Agent Setup

Agent Integrations

Quick answer

Prism detects which supported AI coding agents are installed and wires them up with one click. It backs up the existing config, writes a clearly marked Prism-managed block with the right base URL, token and model catalog, and re-syncs that block whenever your models change.

Last updated 2026-10-06Reviewed against Prism v0.3.26

How one-click setup works

Stage	What happens
Detection	On startup Prism looks for each supported agent's config file or binary and lists the ones it finds in the Agents tab.
Setup	Clicking Setup copies the existing config to <config>.prism-backup, then writes the Prism block or entries.
Auto-sync	Once enabled, Prism rewrites the managed block on startup and whenever you add or remove a model, so agent pickers stay current.
Smart routing	Codex OAuth models are routed through /v1/responses; every other model goes through /v1/chat/completions. Each agent only ever sees the dialect it expects.
Disable	Removes only Prism-tagged entries and restores the non-Prism parts of your config untouched.
MCP access	All 14 agents can also be granted access to the MCP gateway through their own MCP config file.

Nothing is destructive

Prism only ever edits blocks it tagged itself. Your own models, providers and settings elsewhere in the file are preserved, and the pre-setup file is kept as a backup.

Supported agents

Agent	What Prism writes	Config file
Claude Code	ANTHROPIC_* environment keys plus per-tier model mappings	~/.claude/settings.json
Codex Desktop & CLI	A managed provider block and a model catalog JSON	~/.codex/config.toml
Cursor	Base URL and API key overrides in Cursor's model settings	Cursor settings (manual)
Factory Droid	[Prism]-tagged custom model entries	~/.factory/settings.json
OpenCode	prism and prism-responses providers, plus a default model	~/.config/opencode/opencode.json
ZCode	A provider block with limits, modalities and reasoning	~/.zcode/v2/config.json
Grok Build	One [model.prism-*] block per model, with api_backend per model	~/.grok/config.toml
Zed	language_models.openai_compatible settings (JSONC)	Zed settings.json
Pi	Provider entries and a model list	~/.pi/agent/settings.json and models.json
Oh My Pi (OMP)	prism and prism-responses providers	~/.omp/agent/models.yml
Kimi Code	prism and prism-responses providers	$KIMI_CODE_HOME/config.toml or ~/.kimi-code/config.toml
Prime Agent	Provider entries and a model list, with WSL detection	~/.prime/agent (or $PRIME_AGENT_CODING_AGENT_DIR)
Empryo	Provider entries keyed prism/<provider>/<model>	%LOCALAPPDATA%\Empryo\config.json (Windows) or ~/.empryo/config.json
Hermes	prism and prism-responses providers	%LOCALAPPDATA%\hermes\config.yaml (Windows) or ~/.hermes/config.yaml
DeepSeek Harness	Provider routes merged into the llm-pi-ai row, for every profile	$DSH_HOME (default ~/.dsh)

Manual clients

Some clients are not detected or written automatically, but work perfectly with a manual base URL. See their setup pages:

Claude Desktop

Point Claude Desktop at Prism through its gateway provider setting.

OpenAI SDK

Change the base URL in Python or Node and use prism as the key.

Ollama-compatible clients

Reuse port 11434, or move Prism to another port if Ollama is running.

Smart routing in detail

A model's protocol is decided once, in Prism, and then respected by every agent integration. Models served by a Codex OAuth account — and any model explicitly marked api: "responses" — are written into the agent as a Responses endpoint. Everything else is written as a Chat Completions endpoint. In agents that support two provider slots, Prism registers both, which is why you often see prism and prism-responses side by side.

Grant an agent MCP access

The MCP tab's Agent access panel writes a Prism entry into each agent's own MCP config, carrying the header Authorization: Bearer prism. The agent then reaches every server you allowed through /mcp/<agent>, and never sees the upstream credentials. See MCP Gateway for the per-agent file list.

Per-agent setup guides

Claude Code

Route Anthropic Messages to any provider, with per-tier models.

Codex Desktop & CLI

Inject custom models into Codex's native picker.

Cursor

Two-field base URL override, plus MCP.

Factory Droid

[Prism]-tagged custom models in settings.local.json.

OpenCode

ZCode

An Anthropic-kind provider with working server tools.

Grok Build

One [model.prism-*] section per model in config.toml.

Zed

An OpenAI-compatible language model with per-model capabilities.

Pi

Provider blocks plus a generated model list with thinking levels.

Oh My Pi (OMP)

YAML providers with a vision compatibility carve-out.

Kimi Code

TOML providers and one model alias per model.

Prime Agent

models.json providers, with WSL detection on Windows.

Empryo

A custom provider whose token lives in the OS keychain.

Hermes

A managed YAML region with two provider transports.

DeepSeek Harness

One write that covers every DSH profile.