How it works
How your agents use a project's context, skills, and tools to work together through Venom.
The plugin connects your agent.
Install Venom Full in Claude Code or Codex and authorize it for your organization. It installs the venom plugin, which connects your agent to Venom’s MCP server, and adds hooks that ask the server for reminders at key moments. In Claude chat and Cowork, the venom plugin from the plugin directory makes the same connection without hooks. Other clients can use an MCP connection with their own instruction files; see client setup.
The server tells every connected agent how to use Venom: what to read, when to create a task, and what is worth keeping. Your organization’s own practices live in shared skills that admins update without distributing new instructions to every teammate.
In Claude Code and Codex with Venom Full, the agent is shown your projects and unread mentions on the first prompt of a session. When a turn reached a milestone and nothing was logged to Venom, it is asked once to log what changed. Other clients get their guidance from the server’s instructions and the reminder lines in tool responses.
The project brings the work together.
| Part | What your agent uses it for |
|---|---|
| Overview and tasks | Understand the goal, previous decisions, current work, and next steps. |
| Skills | Follow shared practices, such as how your team reviews a change. |
| Connected tools | Use the MCP servers available for the project through the same Venom connection. |
| Artifacts | Save and read substantial results, such as a plan or review. |
People manage this context in the web app. Agents read and update it through tools, using the permissions of the person who authorized the connection. Some connected services require each person to sign in separately.
A session, end to end.
- ReadThe agent reads the task and relevant skills.
- WorkYou and your agent work in your usual tools.
- SaveKeep decisions, useful results, and next steps.
- ContinueA later session reads what was saved.
For example, Jordan’s agent records a design decision and an open security question. Sam’s agent reads them before reviewing the linked code, then adds its findings to the task. Jordan’s next session can continue from that review.
The shared record contains useful context, not a copy of every conversation. Code stays in its repository; a task links to it. A session with nothing useful to add may make no writes.
Why it’s built this way.
- The guidance lives on the server. The plugins are the same for every organization, and Venom Full’s hooks fetch their wording from Venom, so a change reaches every connected agent without reinstalling anything. Your own conventions live in published skills.
- Reminders only when they help. Hooks fire at a few moments, such as after compaction, when a subagent starts, on a new mention, or after a commit or pull request, and add nothing when there is nothing to say. Clients without hooks get short reminders in tool responses.
- Saving less, on purpose. A log is useful only if people and agents read it. Agents save decisions, reasons, blockers, and next steps, and link to code instead of re-explaining it.
- Safe when several agents write at once. Writes carry the version that was read, so a conflicting change is rejected and reconciled rather than overwritten. A request ID makes retries safe. See Write safely with others.
What the reminders do.
The plugin supplies instructions; the agent makes the tool calls. Reminders don’t guarantee a save or recover unsaved chat. Your client’s approvals and disabled-skill settings still apply. Agents are instructed to treat saved material as context, not commands.
For event details and troubleshooting, see Hooks. To try the workflow, follow the Quickstart.