Skip to main content
The Cogniscape CLI captures sessions from five AI coding agents on the developer’s machine: Claude Code, Codex, Cursor, OpenCode, and Kiro (IDE, CLI v3, and Crew). One install covers all of them. Hooks are added for every agent found on the machine. An agent installed later is picked up automatically, within about an hour of the next captured session — there is nothing for the developer to run. The agents do not expose the same data. This page is the single reference for what each one delivers.
This page describes CLI 1.20.2. Machines update themselves within 24 hours of a new release.

​
Capability matrix

How to read the cells:
  • Yes — captured and verified against the real agent.
  • No — … — the agent does not provide the data. Cogniscape cannot capture what the tool never exposes.
  • Not yet — the agent has the data, but the CLI does not read it today.
¹ Only when a specific model is selected. When the agent runs on Auto, it does not say which model answered, so no model is recorded. ² The Kiro Crew session file can carry the selected model, but no real capture with a selected model has been verified yet. Credits are captured. ³ On Codex CLI 0.145.0 or later. Older Codex versions do not fire a session-end hook. ⁴ Sessions without an end event are closed automatically after 2 hours without activity. Every completed turn is already recorded when it ends, so nothing is lost.

​
Missing tokens are not zero cost

When an agent does not report tokens — Cursor and Kiro — or reports a model with no known price, the cost appears as not reported, never as zero. A Cursor session did cost something; Cogniscape simply has no number for it. Kiro sessions show the Kiro credits they used instead of a dollar amount. Keep this in mind when you compare cost or ROI across developers who use different agents.

​
Agent requirements

Most agents need nothing beyond the install. The exceptions:
Kiro runs hooks only in trusted workspaces. Trust the project folder and restart the IDE.
Activity is captured on the v3 engine (kiro-cli chat --v3). On the v2 engine, the built-in default agent does not support hooks. Your own global v2 agents do fire hooks, but Kiro does not send a session id on v2, so only a turn marker is recorded — no summary, tools, model, or credits.
Kiro Crew reads its hooks at startup. Restart Kiro Crew once after installing Cogniscape.
Session end is captured on Codex CLI 0.145.0 or later. Everything else works on earlier versions.
Cogniscape installs a small plugin in ~/.config/opencode/plugins/. At the end of each turn, the CLI reads OpenCode’s local database (~/.local/share/opencode/opencode.db) in read-only mode to collect the summary, tokens, tools, commands, and files.

​
Cloud agents are not captured

The CLI captures agents that run on the developer’s machine, through local hooks. Agents that run in the vendor’s cloud never pass through those hooks, so their sessions are not captured:
  • Cursor background agents and Cursor Projects
  • Claude Code on the web
  • GitHub Copilot coding agent
  • Codex cloud tasks
The work these agents produce still reaches Cogniscape through the GitHub and GitLab webhooks: pushes, pull requests, and reviews. Those events carry no AI session — no summary, tokens, or tool usage.

​
Privacy

Your prompts are never captured, for any agent. See Developer Privacy for exactly what is collected and why.