Evidence
What the home page rests on.
Each claim the home page makes is listed here once, with its release status, where it lives in the code, what proves it and what it does not say. Where we haven’t recorded something yet, it says so.
- Available
- In the free download, for macOS and Windows, today.
- Pro preview
- In the AgentDock Pro developer preview: an unsigned build for Intel Macs.
- In development
- Working on a development build. Not in any download yet.
- Planned
- Not built yet.
How the home page footage was made
Recorded 27 October 2026. Recorded on a private X virtual display (Xvfb) on Linux, never on a visible screen: AgentDock Pro 0.1.6-dev.4, a development build that is not a download, in a 1440 × 900 window at 2× scale, with a throwaway profile and throwaway projects. The agents are the real Claude Code and Codex, signed in with the recording machine's own accounts. A script typed each request and pressed Allow, Approve and Start; everything the agents did and wrote is their own.
A montage of four separate recordings, one scene each, in this order: the Agent Browser, the agent's desktop, tools, AIgency. They were not one run, and no scene borrows from another recording. Inside a scene the cuts are hard and every shot keeps its own constant speed, from real time to 3×; a few end on a held frame. Between scenes the picture dips through black. Each shot is cropped to a rectangle of the window twice, with the same timing: a wide crop, and a closer one for phones. One card is too wide to read whole on a phone, so its phone crop pans across it to its buttons and holds there. Nothing is retouched, composited, reordered within a shot or generated. The script is docs/footage-2026-10-27/build.py.
The opening film · 27.8 s
| At | Scene | Shot | From the recording | Seconds used | Speed | Held at the end |
|---|---|---|---|---|---|---|
| 0.0 s | Browser | request | v-browser | 8.9–10.1 | 1.5× | — |
| 0.8 s | Browser | nodejs | v-browser | 10.1–13.6 | 2× | — |
| 2.6 s | Browser | github | v-browser | 20.8–22.8 | 1.5× | — |
| 3.9 s | Browser | approval | v-browser | 30–31.4 | real time | — |
| 5.3 s | Browser | answer | v-browser | 36.2–37.7 | 1.5× | — |
| 6.3 s | Browser | changes | v-browser | 39.4–40.8 | real time | 0.4 s |
| 8.1 s | Desktop | xcalc | v-desktop | 9.8–12.6 | 2× | — |
| 9.5 s | Desktop | clicks | v-desktop | 13–18.6 | 3× | — |
| 11.4 s | Desktop | result | v-desktop | 18.6–20.2 | real time | 0.6 s |
| 13.6 s | Tools | at-menu | v-tools | 7.25–9.1 | real time | — |
| 15.4 s | Tools | registry | v-tools | 21.6–23.6 | 1.5× | — |
| 16.8 s | Tools | install | v-tools | 24–27.2 | 2× | — |
| 18.4 s | Tools | installed | v-tools | 27.9–29.5 | real time | 0.4 s |
| 20.4 s | AIgency | card | rec-a2 | 32.35–32.95 | real time | 1.6 s |
| 22.6 s | AIgency | plan | rec-a2 | 86.4–88.4 | 1.25× | — |
| 24.2 s | AIgency | task-1 | rec-a2 | 95.5–97.75 | 2.5× | — |
| 25.1 s | AIgency | task-2 | rec-a2 | 121–123.25 | 2.5× | — |
| 26.0 s | AIgency | task-3 | rec-a2 | 158.5–162 | 2.5× | 0.4 s |
Loop: a browser · 3.9 s
| At | Scene | Shot | From the recording | Seconds used | Speed | Held at the end |
|---|---|---|---|---|---|---|
| 0.0 s | Browser | request | v-browser | 8.9–10.1 | 1.5× | — |
| 0.8 s | Browser | nodejs | v-browser | 10.1–13.6 | 2× | — |
| 2.6 s | Browser | github | v-browser | 20.8–22.8 | 1.5× | — |
Loop: you review · 4.2 s
| At | Scene | Shot | From the recording | Seconds used | Speed | Held at the end |
|---|---|---|---|---|---|---|
| 0.0 s | Browser | approval | v-browser | 30–31.4 | real time | — |
| 1.4 s | Browser | answer | v-browser | 36.2–37.7 | 1.5× | — |
| 2.4 s | Browser | changes | v-browser | 39.4–40.8 | real time | 0.4 s |
Loop: a desktop · 5.5 s
| At | Scene | Shot | From the recording | Seconds used | Speed | Held at the end |
|---|---|---|---|---|---|---|
| 0.0 s | Desktop | xcalc | v-desktop | 9.8–12.6 | 2× | — |
| 1.4 s | Desktop | clicks | v-desktop | 13–18.6 | 3× | — |
| 3.3 s | Desktop | result | v-desktop | 18.6–20.2 | real time | 0.6 s |
Loop: tools · 6.8 s
| At | Scene | Shot | From the recording | Seconds used | Speed | Held at the end |
|---|---|---|---|---|---|---|
| 0.0 s | Tools | at-menu | v-tools | 7.25–9.1 | real time | — |
| 1.9 s | Tools | registry | v-tools | 21.6–23.6 | 1.5× | — |
| 3.2 s | Tools | install | v-tools | 24–27.2 | 2× | — |
| 4.8 s | Tools | installed | v-tools | 27.9–29.5 | real time | 0.4 s |
Loop: a team · 7.4 s
| At | Scene | Shot | From the recording | Seconds used | Speed | Held at the end |
|---|---|---|---|---|---|---|
| 0.0 s | AIgency | card | rec-a2 | 32.35–32.95 | real time | 1.6 s |
| 2.2 s | AIgency | plan | rec-a2 | 86.4–88.4 | 1.25× | — |
| 3.8 s | AIgency | task-1 | rec-a2 | 95.5–97.75 | 2.5× | — |
| 4.7 s | AIgency | task-2 | rec-a2 | 121–123.25 | 2.5× | — |
| 5.6 s | AIgency | task-3 | rec-a2 | 158.5–162 | 2.5× | 0.4 s |
- v-browser
- Claude Code (Claude Sonnet 5.5) is asked to open nodejs.org in the Agent Browser, find the current LTS release and its end-of-life date, and add both to NOTES.md. It reads nodejs.org and the Node.js release schedule on GitHub, asks before running a command, and writes the file; the Changes panel then lists it.
- v-desktop
- Claude Code (Claude Sonnet 5.5) is asked to start its own AgentDock desktop, open xcalc and work out 12 × 12 by clicking the buttons. It starts the desktop, launches xcalc, clicks 1, 2, ×, 1, 2, = and reports 144.
- v-tools
- The model picker, the @ menu of skills, tools and agents, and Settings › Capabilities installing sarja/example-adcap 1.0.0 from the live Arcotheon registry, approved in its install dialog.
- rec-a2
- Claude Code (Claude Opus 5.5) asks to use AIgency; the request card is approved, AIgency drafts a three-task plan, Start is pressed and the tasks run (shown up to two of three). The run's end is not used: it reported success although its verification failed (AgentDock-Pro issue #89).
The full session recording
Claude Code adds an option to a small command-line tool, asking before each edit; AgentDock shows the diff; the same conversation is then handed to Codex, which reviews the change. The agents are the real Claude Code and Codex on the recording machine’s own accounts. A script typed the requests and pressed Allow; the waiting is cut out and nothing is sped up.
Available
AgentDock puts Claude Code, Codex and Antigravity in one desktop workspace, with your projects and conversations.
Checked 2026-10-24 · In v0.1.1 (macOS, Intel) and v0.1.2 (Windows)
Where it lives
- AgentDock · tags v0.1.1_macOS, v0.1.2 · src/main (agent adapters), src/renderer
What proves it
- Public GitHub releases of crossainthero-lab/AgentDock
- Agent adapters for all three present at the release tags
What it does not say
Says nothing about Linux, Apple-silicon-native builds or Orbit. macOS is the Intel build, which runs on Apple silicon through Rosetta 2.
AgentDock does not sell model access: your agents run on the accounts you already have with their providers.
Checked 2026-10-24 · In v0.1.1 (macOS, Intel) and v0.1.2 (Windows)
Where it lives
- AgentDock · tags v0.1.1_macOS, v0.1.2 · src/main (agent adapters launch the provider's own CLI/SDK with the user's sign-in)
- AgentDock Pro · private source repository
What proves it
- Download page and docs: 'Bring one agent … signed in with your own account'
- Pro's NVIDIA provider reports 'NVIDIA NIM API key is not configured' until the user supplies their own key; a search of the Pro source and docs found no hosted-inference or included-model code
What it does not say
Holds for the free app and the Pro preview as audited on 2026-10-24. It says nothing about third-party media providers, which also use the user's own accounts and credits, and Pro's NVIDIA delegation has not completed signed-in validation.
Agents ask before they edit files or run commands, and you can deny.
Checked 2026-10-24 · In v0.1.1 (macOS, Intel) and v0.1.2 (Windows)
Where it lives
- AgentDock · tags v0.1.1_macOS, v0.1.2 · src/renderer/components/ApprovalDialog.tsx
What proves it
- ApprovalDialog present at the release tags
- Permission modes are per-agent settings
What it does not say
Permission presets (Restricted / Standard / Expanded) are a Pro preview feature and are not claimed here. A run can also be set to bypass approvals, and says so on screen.
You can hand a conversation from one agent to another, and review and revert what changed.
Checked 2026-10-24 · In v0.1.1 (macOS, Intel) and v0.1.2 (Windows)
Where it lives
- AgentDock · tags v0.1.1_macOS, v0.1.2 · src/main/handoff-service.ts
- AgentDock · tags v0.1.1_macOS, v0.1.2 · src/renderer/components/ChangesDrawer.tsx
What proves it
- Handoff service and Changes drawer present at the release tags; the Changes drawer uses git, which must be installed
What it does not say
In v0.1.1 and v0.1.2 a handoff carries the conversation's context into a new conversation with the other agent; it is not switching agent in place. On the development build the Changes view, the terminal and handoffs are AgentDock Pro features (entitlements: advanced-projects, advanced-workspace); the free downloads have them.
Pro preview
On 10 September 2026, AIgency ran one written brief as a six-task plan across Claude Code and Codex. About ten minutes later there was a command-line todo app that builds, with 58 passing tests as reported by the checking agent.
Checked 2026-10-24
Where it lives
- AgentDock Pro · private source repository
- AgentDock Pro · private source repository
What proves it
- The run's own saved record: objective, plan, per-task timings, the checking agent's command log and written report (run-data)
- Screenshot of the window when it finished (run-screenshot)
- Codex's report states build exit 0 and '58 tests, 58 pass, 0 fail'; its command log shows the built CLI's real output for add, list, complete, delete and three error cases
- Recorded by a real run · src/content/evidence/todo-cli-run.json
Extracted read-only from the local AgentDock profile database by scripts/evidence/extract-aigency-run.py: run 73c5a723-953b-4be4-a189-39ce49b524c9 (objective, six tasks with agents and timings, the checking agent's command log and written report). Machine-specific paths shortened; nothing else changed. - Real screenshot · public/evidence/run-full-2400.webp
Screenshot of the AgentDock window taken on 10 Sep 2026 when the run finished, on a development build of AgentDock Pro (an earlier interface than today's). Derivatives (full, tasks crop, attention crop) built by scripts/evidence/build-evidence-images.py; a crop selects a rectangle and retouches nothing.
Original SHA-256 1b188354098da0c6d363871a15032272222d16d72f45e65662887a0a9beccd00
What it does not say
A small project, chosen so it can be checked. The run record does not show who wrote the six-task plan (every assignment is recorded with source 'manual' and the log has no planner step), so the page claims nothing about AIgency's planner or about it choosing agents from this run. 'Working' rests on the checking agent's report and logged commands; the run's own formal verification came back inconclusive on one criterion (see aigency-inconclusive). Run on a development build with approvals bypassed for the run; the current interface looks different. The project folder has since been edited by later runs, so it is not offered as a download. Until AgentDock-Pro issue #89 is fixed, the status AIgency gives a run is not itself evidence that it worked.
Not yet recorded
A screen recording of a fresh run on the current build (docs/capture-workflow.md, 'Run A'), with a genuine Pro sign-in.
AIgency started each task as soon as the tasks it depended on were done, and ran independent tasks at the same time, here for 2 min 39 s.
Checked 2026-10-24
Where it lives
- AgentDock Pro · private source repository
What proves it
- Task start and end times in the run record (run-data): the CLI entry point (Codex) and the tests (Claude Code) overlap
- Run concurrency was set to 2
- Recorded by a real run · src/content/evidence/todo-cli-run.json
Extracted read-only from the local AgentDock profile database by scripts/evidence/extract-aigency-run.py: run 73c5a723-953b-4be4-a189-39ce49b524c9 (objective, six tasks with agents and timings, the checking agent's command log and written report). Machine-specific paths shortened; nothing else changed. - Real screenshot · public/evidence/run-full-2400.webp
Screenshot of the AgentDock window taken on 10 Sep 2026 when the run finished, on a development build of AgentDock Pro (an earlier interface than today's). Derivatives (full, tasks crop, attention crop) built by scripts/evidence/build-evidence-images.py; a crop selects a rectangle and retouches nothing.
Original SHA-256 1b188354098da0c6d363871a15032272222d16d72f45e65662887a0a9beccd00
What it does not say
The overlap is two tasks, in one shared folder. Isolated git worktrees exist in AIgency but this run did not use them. Task assignment in this run is recorded with source 'manual'; do not describe it as automatic model routing.
Not yet recorded
A recording of a deliberately parallel job (three independent modules, then integration and verification) at concurrency 3.
When one required check came back without a result, AgentDock marked the run as needing attention instead of calling it a success, and asked a person to decide.
Checked 2026-10-24
Where it lives
- AgentDock Pro · private source repository
What proves it
- Criterion 'CLI supports all four commands' recorded as blocked: 'The verifying agent did not report a result for this criterion.' (run-data)
- The 'Needs your attention' panel in the screenshot (run-screenshot)
- Recorded by a real run · src/content/evidence/todo-cli-run.json
Extracted read-only from the local AgentDock profile database by scripts/evidence/extract-aigency-run.py: run 73c5a723-953b-4be4-a189-39ce49b524c9 (objective, six tasks with agents and timings, the checking agent's command log and written report). Machine-specific paths shortened; nothing else changed. - Real screenshot · public/evidence/run-full-2400.webp
Screenshot of the AgentDock window taken on 10 Sep 2026 when the run finished, on a development build of AgentDock Pro (an earlier interface than today's). Derivatives (full, tasks crop, attention crop) built by scripts/evidence/build-evidence-images.py; a crop selects a rectangle and retouches nothing.
Original SHA-256 1b188354098da0c6d363871a15032272222d16d72f45e65662887a0a9beccd00
What it does not say
Not a claim that verification is reliable in general: here it flagged a gap that the checking agent's notes suggest was a reporting miss, not a failed command. A later run did the opposite: its verification task reported a failure, yet the run was shown as completed (AgentDock-Pro issue #89, open). Until that is fixed, a run's status is not a verified result.
AgentDock reports a failed or inconclusive check and leaves the next step to you; it does not start repairs by itself.
Checked 2026-10-24
Where it lives
- AgentDock Pro · private source repository
What proves it
- Source comment and behaviour on Pro main; the run above stopped for a decision with a repair iteration offered, not started
- Real screenshot · public/evidence/run-full-2400.webp
Screenshot of the AgentDock window taken on 10 Sep 2026 when the run finished, on a development build of AgentDock Pro (an earlier interface than today's). Derivatives (full, tasks crop, attention crop) built by scripts/evidence/build-evidence-images.py; a crop selects a rectangle and retouches nothing.
Original SHA-256 1b188354098da0c6d363871a15032272222d16d72f45e65662887a0a9beccd00
What it does not say
Replaces the older marketing wording 'repair loops', which implied the opposite. In AgentDock-Pro issue #89 (open), a failed verification was reported in its task, but the run was still shown as completed, so a failure can be missed if you read only the run's status.
AIgency can give parallel tasks their own isolated git worktrees.
Checked 2026-10-24
Where it lives
- AgentDock Pro · private source repository
What proves it
- Worktree service on Pro main; branches aigency/<run>/<task> were created in later real runs
What it does not say
Merging the branches back is an agent task, not an automatic merge; the service only previews conflicts.
Not yet recorded
Show the branches and 'git worktree list' from a real run.
A built-in browser that an agent can drive, with URL limits, approvals and human takeover.
Checked 2026-10-27
Where it lives
- AgentDock Pro · private source repository
What proves it
- Acceptance run 2026-10-09 on the Intel developer preview: navigation, a refused forbidden URL, takeover and resume, against a local fixture page
- Recorded 27 Oct 2026 on the 0.1.6-dev.4 development build: Claude Code read nodejs.org and the Node.js release schedule on GitHub in the Agent Browser, asked before running a command, and wrote what it found to NOTES.md (footage-film, footage-chapters, footage-log)
- Real recordings, edited and sped up · public/film/hero-2400.mp4
The home page's opening film: a montage of four separate recordings made on 27 Oct 2026 on a private virtual display (Xvfb) of AgentDock Pro 0.1.6-dev.4, a development build, with a throwaway profile and projects, one scene per recording and never interleaved. The real Claude Code and Codex on the recording machine's own accounts; a script typed the requests and pressed Allow, Approve and Start. Hard cuts inside a scene and a dip through black between scenes; each shot keeps one constant speed (real time to 3×) shown in footage-2026-10-27.json, cropped and scaled; nothing retouched or generated. hero-1600.mp4 is a re-encode of the same cut; hero-m.mp4 is the same shots and timing with closer crops for phones. - Real recordings, edited and sped up · public/film/ch-browser-1600.mp4
The loops beside each part of the home page: the same shots as the opening film at the same speeds, each loop from a single recording. footage-2026-10-27.json lists their shots and speeds; each poster is the first frame of its loop, and the -m files are the closer phone crops. - Recorded by a real run · src/content/evidence/footage-2026-10-27.json
Every shot of the opening film and the loops: which recording, which seconds of it, the speed-up, the held frame, and what the recording shows. Written by the script that cut them (docs/footage-2026-10-27/build.py).
What it does not say
The recording is from a development build whose interface is newer than the preview's. Take control is visible on screen but was not exercised in the recording. Says nothing about sites that need a sign-in.
Media production: storyboards and motion graphics directed by an agent.
Checked 2026-10-24
Where it lives
- AgentDock Pro · private source repository
What proves it
- Real but test-grade outputs only (a 2-second title render); paid generation has never been exercised
What it does not say
No generated-video or voiceover claim: paid image, video and voice generation has never been exercised. Not shown or demonstrated on the home page.
Not yet recorded
A real, free-tier production recorded end to end, with the QC report.
AgentDock Pro can be downloaded today only as an unsigned developer preview for Intel Macs.
Checked 2026-10-25
Where it lives
- AgentDock · GitHub pre-release pro-developer-preview-0.1.6-dev.3 (2026-10-08) · AGDC-P Developer Build 01 Intel Mac DMG and SHA256SUMS
What proves it
- The pre-release downloads anonymously (checked 2026-10-24); its notes say 'unsigned, Intel-only, and not production-ready' and that automatic update installation is unavailable
- The free v0.1.1 and v0.1.2 downloads contain no Sarja sign-in or Pro code, so they cannot activate Pro
- The private Pro releases (including the Windows, Apple-silicon and Linux builds) are not publicly accessible
What it does not say
Says nothing about a signed or multi-platform Pro release, which does not exist. Nothing is emailed after payment and the account page does not hold a download; the pricing and account pages link the preview and say so.
NVIDIA NIM models, used with your own NVIDIA API key.
Checked 2026-10-27
Where it lives
- AgentDock Pro · private source repository
What proves it
- NVIDIA provider adapter and credential service in the developer preview; it reports 'NVIDIA NIM API key is not configured' until a key is entered
What it does not say
Signed-in end-to-end use has not been validated or recorded. AgentDock does not provide the key or the model access.
AIgency drafts a plan you can review and edit, and nothing runs until you start it.
Checked 2026-10-27
Where it lives
- AgentDock Pro · private source repository
What proves it
- The plan card recorded on 27 Oct 2026 before Start was pressed (aigency-plan-still)
- The same run in the opening film and the team loop: the request card, the plan, Start, and the tasks running (footage-film, footage-chapters)
- Real screenshot · public/session/aigency-plan-1080.webp
Crop of a frame recorded on 27 Oct 2026 on the same development build, when Claude Code asked AIgency to plan a three-task job in the same throwaway project: the plan card AgentDock shows before anything runs. Cropped only.
Original SHA-256 12c91e414c07fd8267e12bc1073caef0d548aa287a398c39542af08971eab78f - Real recordings, edited and sped up · public/film/hero-2400.mp4
The home page's opening film: a montage of four separate recordings made on 27 Oct 2026 on a private virtual display (Xvfb) of AgentDock Pro 0.1.6-dev.4, a development build, with a throwaway profile and projects, one scene per recording and never interleaved. The real Claude Code and Codex on the recording machine's own accounts; a script typed the requests and pressed Allow, Approve and Start. Hard cuts inside a scene and a dip through black between scenes; each shot keeps one constant speed (real time to 3×) shown in footage-2026-10-27.json, cropped and scaled; nothing retouched or generated. hero-1600.mp4 is a re-encode of the same cut; hero-m.mp4 is the same shots and timing with closer crops for phones. - Real recordings, edited and sped up · public/film/ch-browser-1600.mp4
The loops beside each part of the home page: the same shots as the opening film at the same speeds, each loop from a single recording. footage-2026-10-27.json lists their shots and speeds; each poster is the first frame of its loop, and the -m files are the closer phone crops.
What it does not say
Says nothing about how that recorded run turned out, and the footage stops before its end. Its verification task could not see the other tasks' changes, which ran in separate worktrees, and reported a failure, yet the run was shown as completed (AgentDock-Pro issue #89); the home page does not cite it as a result.
In development
Installing signed Arcotheon capabilities (.adcap) from inside AgentDock, with the permissions you approve.
Checked 2026-10-27
Where it lives
- AgentDock Pro · private source repository
What proves it
- Recorded 27 Oct 2026: Settings › Capabilities installed sarja/example-adcap 1.0.0 from the live Arcotheon registry after its install dialog was approved (footage-film, footage-log)
- Real recordings, edited and sped up · public/film/hero-2400.mp4
The home page's opening film: a montage of four separate recordings made on 27 Oct 2026 on a private virtual display (Xvfb) of AgentDock Pro 0.1.6-dev.4, a development build, with a throwaway profile and projects, one scene per recording and never interleaved. The real Claude Code and Codex on the recording machine's own accounts; a script typed the requests and pressed Allow, Approve and Start. Hard cuts inside a scene and a dip through black between scenes; each shot keeps one constant speed (real time to 3×) shown in footage-2026-10-27.json, cropped and scaled; nothing retouched or generated. hero-1600.mp4 is a re-encode of the same cut; hero-m.mp4 is the same shots and timing with closer crops for phones. - Recorded by a real run · src/content/evidence/footage-2026-10-27.json
Every shot of the opening film and the loops: which recording, which seconds of it, the speed-up, the held frame, and what the recording shows. Written by the script that cut them (docs/footage-2026-10-27/build.py).
What it does not say
Arcotheon itself is a live, separate site (agdc.dev/arcotheon). In the development build a capability's tools run only on Linux, where its sandbox exists, and the registry's root key is still a development key.
A desktop of the agent's own: an isolated Linux desktop it can start, run programs in and click through, with only the project folder shared.
Checked 2026-10-27
Where it lives
- AgentDock Pro · private source repository
What proves it
- Recorded 27 Oct 2026: Claude Code started its desktop, opened xcalc, clicked 1, 2, ×, 1, 2, = and reported 144 (footage-film, footage-chapters, footage-log)
- Real recordings, edited and sped up · public/film/hero-2400.mp4
The home page's opening film: a montage of four separate recordings made on 27 Oct 2026 on a private virtual display (Xvfb) of AgentDock Pro 0.1.6-dev.4, a development build, with a throwaway profile and projects, one scene per recording and never interleaved. The real Claude Code and Codex on the recording machine's own accounts; a script typed the requests and pressed Allow, Approve and Start. Hard cuts inside a scene and a dip through black between scenes; each shot keeps one constant speed (real time to 3×) shown in footage-2026-10-27.json, cropped and scaled; nothing retouched or generated. hero-1600.mp4 is a re-encode of the same cut; hero-m.mp4 is the same shots and timing with closer crops for phones. - Real recordings, edited and sped up · public/film/ch-browser-1600.mp4
The loops beside each part of the home page: the same shots as the opening film at the same speeds, each loop from a single recording. footage-2026-10-27.json lists their shots and speeds; each poster is the first frame of its loop, and the -m files are the closer phone crops. - Recorded by a real run · src/content/evidence/footage-2026-10-27.json
Every shot of the opening film and the loops: which recording, which seconds of it, the speed-up, the held frame, and what the recording shows. Written by the script that cut them (docs/footage-2026-10-27/build.py).
What it does not say
Linux only: the sandbox it runs in does not exist on macOS or Windows, and there is no Linux download yet. A Pro feature in the development build's entitlements.
Tools and capabilities an agent can use: skills, MCP servers, your own command-line and REST tools, agents to delegate to, and the @ menu that reaches all of them.
Checked 2026-10-27
Where it lives
- AgentDock Pro · private source repository
What proves it
- Recorded 27 Oct 2026: the @ menu listing skills, tools and agents, and Tools & Capabilities (footage-film, footage-chapters, footage-log)
- Real recordings, edited and sped up · public/film/hero-2400.mp4
The home page's opening film: a montage of four separate recordings made on 27 Oct 2026 on a private virtual display (Xvfb) of AgentDock Pro 0.1.6-dev.4, a development build, with a throwaway profile and projects, one scene per recording and never interleaved. The real Claude Code and Codex on the recording machine's own accounts; a script typed the requests and pressed Allow, Approve and Start. Hard cuts inside a scene and a dip through black between scenes; each shot keeps one constant speed (real time to 3×) shown in footage-2026-10-27.json, cropped and scaled; nothing retouched or generated. hero-1600.mp4 is a re-encode of the same cut; hero-m.mp4 is the same shots and timing with closer crops for phones. - Real recordings, edited and sped up · public/film/ch-browser-1600.mp4
The loops beside each part of the home page: the same shots as the opening film at the same speeds, each loop from a single recording. footage-2026-10-27.json lists their shots and speeds; each poster is the first frame of its loop, and the -m files are the closer phone crops. - Recorded by a real run · src/content/evidence/footage-2026-10-27.json
Every shot of the opening film and the loops: which recording, which seconds of it, the speed-up, the held frame, and what the recording shows. Written by the script that cut them (docs/footage-2026-10-27/build.py).
What it does not say
Which skills and tools appear depends on what is installed on the machine. A Pro feature in the development build's entitlements.
Automations: scheduled prompts and a calendar of what will run.
Checked 2026-10-27
Where it lives
- AgentDock Pro · private source repository
What proves it
- Calendar and scheduler services, and the Calendar view in the development build's sidebar
What it does not say
Not recorded or shown on the home page beyond its name; email reports are not claimed.
Not yet recorded
A scheduled prompt created and run on the development build.
In one recorded conversation, Claude Code made a change while asking before it edited files and ran commands, AgentDock showed the diff with a Revert button, and the same conversation then switched to Codex, which reviewed the change without editing anything.
Checked 2026-10-27
Where it lives
- AgentDock Pro · private source repository
What proves it
- The recording and its log (session-recording, session-log)
- Crops of the approval prompt, the Changed Files panel and Codex's review (session-stills)
- Real recording · public/session/session-2880.mp4
Screen recording made on 27 Oct 2026 of AgentDock Pro 0.1.6-dev.4 (a development build, not a public download) driving the real Claude Code and Codex, each signed in with the recording machine's own account, in a throwaway Node.js project. A Playwright script typed the two requests and pressed Allow. Waits were cut with hard cuts; nothing was sped up or retouched. The 1920 file is a re-encode of the same cut; the poster is one of its frames. - Recorded by a real run · src/content/evidence/session-2026-10-27.json
The recording script's own timestamps (seconds after the first request was sent) and the summary lines AgentDock printed for each agent's turn, copied from the window. The prompts are verbatim. - Real screenshot · public/session/approval-1560.webp
Rectangular crops, at the window's own 2x pixels, of frames from the same 27 Oct 2026 recording: the approval prompt, the Changed Files panel with its diff, Codex's review, and the model and permission menus opened before the first request. Cropped only; nothing retouched.
What it does not say
Recorded on a development build whose interface is newer than the current downloads. Approvals, the Changes view and handing a conversation to another agent are in the public releases (see approvals and handoff-and-changes), with the older interface. A script pressed Allow; the agents' work and words are their own. It does not claim Codex's review was right, only what it said.
Planned
A signed Pro download for Windows, Apple silicon and Linux, handed to paying customers from their account.
Checked 2026-10-25
Where it lives
- AgentDock Pro · private source repository
What proves it
- Pro's own release-readiness notes say a public release is not ready; no stable Pro release has been published
A Linux build.
Checked 2026-10-24
Where it lives
- AgentDock · README: AppImage 'experimental'; website releases.ts: coming-soon
What proves it
- No Linux asset in any public release