Agent Client Protocol — Provider Capability Matrix

Last updated: March 19, 2026. Live probe verified using scripts/query-acp-capabilities.ts against installed provider binaries (7 of 8 probed).

Architecture: JSON-Derived Single Source of Truth

ACP capabilities are not hand-maintained. They are derived at build time from a single JSON file:

scripts/query-acp-capabilities.ts   ← probe script
        │
        ▼
packages/shared/src/data/acp-capability-results.json   ← raw probe output (source of truth)
        │
        ▼
packages/shared/src/domain/provider.ts   ← imports JSON, derives ACP_CAPABILITIES
        │
        ▼
UI components (FindingsPage, ProvidersSection)

To refresh probe data:

bash

npx tsx scripts/query-acp-capabilities.ts

This writes to packages/shared/src/data/acp-capability-results.json. The ACP_CAPABILITIES constant in provider.ts imports this JSON and transforms it via deriveCapabilities(). A small CAPABILITY_OVERRIDES map supplies non-probe metadata (authMethod, systemPromptMethod) that cannot be determined automatically.

Protocol Overview

The Agent Client Protocol (ACP) is a standardized protocol for client-agent communication. Flightdeck uses ACP as its unified adapter layer — all 8 provider backends are accessed through the same ACP interface, enabling provider-agnostic agent orchestration.

ACP defines capability negotiation during session initialization: the client advertises what it supports (filesystem access, terminal, etc.) and the server responds with agent capabilities (image support, audio, MCP servers, session resume).

ACP SDK Type Definitions

The @agentclientprotocol/sdk package defines the following capability types exchanged during initialization:

`ClientCapabilities` — What the client advertises to agents

typescript

ClientCapabilities {
  fs?: {
    readTextFile?: boolean;   // Client can read files
    writeTextFile?: boolean;  // Client can write files
  };
  auth?: {
    terminal?: boolean;       // Client supports terminal-based auth flows
  };
  terminal?: boolean;         // Client has terminal access
  _meta?: Record<string, unknown>;
}

`AgentCapabilities` — What the agent reports back

typescript

AgentCapabilities {
  loadSession?: boolean;            // Agent supports session resume
  promptCapabilities?: PromptCapabilities;
  mcpCapabilities?: McpCapabilities;
  sessionCapabilities?: SessionCapabilities;
}

`PromptCapabilities` — Supported content types in prompts

typescript

PromptCapabilities {
  image?: boolean;           // Supports image attachments
  audio?: boolean;           // Supports audio input/output
  embeddedContext?: boolean;  // Supports inline file/resource embedding
}

`McpCapabilities` — MCP server connectivity

typescript

McpCapabilities {
  http?: boolean;  // Supports HTTP-based MCP servers
  sse?: boolean;   // Supports SSE-based MCP servers
}

`SessionCapabilities` — Session lifecycle operations

typescript

SessionCapabilities {
  resume?: {};  // Supports resuming previous sessions
  close?: {};   // Supports explicit session close
  fork?: {};    // Supports forking a session
  list?: {};    // Supports listing previous sessions
}

Opportunities: Capability Utilization

✅ Resolved

Live probe data for all 7 installed providers — no more assumptions or static guesses
ACP_CAPABILITIES derived from probe JSON — single source of truth, no hand-maintained constants that drift
Gemini resume preset clarified — the old supportsResume field was ambiguous. It has been renamed to supportsLoadSession (which IS true for Gemini — it advertises loadSession). Gemini does NOT have sessionCapabilities.resume (no session resume/fork), which is what the old preset incorrectly implied.

⚠️ Remaining Gaps

1. Empty client capabilities. Flightdeck sends clientCapabilities: {} during session initialization (AcpAdapter.ts:330). Agents don't know the client has filesystem and terminal access. Should be { fs: { readTextFile: true, writeTextFile: true }, terminal: true }.

2. Only supportsImages consumed at runtime. From the rich AgentCapabilities response, only promptCapabilities.image is read (AcpAdapter.ts:113):

typescript

get supportsImages(): boolean {
  return this.agentCapabilities?.promptCapabilities?.image ?? false;
}

3. Captured but unused fields: loadSession, audio, embeddedContext, mcpCapabilities, and sessionCapabilities are stored on the adapter instance but not surfaced in the UI or used in adapter logic.

Static Presets vs Runtime Capabilities

There are two distinct sources of capability information:

Static presets (PROVIDER_REGISTRY in packages/shared/src/domain/provider.ts) — Hardcoded metadata about each provider CLI: binary name, supported models, auth method, resume support. These are compile-time constants used for provider selection and model resolution.
Runtime capabilities (AgentCapabilities from ACP initialize response) — Actual capabilities reported by the running agent process. These may differ from static presets if a provider updates its capabilities. Currently only supportsImages is read at runtime.

The static presets drive the UI (model selectors, provider badges, setup wizard). The runtime capabilities should ideally augment or override static presets but currently do not.

Provider Details

🐙 GitHub Copilot

Field	Value
Binary	`copilot --acp --stdio`
Probe Version	v1.0.9
Auth	`gh auth status` (GitHub CLI login)
Resume	✅ (loadSession + session list, no fork/resume)
Images	✅
Audio	❌
MCP	❌ (no mcpCapabilities)
Embedded Context	✅
Model Selection	`--model <name>` flag
Models	Anthropic (sonnet, haiku, opus), OpenAI (gpt-4.1), Google (gemini-3-pro-preview), xAI
Model Tiers	fast: `claude-haiku-4.5`, standard: `claude-sonnet-4.6`, premium: `claude-opus-4.6`
System Prompt	First user message
Agent File	`.agent.md`
Unique	Multi-backend access, `--agent=<name>` flag, widest model selection

🟠 Claude (Anthropic)

Field	Value
Binary	`claude-agent-acp`
Probe Version	v0.21.0
Auth	`ANTHROPIC_API_KEY` env var
Resume	✅ (full: fork + list + resume — richest session support)
Images	✅
Audio	❌
MCP	✅ HTTP + SSE
Embedded Context	✅
Model Selection	`--model <alias>` (uses short aliases: `default`, `opus`, `haiku`)
Models	Anthropic only (sonnet, haiku, opus)
Model Tiers	fast: `haiku`, standard: `default` (sonnet), premium: `opus`
System Prompt	`_meta.systemPrompt` extension (unique to Claude)
Agent File	`CLAUDE.md`
Unique	Only provider with full session fork/resume, `_meta.systemPrompt`, `promptQueueing`, model aliases

💎 Google Gemini CLI

Field	Value
Binary	`gemini --acp`
Probe Version	v0.34.0
Auth	`GEMINI_API_KEY` env var (4 auth methods: OAuth, API key, Vertex AI, Gateway)
Resume	⚠️ (loadSession ✅, but no sessionCapabilities — no resume/fork)
Images	✅
Audio	✅ (only provider with audio support)
MCP	✅ HTTP + SSE
Embedded Context	✅
Model Selection	`--model <name>` flag
Models	Google only (gemini-3.1-pro, gemini-3.1-flash, gemini-3.1-flash-lite)
Model Tiers	fast: `gemini-3.1-flash-lite`, standard: `gemini-3.1-flash`, premium: `gemini-3.1-pro`
System Prompt	First user message
Agent File	`.gemini/agents/*.md`
Unique	Google-native models, audio support, directory-based agent files

🤖 Codex (OpenAI)

Field	Value
Binary	`codex-acp`
Probe Version	v0.9.5
Auth	`OPENAI_API_KEY` env var (3 auth methods: ChatGPT login, CODEX_API_KEY, OPENAI_API_KEY)
Resume	❌ (session list only, no resume/fork)
Images	✅
Audio	❌
MCP	⚠️ HTTP only (no SSE)
Embedded Context	✅
Model Selection	Config-style: `-c model=<name>`
Models	OpenAI only (gpt-5.x-codex, gpt-5.4)
Model Tiers	fast: `gpt-5.1-codex-mini`, standard: `gpt-5.3-codex`, premium: `gpt-5.4`
System Prompt	First user message
Unique	Config-based model selection; HTTP-only MCP

🌙 Kimi CLI (Moonshot AI)

Field	Value
Binary	`kimi acp`
Probe Version	v1.24.0
Auth	`kimi login` (Moonshot account)
Resume	✅ (session list + resume, loadSession)
Images	✅
Audio	❌
MCP	⚠️ HTTP only (no SSE)
Embedded Context	✅
Model Selection	`--model <name>` flag
Models	Moonshot native (kimi-latest, moonshot-v1-8k)
Model Tiers	fast: `moonshot-v1-8k`, standard: `kimi-latest`, premium: `kimi-latest`
System Prompt	First user message
Unique	Moonshot AI models, terminal-based login auth

🔮 Qwen Code (Alibaba)

Field	Value
Binary	`qwen --acp --experimental-skills`
Probe Version	v0.12.6
Auth	Qwen OAuth (free daily requests) or OPENAI_API_KEY
Resume	✅ (session list + resume, loadSession)
Images	✅
Audio	✅ (one of two providers with audio support)
MCP	❌
Embedded Context	✅
Model Selection	`--model <name>` flag
Models	Qwen (qwen-coder-plus-latest) + OpenAI backend
Model Tiers	fast/standard/premium: `qwen-coder-plus-latest` (single model currently)
System Prompt	First user message
Unique	Audio support, Qwen OAuth with free tier, dual-backend (Qwen + OpenAI)

↗️ Cursor (NOT PROBED — binary not installed)

Field	Value
Binary	`agent acp`
Auth	`CURSOR_API_KEY` env var
Resume	Unknown (static preset only)
Model Selection	Not configurable via CLI (`modelArgStrategy: 'none'`)
Models	Multi-backend: Anthropic, OpenAI, Google
Model Tiers	fast: `claude-haiku-4.5`, standard: `claude-sonnet-4.6`, premium: `claude-opus-4.6`
System Prompt	`.cursorrules` file
Unique	Multi-backend like Copilot; model selection managed by Cursor, not CLI args; `.cursorrules` for system prompt

🔓 OpenCode

Field	Value
Binary	`opencode acp`
Probe Version	v1.2.27
Auth	Self-managed (`opencode auth login` — handles own API keys)
Resume	✅ (full: fork + list + resume — matches Claude)
Images	✅
Audio	❌
MCP	✅ HTTP + SSE
Embedded Context	✅
Model Selection	Not configurable via CLI (`modelArgStrategy: 'none'`)
Models	Multi-backend: Anthropic, OpenAI, Google, local models
Model Tiers	fast: `anthropic/claude-haiku-4-5`, standard: `anthropic/claude-sonnet-4-6`, premium: `anthropic/claude-opus-4-6`
System Prompt	First user message
Model Name Format	Prefixed: `anthropic/claude-sonnet-4-6`, `openai/gpt-5.2`, `google/gemini-2.5-pro`
Unique	Full session management (like Claude), supports local models, provider-prefixed model names, self-managed auth

Capability Matrix Summary (Probe-Verified)

Provider	Version	Resume	Images	Audio	MCP	Embedded Ctx	Session Caps	Auth Method	Multi-Backend
Copilot	v1.0.9	✅	✅	❌	❌	✅	list	GitHub CLI	✅ (4 backends)
Claude	v0.21.0	✅	✅	❌	✅ http+sse	✅	fork+list+resume	API key	❌
Gemini	v0.34.0	⚠️ loadSession only	✅	✅	✅ http+sse	✅	none	API key (4 methods)	❌
Codex	v0.9.5	❌	✅	❌	⚠️ http only	✅	list	API key (3 methods)	❌
Kimi	v1.24.0	✅	✅	❌	⚠️ http only	✅	list+resume	Moonshot login	❌
Qwen Code	v0.12.6	✅	✅	✅	❌	✅	list+resume	Qwen OAuth / OPENAI_API_KEY	✅ (2 backends)
Cursor	—	✅*	—	—	—	—	—	API key	✅ (3 backends)
OpenCode	v1.2.27	✅	✅	❌	✅ http+sse	✅	fork+list+resume	opencode auth login	✅ (4 backends + local)

*Cursor not probed (binary not installed). Resume status from static preset only.

Key Findings

1. Empty Client Capabilities

Flightdeck sends clientCapabilities: {} — agents don't know the client has filesystem and terminal access. This should be { fs: { readTextFile: true, writeTextFile: true }, terminal: true }.

2. Only `supportsImages` Consumed

Of the rich AgentCapabilities response, only promptCapabilities.image is read. loadSession, audio, embeddedContext, mcpCapabilities, and sessionCapabilities are captured but unused.

3. Session Resume Support Varies Widely

Probe-verified session capabilities across all 7 providers:

Full session management (fork+list+resume): Claude, OpenCode
List + resume (no fork): Kimi, Qwen Code
List only (no resume/fork): Copilot, Codex
No session capabilities: Gemini (probe-verified — preset was incorrectly true, now fixed)

4. MCP Server Support Varies

Probe verified: Claude, Gemini, and OpenCode support both HTTP and SSE MCP transports. Codex and Kimi support HTTP only (no SSE). Copilot and Qwen Code have no MCP capabilities. No current Flightdeck configuration leverages MCP server passthrough.

5. Claude Uses `_meta.systemPrompt` Extension

Claude is the only provider receiving system prompts via _meta.systemPrompt in the session metadata (AcpAdapter.ts:149-150). All others receive it as the first user message.

6. Model Selection Strategies Vary

Three distinct strategies exist across providers:

Flag (--model <name>): Copilot, Claude, Gemini, Kimi, Qwen Code
Config (-c model=<name>): Codex
None (managed externally): Cursor, OpenCode

7. Audio Support Is Rare

Only Gemini and Qwen Code support audio content in prompts. All other providers are image-only for media.

8. OpenCode Matches Claude for Session Capabilities

OpenCode (v1.2.27) now supports full session management (fork+list+resume), matching Claude as the only providers with complete session lifecycle support. Both also support MCP HTTP+SSE.

Recommendations

Advertise client capabilities — Send { fs: { readTextFile: true, writeTextFile: true }, terminal: true } in clientCapabilities so providers can offer richer agent cooperation.
Surface capabilities in UI — Show which capabilities each agent's provider supports in the agent detail panel, giving operators visibility into what each agent can do.
Leverage MCP server passthrough — Allow project configuration to specify MCP servers that agents should connect to, enabling tool augmentation (database access, API integrations, custom tools).
Handle resume gaps gracefully — Both Gemini and Codex don't support session resume. The UI should indicate this limitation and offer context-replay as an alternative.
Consume runtime capabilities — Use the full AgentCapabilities response to override static presets, enabling the system to adapt to provider updates without code changes.
Unify model selection — Abstract the 3 model selection strategies behind a consistent interface so new providers can be added without adapter changes.

Technical Reference

File	Purpose
`packages/shared/src/data/acp-capability-results.json`	Raw probe output — single source of truth for capability data
`packages/shared/src/domain/provider.ts`	Central `PROVIDER_REGISTRY` + derived `ACP_CAPABILITIES` from probe JSON
`scripts/query-acp-capabilities.ts`	Probe script — runs all providers, writes JSON
`packages/server/src/adapters/AcpAdapter.ts`	Core ACP adapter — session init, capability negotiation, message routing
`packages/server/src/adapters/types.ts`	TypeScript types for `AgentAdapter`, `ToolCallInfo`, capability interfaces
`packages/server/src/adapters/AdapterFactory.ts`	Factory that creates provider-specific adapter instances
`packages/server/src/adapters/presets.ts`	Provider presets derived from shared registry
`packages/server/src/adapters/ModelResolver.ts`	Model name resolution, tier mapping, and provider-specific prefixing
`@agentclientprotocol/sdk`	ACP SDK — Zod schemas for all capability types

Agent Client Protocol — Provider Capability Matrix ​

Architecture: JSON-Derived Single Source of Truth ​

Protocol Overview ​

ACP SDK Type Definitions ​

ClientCapabilities — What the client advertises to agents ​

AgentCapabilities — What the agent reports back ​

PromptCapabilities — Supported content types in prompts ​

McpCapabilities — MCP server connectivity ​

SessionCapabilities — Session lifecycle operations ​

Opportunities: Capability Utilization ​

✅ Resolved ​

⚠️ Remaining Gaps ​

Static Presets vs Runtime Capabilities ​

Provider Details ​

🐙 GitHub Copilot ​

🟠 Claude (Anthropic) ​

💎 Google Gemini CLI ​

🤖 Codex (OpenAI) ​

🌙 Kimi CLI (Moonshot AI) ​

🔮 Qwen Code (Alibaba) ​

↗️ Cursor (NOT PROBED — binary not installed) ​

🔓 OpenCode ​

Capability Matrix Summary (Probe-Verified) ​

Key Findings ​

1. Empty Client Capabilities ​

2. Only supportsImages Consumed ​

3. Session Resume Support Varies Widely ​

4. MCP Server Support Varies ​

5. Claude Uses _meta.systemPrompt Extension ​

6. Model Selection Strategies Vary ​

7. Audio Support Is Rare ​

8. OpenCode Matches Claude for Session Capabilities ​

Recommendations ​

Technical Reference ​

Agent Client Protocol — Provider Capability Matrix

Architecture: JSON-Derived Single Source of Truth

Protocol Overview

ACP SDK Type Definitions

`ClientCapabilities` — What the client advertises to agents

`AgentCapabilities` — What the agent reports back

`PromptCapabilities` — Supported content types in prompts

`McpCapabilities` — MCP server connectivity

`SessionCapabilities` — Session lifecycle operations

Opportunities: Capability Utilization

✅ Resolved

⚠️ Remaining Gaps

Static Presets vs Runtime Capabilities

Provider Details

🐙 GitHub Copilot

🟠 Claude (Anthropic)

💎 Google Gemini CLI

🤖 Codex (OpenAI)

🌙 Kimi CLI (Moonshot AI)

🔮 Qwen Code (Alibaba)

↗️ Cursor (NOT PROBED — binary not installed)

🔓 OpenCode

Capability Matrix Summary (Probe-Verified)

Key Findings

1. Empty Client Capabilities

2. Only `supportsImages` Consumed

3. Session Resume Support Varies Widely

4. MCP Server Support Varies

5. Claude Uses `_meta.systemPrompt` Extension

6. Model Selection Strategies Vary

7. Audio Support Is Rare

8. OpenCode Matches Claude for Session Capabilities

Recommendations

Technical Reference