ActVox can connect supported sources, platforms, and agent tooling to a workspace. Connections are scoped to the workspace and subject to the connected provider’s permissions, availability, quotas, and policy.

### Connect your AI

Open **Connectors** in the sidebar and choose **Connect your AI**.

1. Pick the AI you use. Some run in your browser (Claude, ChatGPT); others run in your terminal (Claude Code, Codex).
2. For a terminal client, copy the prompt it shows and paste it into the client. It sets up the connection itself and opens an ActVox window. For a browser client, add ActVox to it as the steps show.
3. In the ActVox window, choose this workspace and click **Allow access**. Read and write are chosen for you; **Choose permissions** changes them.
4. The flow says the AI is connected. Copy the first task it offers and paste it into the AI. It asks you a few questions, then saves its plan here as a page, and **Open it** takes you there.

Each connected AI has its own **Settings**: rename it, turn on **Mentions and tasks** so it can be @mentioned and assigned work, and choose who can use it.

### Availability vocabulary

The UI uses these exact availability states: `supported`, `canary_passed`, `verified`, `verified_read`, `documented`, `unavailable` and `blocked`. `verified` and `verified_read` are permanent approvals based on real production connections: `verified` means a write was confirmed in production, `verified_read` means reads only. A label says how far ActVox has verified a client end to end. It is not the access level: a connection's permissions are what you approve at consent. The state is documented separately for each product: MCP grants authenticated tools and workspace data; a skill supplies instructions for a local agent.

### AI-agent availability

| AI agent | MCP | Skill |
| --- | --- | --- |
| **ChatGPT** | **verified** | — |
| **Claude (hosted)** | **verified** | — |
| **Claude Code** | **verified read** | **supported** |
| **Codex** | **verified** (through ChatGPT’s connector) | **supported** |
| **Hermes** | **verified** through the ActVox CLI | **supported** |
| **Grok** | **verified** | — |
| **OpenClaw** | **verified** through the ActVox CLI | No catalog skill or installer |

Setup steps for each agent are on the [Agents page](/agents#agents-connect).

### Keep the three stores separate

Profile, workspace instructions, and the Knowledge Base are three separate stores. A connection does not merge them into one memory blob.

1. Profile stores the business portrait and identity fields.
2. Instructions store how the agent should work with you.
3. The Knowledge Base stores sources such as sites and documents; indexing can spend credits.

### ChatGPT — MCP — verified

- Add ActVox as a custom connection in ChatGPT Developer mode, then authorize the hosted MCP endpoint `https://api.actvox.com/mcp` with OAuth.
- Start with read access for one workspace. ChatGPT has no ActVox skill installer; use the Codex tile for a local skill.

### Claude (hosted) — MCP — verified

- In Claude, open Customize → Connectors → **`Add custom connector`**, paste `https://api.actvox.com/mcp`, and choose **`Sign in now`** with Claude’s published OAuth identity.
- Team and Enterprise owners can add the connector once in organization settings, but each member still signs in and consents individually. Claude Code is a separate tile.
- Approved permanently on 2026-10-03 from production use: live connections wrote and the result was confirmed by read-back, so its setup panel reads **Integration: Tested**.

### Claude Code — MCP `verified_read`; skill supported

- The hosted MCP path is approved permanently at **verified read**: live Claude Code connections read workspace data in production, and no write through its own connection has been confirmed yet. Run `claude mcp add --transport http actvox https://api.actvox.com/mcp`, then `claude mcp login actvox` and complete consent.
- The installable skill is supported for the `claude` target. Use the Claude Code tile for the installer, key setup, and local verification.

### Codex — skill supported; MCP verified

- The installable skill is supported for the `codex` target. Use the Codex tile to install it, issue a workspace key, and run the read-only doctor and capabilities checks.
- Hosted MCP is approved permanently: Codex’s production use runs through ChatGPT’s connector and includes writes confirmed by read-back. To connect the Codex CLI with its own identity, run `codex mcp add actvox --url https://api.actvox.com/mcp`, then `codex mcp login actvox` and complete consent. A local skill install is not hosted MCP support.

### Hermes — skill supported; verified through the ActVox CLI

- The installable skill is supported for the `hermes` target. Use its tile for the skill installer and workspace-key setup.
- Hermes works with ActVox through the skill and the ActVox CLI (a workspace key or a CLI login). That path is approved permanently on production writes, and its setup panel reads **Integration: Tested over the ActVox CLI**; the tile itself reads **Connected** once you connect one here.
- Hermes’ own hosted MCP connection stays available: run `hermes mcp add actvox --url https://api.actvox.com/mcp --auth oauth`, then `hermes mcp login actvox` and complete consent. The skill’s instructions do not grant workspace data by themselves; keep the key in the agent’s protected environment.

### Grok — MCP — verified

- In grok.com, open Connectors → New Connector → Custom and add `https://api.actvox.com/mcp` with the OAuth values shown on the Grok tile.
- Grok has no ActVox CLI or skill installer. The hosted connection is approved permanently: live Grok connections wrote in production and the result was confirmed by read-back.

### OpenClaw — verified through the ActVox CLI; no installer

- OpenClaw works with ActVox through the ActVox CLI (the ActVox CLI tile: a CLI login or a workspace key). That path is approved permanently on production writes, and its setup panel reads **Integration: Tested over the ActVox CLI**; the tile itself reads **Connected** once you connect one here.
- Its own hosted MCP connection stays available: run `openclaw mcp add actvox --url https://api.actvox.com/mcp --transport streamable-http --auth oauth`, then `openclaw mcp login actvox` and complete consent. There is no catalog skill or installer target for it.
- The manifest still lists `local_skill` as a limitation, but OpenClaw has no catalog `skill: true` flag and is not an installable skill client. Do not claim it is supported or add an installer.

### Short first test

After a supported skill or an available MCP connection is ready, send this read-only Knowledge Base smoke test:

> `Run a read-only Knowledge Base smoke test. Tell me which workspace you can access and list the titles of the first three Knowledge Base sources. Do not write or save any personal information unless I explicitly confirm the exact fields and values first.`

This is a smoke test, not the five-step onboarding. If onboarding follows, ask before each write and read back the profile, instructions, and Knowledge Base changes separately.

### Manage and remove

- For MCP, the **Connectors** page shows the grant, last call, and verified effect; pause, resume, or revoke access there, then remove the connection in the provider.
- For a skill, revoke its workspace key on the Connectors page and remove the installed skill from the local agent.
- If a consent page says the client is not approved, the current tile policy is authoritative. Retrying does not change a documented, unavailable, or blocked state.

### Your agents and your role

- Anyone with the editor role or above can connect their own agent, through a consent page or with a key on a connector tab. You name the agent when you connect it. It is private to you until an owner or admin shares it with the workspace, and only you see its connections.
- An agent never does more than the person who connected it can do now. An editor's agent cannot share, publish or change workspace settings; if you become a viewer, your agent can only read; if you leave the workspace, it stops working.
- On Free, AI connectors count for the whole organization: a product another member already connected adds no new connector.

### Automation boundaries

An available connection does not mean every action is automatic or approved. Publishing, paid activity, destructive updates, and high-impact external actions can require explicit approval. ActVox does not guarantee external platform availability or campaign outcomes.

---