmlane/acp/agents.json on the computer. Helmlane reads it the next time it starts, keeps your choices, and deletes the file. Commands come only from that table and that file — never from anything an agent read or a webhook sent.

~/.helmlane/acp/agents.json

{
  "agents": [
    { "launcher": "gemini" },
    {
      "slug": "my-agent",
      "displayName": "My Agent",
      "command": "/opt/tools/my-agent",
      "args": ["--acp"]
    }
  ],
  "disabled": ["kimi"]
}

Signed-in check	`session/new` answering ACP `auth_required`
Shown as unknown	until a session has been attempted
Sign in from phone	Not supported: no reviewed login surface

Adapter SDK plugins

For agents that do not speak ACP, the adapter SDK (@helmlane/adapter-sdk) builds a plugin that appears asext:<publisher>/<name>. Install one by placing it in its own folder under ~/.helmlane/adapters/ with a manifest:

~/.helmlane/adapters/review-bot/helmlane-adapter.json

{
  "id": "ext:acme/review-bot",
  "entry": "dist/index.js"
}

A plugin runs inside the Helmlane agent with its privileges, so the loader holds it to the same standard ssh holds ~/.ssh:

the folder and entry file must belong to you and be writable by nobody else;
the entry must stay inside its own folder after following links;
the manifest and the adapter must agree on the id, and a plugin can never impersonate a built-in or an ACP agent;
a plugin built for a newer Helmlane agent is refused instead of half-running.

A plugin that fails any check is reported with the reason and not loaded. Like ACP agents, a loaded plugin starts with every capability off until a conformance run proves it.

Signed-in check	None: the SDK has no sign-in concept
Shown as unknown	always: the SDK has no login concept
Sign in from phone	Not supported

Agents and capabilities

How to read this#

At a glance#

Sign in

How Helmlane knows

Available (13)

Not available (3)

Sign in

How Helmlane knows

Available (16)

Not available (0)

Sign in

How Helmlane knows

Available (15)

Not available (1)

Sign in

How Helmlane knows

Available (12)

Not available (4)

Sign in

How Helmlane knows

Available (14)

Not available (2)

Sign in

How Helmlane knows

Available (11)

Not available (5)

Sign in

How Helmlane knows

Available (11)

Not available (5)

Sign in

How Helmlane knows

Available (12)

Not available (4)

Sign in

How Helmlane knows

Available (13)

Not available (3)

Sign in

How Helmlane knows

Available (15)

Not available (1)

Sign in

How Helmlane knows

Available (11)

Not available (5)

Sign in

How Helmlane knows

Available (11)

Not available (5)

Any ACP agent#

Adapter SDK plugins#

How to read this

At a glance

Any ACP agent

Adapter SDK plugins