You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Move from terminal-only interaction to a shared browser workflow where the agent can propose, the user can inspect and edit, and only then the action continues.
That is too narrow for high-risk or high-context work.
32
+
33
+
AgentClick extends the interaction into a browser UI so the agent can hand off a structured review surface for things like:
34
+
35
+
- emails and inbox triage
36
+
- shell commands and risky actions
37
+
- plans and trajectories
38
+
- forms and selections
39
+
- memory review
40
+
41
+
The goal is simple: keep the speed of terminal agents, but add a real review layer before the agent commits to irreversible work.
10
42
11
43
---
12
44
13
-
## Why AgentClick
45
+
## Why It Helps
14
46
15
-
-**Not just approve/deny** -- edit the email subject, change the command, modify the payload before it sends.
16
-
-**Preference learning** -- delete a paragraph and tell AgentClick why. It writes the rule to disk so your agent never makes the same mistake again.
17
-
-**Framework-agnostic** -- works with OpenAI, Anthropic, LangChain, or any HTTP-capable agent. Just POST and long-poll.
47
+
-**Edit before execution**: the user can change the draft, command, or payload instead of only approve/reject.
48
+
-**Shared visual context**: the agent and user move from raw terminal text to a purpose-built UI.
49
+
-**Preference learning**: feedback from review can be persisted so the agent improves over time.
50
+
-**Framework-agnostic**: anything that can `POST` JSON and poll an HTTP endpoint can use it.
18
51
19
52
---
20
53
21
54
## Quick Start
22
55
23
56
```bash
24
57
npm install -g @harvenstar/agentclick
25
-
agentclick # local only
26
-
agentclick --remote # + public URL for phone / remote access
58
+
agentclick
27
59
```
28
60
29
-
`--remote` automatically downloads and starts a [Cloudflare Quick Tunnel](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/do-more-with-tunnels/trycloudflare/) — no account needed. A public HTTPS URL is printed in the terminal; open it on any device.
30
-
31
-
Then test it with a mock session:
61
+
For remote access from another device:
32
62
33
63
```bash
34
-
curl -X POST http://localhost:3001/api/review \
35
-
-H "Content-Type: application/json" \
36
-
-d '{"type":"code_review","sessionKey":"test","payload":{"command":"rm -rf /tmp/old-cache","cwd":"/home/user","explanation":"Clean up stale cache directory","risk":"medium"}}'
64
+
agentclick --remote
37
65
```
38
66
39
-
A browser tab opens automatically. Review, approve or reject, and close the tab.
67
+
`--remote` automatically downloads and starts a [Cloudflare Quick Tunnel](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/do-more-with-tunnels/trycloudflare/). It prints a public HTTPS URL you can open on your phone or another machine.
40
68
41
69
---
42
70
43
-
## How It Works
71
+
## Use It With An Agent
44
72
45
-
1.**Agent POSTs structured data** to `http://localhost:3001/api/review` with a session key.
46
-
2.**User reviews, edits, and approves** in the browser -- paragraph-level delete/rewrite for emails, approve/reject for commands and actions.
47
-
3.**Agent receives the result** via long-poll (`GET /api/sessions/:id/wait`) and continues execution.
73
+
If your agent can work with local repos and local skills, keep the instructions short.
48
74
49
-
No WebSockets, no framework plugins. One HTTP endpoint in, one HTTP endpoint out.
75
+
Example:
50
76
51
-
---
77
+
1. Ask the agent to download or clone AgentClick.
78
+
2. Ask the agent to load the skill from this repo.
79
+
3. Ask the agent to use AgentClick for review-heavy tasks.
| Framework-agnostic | Yes | Yes | LangGraph only | Vercel only |
61
-
| Self-hosted | Yes | Yes | Yes | Cloud |
62
-
63
-
---
64
-
65
-
## Session Types
83
+
```text
84
+
Download agentlayer-io/AgentClick, load its SKILL.md, start it locally, and use it whenever you need a browser review UI instead of only terminal output.
85
+
```
66
86
67
-
-**email_review** -- two-column inbox and draft editor. Users can delete paragraphs with reasons, request rewrites, toggle intent suggestions, and confirm or regenerate.
68
-
-**code_review** -- displays the shell command, working directory, affected files as a collapsible tree, and risk level. Approve or reject with an optional note.
69
-
-**action_approval** -- generic high-risk action gate. Shows action description, detail, and risk badge. Approve or reject with an optional note.
70
-
-**form_review** -- editable key-value form. Each field can be free-text or a dropdown. Agent receives the edited values on completion.
71
-
-**selection_review** -- single or multi-select option picker. Agent receives the chosen option IDs.
72
-
-**trajectory_review** -- DAG visualization of a multi-step agent execution. Users can mark incorrect steps, provide per-step guidance, set a resume point, and request a retry. Guidance with "Remember this" checked is persisted to `MEMORY.md` for future runs.
73
-
-**plan_review** -- step tree with DAG layout for reviewing an agent's proposed plan before execution. Users can edit step labels, insert or remove steps, add constraints, choose an alternative plan, or request a regeneration.
74
-
-**memory_review** -- grouped file browser for reviewing agent memory modifications. Users decide which files to include or disregard, accept or reject proposed edits, and leave a global note for the agent.
87
+
If the agent supports sub-skills, the root [`SKILL.md`](./SKILL.md) routes to the right one automatically.
75
88
76
89
---
77
90
78
91
## Skill Layout
79
92
80
-
The root `SKILL.md` is a router only. Detailed review workflows live in:
81
-
82
-
-`skills/clickui-approve/`
83
-
-`skills/clickui-code/`
84
-
-`skills/clickui-email/`
85
-
-`skills/clickui-plan/`
86
-
-`skills/clickui-trajectory/`
87
-
-`skills/clickui-memory/`
88
-
89
-
Use the sub-skill that matches the review type instead of expanding the root skill file.
93
+
The root skill is [`SKILL.md`](./SKILL.md). It routes to the right sub-skill.
90
94
91
-
---
92
-
93
-
## API
94
-
95
-
| Method | Endpoint | Description |
95
+
| Skill | Path | Purpose |
96
96
|---|---|---|
97
-
| POST |`/api/review`| Agent creates a review session. Returns `{ sessionId, url }`. Browser opens automatically unless `noOpen: true` is passed. |
98
-
| POST |`/api/review/batch`| Create multiple sessions at once. Pass `{ sessions: [...] }`. Returns `{ sessions: [{ sessionId, url }] }`. All sessions are silent (no browser open). |
99
-
| GET |`/api/sessions/:id`| Fetch session data (payload, status, result). |
100
-
| GET |`/api/sessions/:id/wait`| Long-poll. Blocks up to 5 minutes until the user completes the review. |
101
-
| POST |`/api/sessions/:id/complete`| UI submits the user's decision. Triggers preference learning and agent callback. |
102
-
| PUT |`/api/sessions/:id/payload`| Agent updates payload after a rewrite cycle. Only valid when session status is `rewriting`. |
97
+
| Router |`SKILL.md`| Entry point that routes the agent to the right review workflow. |
98
+
| Action Approval |`skills/clickui-approve/`| Approve or reject risky actions before execution. |
99
+
| Code Review |`skills/clickui-code/`| Review shell commands, diffs, and code-related actions in UI. |
100
+
| Email Review |`skills/clickui-email/`| Review inbox items, drafts, replies, and live email sessions. |
101
+
| Plan Review |`skills/clickui-plan/`| Inspect and revise proposed plans before the agent runs them. |
0 commit comments