OpenTacit operates the same on all surfaces. This chapter tells you about this shared behavior. The chapters that follow give the details of each task.
OpenTacit connects to your work through surfaces. Most surfaces are harnesses: AI coding tools with a small OpenTacit integration. The integration lets OpenTacit monitor a session and give an occasional suggestion. One surface is different: a chat connector that you add to Claude or ChatGPT. The connector installs nothing, and it lets the model pull from the playbook on demand (Connect a chat tool). The behavior in this chapter applies on all surfaces that can supply it. The surfaces are different only in how much of this behavior each one supports.
OpenTacit monitors each session silently and at no cost. It gives suggestions rarely, and only with evidence. It answers directly each time you ask.
The two modes of operation
OpenTacit operates in two modes:
- OpenTacit listens. This mode is automatic and always on. OpenTacit monitors the session. When OpenTacit has information of sufficient value, it gives one suggestion that you did not request. This mode uses your attention, so OpenTacit limits it strictly: a small number of suggestions in each session, with no repeats. → Work with suggestions.
- You ask OpenTacit. Use
@tacitin a prompt, a/tacit:command, or a tool that your agent calls. Because you asked, this mode has no cost and no limit. → Search and review on demand.
This difference is important because a surface can support one mode without the other. A harness supports the two modes. A chat connector supports only the second mode. The connector has no hooks, so no suggestion comes unless you or the model asks. This one fact causes most of the differences between surfaces.
In the two modes, OpenTacit does not make a turn slower and does not add unwanted output. All heavy operations occur in the background. If the registry or the model is slow or not available, OpenTacit becomes silent. It does not cause a delay, and it does not put an error into your work.
What the connection installs
When you connect a harness, OpenTacit installs a small integration package. The package has a maximum of four parts. All parts read one settings file (~/.config/tacit/agent.env), so they use the same registry and key:
- Lifecycle hooks — these let OpenTacit monitor your session and give an occasional suggestion. The hooks do not change your prompt or your tools.
- Skills / slash commands — commands that you start by name to pull from the playbook. They include search, review, audit, contribute, drafts, insights, and org. They also include the operational commands status, test, and help. The chapters that follow describe them.
- An MCP server — this gives your agent direct access to OpenTacit tools (
tacit_search,tacit_metrics,tacit_drafts,tacit_draft_action). On hosts that support panels, the tools show as interactive panels. On all other hosts, they show as text. - A status-line segment — on harnesses that have a status line, the segment shows session data (the status line).
To disconnect, run tacit disconnect. The command removes all the parts that the connection installed, and no other files.
A chat connector is different: it installs nothing on your machine. You add one URL and sign in. Only the pull tools are available. To remove the connector, delete it in your chat app. See Connect a chat tool.
Where OpenTacit can and can't render
The behavior is the same on all surfaces. The difference is where a pushed suggestion can appear. OpenTacit sends the ◆ block as hook output, and a surface without hooks cannot show the block. This is the one important difference between surfaces:
| Surface | Pushed ◆ suggestions | If they cannot appear |
|---|---|---|
| Claude Code (terminal), Codex, Amp, pi/omp, opencode | Yes, below the answer (on Amp, with your next message) | — |
| Cursor | Yes, but always with your next prompt | there is a delay of one turn |
| Claude Code on web or phone | No. These clients do not show hook output | use pull: @tacit, /tacit:review, or the tools; all show as usual chat text |
| GitHub Copilot CLI | No | OpenTacit continues to monitor; pull with the tacit-* skills or the tacit-review agent |
| Chat connector (Claude, ChatGPT) | No. A connector has no hooks, so OpenTacit does not push or monitor | this surface is pull-only by design; ask in words, or use the tools, which show in the chat |
The pull path operates on all surfaces, also where the push path cannot. To make sure that OpenTacit is connected, ask it: @tacit status.
What stays private
This design is intentional. It is the reason that OpenTacit can operate in your editor and keep your privacy:
- The registry stores no identity and no conversation text. The data that leaves your machine is a structured characterization (the task type, the tools in use, your cohort, for example
team:revops) plus feedback events. OpenTacit uses the retrieval text only to find techniques, and it does not store this text. - All reports are by cohort, never by person. A line such as "helped 94% · n=120 · team:revops" is a cohort statistic.
- Your record of the suggestions that you adopted or dismissed stays only on your machine (
~/.tacit-technique-memory.json). Its function is to prevent repeated suggestions. - OpenTacit removes secrets and PII from all text that you contribute, before the text leaves your machine.
Give feedback gives the full contents of a feedback event, and describes the optional sketch opt-in.
More information
- Work with suggestions — the ◆ block,
@tacit, and the status line. - Search and review on demand — how to pull from the playbook on demand.
- Contribute a technique — capture a move that worked, so colleagues can use it again.
- Give feedback — how OpenTacit records the evidence for each technique.