Coding agent tracing
Connect coding agents with one setup command and the same credentials, privacy controls and diagnostics. Compare available local integrations and unsupported surfaces.
Connect your coding agent to Future AGI with one setup command. See each prompt, its model activity and tools in Observe, with conversation turns grouped in Sessions.
Choose your agent
Every available integration uses npx @traceai/agents setup <agent> --project <name>. The examples below send traces to a project named coding-agents. Install and sign in to the agent first; tracing does not provide model access.
Availability is specific to the surface listed. A supported CLI does not imply support for its desktop app, hosted agent or gateway. Follow the linked guide for supported host versions, operating systems and the correct run command.
| Agent | Setup command | What is captured | Status |
|---|---|---|---|
| Claude Code CLI | npx @traceai/agents setup claude --project coding-agents | Prompts, reconstructed generations, tokens, tools and subagents. | Available |
| Codex CLI | npx @traceai/agents setup codex --project coding-agents | Prompts, model attempts, tokens, tools and nested subagents. | Available |
| Cursor IDE | npx @traceai/agents setup cursor --project coding-agents | Prompts, reconstructed generations, turn usage and tools. | Available |
| Cursor interactive CLI | npx @traceai/agents setup cursor --project coding-agents | Shared IDE hooks; late transcripts complete generations. Print mode is partial. | Available |
| GitHub Copilot CLI | npx @traceai/agents setup copilot --project coding-agents | Native model calls, tokens, tools and emitted subagent relationships; use run copilot. | Available |
| Antigravity CLI | npx @traceai/agents setup antigravity --project coding-agents | Prompts, transcript generations, tokens, tools and native subagents. | Available |
| Grok Build headless CLI | npx @traceai/agents setup grok --project coding-agents | Prompts, reconstructed responses, turn usage, tools and subagents. | Available |
| OpenCode V2 | npx @traceai/agents setup opencode --project coding-agents | Prompts, model steps, usage, tools and native subagents. | Available |
| OpenCode V1 | npx @traceai/agents setup opencode-v1 --project coding-agents | Prompts, model steps, usage and tools; explicit task-child links. | Available |
| Goose CLI | npx @traceai/agents setup goose --project coding-agents | Prompts, saved model generations, usage and tools. | Available |
| Hermes Agent CLI | npx @traceai/agents setup hermes --project coding-agents | Prompts, provider attempts, usage, tools and in-process subagent relationships. | Available |
| OpenClaw local agent | npx @traceai/agents setup openclaw --project coding-agents | Native model and tool spans beneath a prompt; use run openclaw. | Available |
| Pi | npx @traceai/agents setup pi --project coding-agents | Prompts, model requests, usage, tools and automatic compaction. | Available |
| Oh My Pi | npx @traceai/agents setup oh-my-pi --project coding-agents | Prompts, reconstructed generations, usage, tools and task subagents. | Available |
| VS Code Copilot | npx @traceai/agents setup copilot-vscode --project coding-agents | Copilot Chat model calls, tokens and tools through the Future AGI VS Code extension; reload VS Code after setup. | Available |
| Antigravity desktop app (macOS) | npx @traceai/agents setup antigravity --project coding-agents | Prompts, model steps, Gemini usage and tools; earlier model steps may lack text. | Available |
| Antigravity desktop app (Windows, Linux) | — | The plugin setup is tested on macOS only. | Not yet available |
| Cursor Cloud | — | Local hooks do not enable hosted-agent tracing. | Not yet available |
| Hosted Codex | — | Local hooks do not run in hosted tasks. | Not yet available |
| GitHub hosted coding agent | — | Copilot CLI setup does not configure hosted tasks. | Not yet available |
| Goose Desktop | — | CLI setup does not configure the desktop app. | Not yet available |
| Hermes desktop / messaging gateways | — | The CLI integration does not cover these surfaces. | Not yet available |
| OpenClaw Gateway / channels / TUI / Control UI | — | The supported route is local agent --local through the launcher. | Not yet available |
| Grok interactive TUI | — | Use the supported headless CLI route instead. | Not yet available |
| Gemini CLI | — | The shared setup integration is not available yet. | Not yet available |
| Kiro CLI / IDE | — | The shared setup integrations are not available yet. | Not yet available |
| Augment CLI / IDE | — | The shared setup integrations are not available yet. | Not yet available |
| Devin Local / Cascade / hosted | — | The shared setup integrations are not available yet. | Not yet available |
| Muse / Cowork | — | No shared setup route is available yet. | Not yet available |
| DeepSeek Harness | — | No coding-agent setup route is available yet. | Not yet available |
Three steps, every time
Get your Future AGI keys
Create an API key and secret key on the keys page. Keep them out of agent conversations, command arguments and repository files.
Set up tracing
Run the command for your agent from its repository. Enter the keys at the hidden prompts and review the destination before approving reuse of saved credentials. Complete any native plugin trust prompt yourself; setup does not bypass the agent’s permission checks.
Credentials are saved once in ${XDG_CONFIG_HOME:-$HOME/.config}/futureagi/agents/credentials.json, protected with owner-only permissions. Agent preferences remain separate from the shared key store.
Warning
Content capture is on by default: prompts, model responses and tool input/output can include source code and secrets the agent reads. Add --no-capture-content to setup to save a content-off preference, or set FI_AGENT_CAPTURE_CONTENT=false in the agent’s launch environment for a process-level opt-out. Structure and available usage remain visible.
Run the agent and open the trace
Restart the agent and use the run command in its guide. Copilot CLI and OpenClaw require the maintained npx @traceai/agents run launcher; other supported local routes use their native command after setup. Ask the agent to summarize a non-sensitive README, let it finish, and open Observe in Future AGI. Some hosts deliver only on the next prompt, session exit or after a late transcript; each guide explains when to expect the trace.
Shared controls
Replace <agent> with the setup name from the table:
npx @traceai/agents status <agent> --check
npx @traceai/agents doctor <agent>
npx @traceai/agents disable <agent>
npx @traceai/agents enable <agent>
npx @traceai/agents update <agent>
npx @traceai/agents uninstall <agent>
Status separates configuration, native registration, observed activation and export delivery. Doctor checks configuration and connectivity. HTTP acceptance is not proof that a trace is visible: open Observe to confirm it.
Commands read the current repository by default. --user-only ignores repository overrides for that command; --project-dir <path> selects another repository. Do not combine them. --project <name> names the Future AGI destination, not an installation scope. Native user/project installation support differs by agent; follow its scope section instead of assuming --scope project is supported everywhere.
All maintained setup routes use FI_API_KEY, FI_SECRET_KEY, FI_PROJECT_NAME, FI_USER_ID, FI_AGENT_CAPTURE_CONTENT, FI_AGENT_ENABLED, FI_BASE_URL, FI_OTLP_ENDPOINT and FI_AGENT_CONFIG_DIR. Supply both credential variables together. Saved credentials require approval before a project or destination change. A repository cannot increase your user-level content permission.
Read a trace consistently
A prompt is the unit of work; a session groups prompts. Available model activity and tool calls appear beneath the prompt, with explicit subagent relationships when the host provides them. Model requests and transcript-reconstructed generations are not interchangeable: reconstructed timestamps do not establish network latency, and turn-level usage is not a per-request token count. Each guide explains its source.
Usage is counted once, not on both a parent summary and its children. Missing usage, cost, ancestry or outcomes are not guessed. The shared outcome labels are success, error, interrupted and incomplete. Failed or interrupted work is distinguished from success; incomplete lifecycle data is not a successful result. Host-specific diagnostics remain visible rather than being converted into invented model calls.
Migrate from copied hooks
Remove only the old Future AGI hook or collector registration described in your agent’s guide before setting up the maintained package. Keep unrelated plugins, hooks and permission guards. Do not run two exporters for the same agent session. The old Hook-to-OTLP emitter recipe is retired; no new setup requires copying its source.
Privacy
Redaction is best effort, not a guarantee that arbitrary secrets or personal data are removed. Content fields are capped at 20,000 characters; a structured (JSON) value longer than that may be omitted as a whole rather than shortened, so that no partial secret survives the cut. Use content-off mode for sensitive work. Structural metadata can still contain model names, tool names and session identifiers. Disable and uninstall do not delete data already accepted by Future AGI or change your agent’s own session-history retention.
Exports are bounded and fail open. Agent work continues when tracing fails, but forced termination or an unavailable destination can lose telemetry. The integrations do not promise exactly-once delivery.
Building your own agent
Application SDK instrumentation is separate from tracing a coding-agent installation. See the Codex SDK, Cursor SDK, Pi SDK, or Claude Agent SDK guides.
Questions & Discussion