Supported Coding Agents Matrix
Twelve evidence-measured adapters backed by automated Linux sandbox probe captures.
The Evidence Evaluation Standard
Most multi-agent tooling relies on vendor marketing sheets or outdated README entries. Agent Profile rejects assumptions. Our adapters are measured against an automated Linux sandbox harness that inspects filesystem deltas, process strace events, and environment inheritance across unauthenticated launches.
The adapter mechanism has been empirically observed relocating files inside our test sandbox. The agent launches successfully, respects the isolation variable, and exhibits predictable behavior.
The adapter relocates primary configuration, but certain credential or state subsystems require network authentication or vendor login steps not yet fully mapped in offline sandbox tests.
The 12-Agent Capability Matrix
Status as verified against agent-profile probe captures:
| Agent | Support Tier | Isolation Mechanism | Config | Credentials | State |
|---|---|---|---|---|---|
| Claude Code | Proven | CLAUDE_CONFIG_DIR (env) |
Supported | Conditional | Not Guaranteed |
| OpenAI Codex | Proven | CODEX_HOME (env) |
Supported | Conditional | Conditional |
| Aider | Proven | --config <file> (arg) |
Not Guaranteed | Not Supported | Not Supported |
| Gemini CLI | Proven | GEMINI_CLI_HOME (env) |
Supported | Unknown | Supported |
| OpenCode | Proven | OPENCODE_CONFIG_DIR (env) |
Supported | Unknown | Not Supported |
| Cursor CLI | Experimental | CURSOR_CONFIG_DIR (env) |
Supported | Unknown | Unknown |
| Copilot CLI | Experimental | COPILOT_HOME (env) |
Unknown | Unknown | Unknown |
| Amp (Ampere) | Proven | --settings-file <file> (arg) |
Not Guaranteed | Unknown | Not Guaranteed |
| Cline CLI | Experimental | CLINE_DATA_DIR (env) |
Unknown | Unknown | Unknown |
| Continue | Experimental | --config <file> (arg) |
Unknown | Unknown | Unknown |
| Windsurf | Experimental | WINDSURF_CONFIG_DIR (env) |
Unknown | Unknown | Unknown |
| Kiro CLI | Experimental | KIRO_HOME (env) |
Supported | Unknown | Not Supported |
Deep Dive: Detailed Evidence Citations
Claude Code (Anthropic)
Mechanism: Environment variable CLAUDE_CONFIG_DIR
Config Isolation: Supported. User settings and tool configuration files live in the relocated config directory. Note that project-local .claude/settings*.json and .mcp.json layer on top within the repository.
Credential Isolation: Conditional. Ambient environment variables (e.g. ANTHROPIC_API_KEY) override disk settings. The per-directory .credentials.json and macOS Keychain entries follow documented paths.
State Isolation: Not guaranteed. Community reports show history files moving with the directory, but vendor documentation does not formally guarantee complete database relocation.
OpenAI Codex CLI
Mechanism: Environment variable CODEX_HOME
Config Isolation: Supported. config.toml and named configuration profiles live inside CODEX_HOME.
Credential Isolation: Conditional. OPENAI_API_KEY, CODEX_API_KEY, and CODEX_ACCESS_TOKEN bypass disk storage when present in the ambient environment.
State Isolation: Conditional. CODEX_SQLITE_HOME relocates the state database when explicitly configured.
Gemini CLI (Google)
Mechanism: Environment variable GEMINI_CLI_HOME
Config Isolation: Supported. The probe harness observed filesystem creation at + d /target/.gemini and + f /target/.gemini/projects.json.
State Isolation: Supported. History and temporary execution files relocated cleanly into + d /target/.gemini/history and + d /target/.gemini/tmp.
Aider
Mechanism: CLI argument injection --config <file>
Config Isolation: Not guaranteed. Aider's configuration cascade merges the user-level ~/.aider.conf.yml, repository-level files, and .env files.
State Isolation: Not supported. Conversation history files (.aider.chat.history.md) are written directly into the working directory.
Auditing Local Adapters on Your Machine
You can inspect the exact capability matrix and verified executable paths on your own system at any time:
# Display capability matrix and PATH resolution for all 12 agents:
agent-profile agents
# Machine-readable JSON output for automated validation scripts:
agent-profile agents --json
# Run health diagnostics on a specific agent adapter:
agent-profile claude doctor
agent-profile claude doctor work --json