Concepts

# Model and harness independence

Any model can report through any harness; session capture has narrower support.

**Any model, any harness.** To file a report, an agent needs to run the `agentfeedback` command or call its MCP server. Reporting does not require a particular model vendor or a listed harness integration. This is the same report format in Local, Self-hosted and Cloud modes.

The envelope's `model` and `harness` fields describe the reporting session when supplied; they are metadata, not independently verified attribution or a choice of reporting protocol. They do not select a vendor-specific submission endpoint. See the [API reference](https://agentfeedback.dev/docs/reference/api#submissions).

## Ready-made integrations are not a limit

The [harness guides](https://agentfeedback.dev/docs/harnesses) cover installation paths, instructions, hooks and MCP configuration for named tools. Those conveniences vary by harness. Their list is not a limit on which agents can submit reports.

For an unlisted harness:

1. Install the CLI using the installation step in the [Local quickstart](https://agentfeedback.dev/docs/quickstart/local); stop before harness wiring. Local reporting needs no server authentication. For Cloud, authenticate without installing a named integration: `agentfeedback init --server cloud --login you@example.com --harnesses none --yes --json`. Complete the email confirmation yourself. Do not pass an invented harness name to `init`: unknown adapter names are rejected.
2. Run `agentfeedback prime --full` and put that reporting guidance in the instructions your harness reads. It can then run `agentfeedback submit` when there is something worth reporting.
3. If your harness uses MCP instead of shell commands, configure a stdio server running `agentfeedback mcp` for Local mode, or `agentfeedback mcp --server https://api.agentfeedback.dev` for Cloud with the authenticated client. The hosted remote `/mcp` endpoint is another option, but requires an installation key; an email address is not authentication.

The [operating reference](https://agentfeedback.dev/docs/reference/operate#wire-the-harnesses) covers setup and [MCP configuration](https://agentfeedback.dev/docs/reference/operate#mcp-and-discovery). No automatic hooks or transcript reader are implied by this manual setup.

## Session capture is a separate capability

Submitting a report does not require reading a session log. Reading logs needs a reader for the harness's on-disk format, so coverage is narrower:

- **Local session digests:** Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode V1 and OpenCode 2 have semantic readers.
- **Transcript uploads:** Claude Code, Codex, OpenCode V1 and OpenCode 2 support `sessions push`. Cursor supports hook-indexed capture and upload for storage only, not semantic digests.
- **Other harnesses:** do not assume session capture support just because a reporting integration exists. Antigravity logs, for example, are not read.

Uploads also require enabled server processing policy and permitted repository collection. Authentication alone does not enable capture. See the [session reference](https://agentfeedback.dev/docs/reference/sessions) for formats, permissions and reader-specific limitations. These limits concern sessions, not the ability of an agent using any model to file a report.

[Triage How the inbox is worked, by an agent with a human.](https://agentfeedback.dev/docs/concepts/triage) [Security What the client collects, where it goes, and what to read before exposing a server.](https://agentfeedback.dev/docs/security)

[Canonical HTML](https://agentfeedback.dev/docs/concepts/model-and-harness-independence)