Skip to main content
Most ML work now happens in a conversation with an agent. The agent plugin makes that conversation part of the record rather than something that evaporates when the terminal closes.
Supported hosts: Claude Code, Codex, pi.

Two plugins, two decisions

They are separate because they are separate consent decisions.

probe-research

Tracking. Skills that teach the agent the experiment workflow, and a read-only MCP server it searches your lab’s history with.Writes go through the probe CLI, never through MCP.

probe-research-tap

Session capture. Streams this device’s sessions to the team knowledgebase: your prompts, the agent’s replies, and the shell commands you run.Per-device, disclosed twice before it starts, off at any time.

What the agent can do with them

Before proposing a direction, the agent can check whether somebody already tried it — across runs, notes, sessions, papers and connected sources. That is the read half, and it is available even with writes turned off.
Create the project and experiment explicitly, open runs, route files to artifacts and numbers to metrics, write decisions into notes, and close with the real lifecycle outcome. At the moment it happens — not summarised at the end of the session.
/instrument-code wires probe.init and probe.log into a trainer, an eval or a sweep, including trainers that already have their own tracker.
/visualize-progress renders the tracked state, the gaps, and the arc as one timeline with the next action named — every line read back from Probe, never from the session’s own memory.

The switch

One control governs the whole conversation.
Only the researcher moves the switch. Typed bare by you it toggles between on and read; off is not on that cycle and is reached only by typing it. An agent invoking the skill itself loads guidance and never changes the state.
The switch has three scopes, and each has a natural home.
Type it in the agent:
Or from a shell, against the current session:
probe session status returns the full decision, including which default applied and whether the capture daemon is actually running — so “tracking is on” and “anything is being recorded” can be told apart.

What each surface is allowed to do

The read/write split is structural, not a policy setting. An agent with only the MCP server cannot change anything.

The managed rules block

The install writes a managed Probe block into each selected agent’s global instruction file, prompting it to search Probe and track research there. Everything outside the block is left alone.
Tick it on or off. The wizard applies it to every agent you have configured, and leaves untouched rows exactly as they were.

Skills

Every skill, when it fires, and what it does.

MCP reference

Six tools, token budgets, pagination, entity views.

Session capture

What leaves the machine, and how to stop it.

Status line

The tracked/untracked indicator in Claude Code.