openhuman/gitbooks/developing/providers/claude-code.md

Claude Code CLI provider

OpenHuman can route any chat workload through Anthropic's claude CLI instead of calling the Anthropic HTTP API directly. The CLI handles model selection, auth, and prompt-cache management; OpenHuman drives it as a child process per turn, parses its stream-json output, and re-exposes its own read-only tools back into the CLI over MCP so the model can reach native OpenHuman state (memory, threads, channels, people).

Locked decisions live in .planning/claude-code-provider/PLAN.md §13.

Requirements

• Claude Code CLI ≥ 2.0.0 on PATH (or OPENHUMAN_CLAUDE_CLI=/abs/path/to/claude).
• An Anthropic API key in ANTHROPIC_API_KEY, or a pre-existing ~/.claude/.credentials.json from claude login.
• The openhuman-core binary on disk — OpenHuman spawns openhuman-core mcp as a stdio MCP server so the CLI can call OpenHuman tools. The path is discovered via std::env::current_exe().

Routing a workload through the CLI

The factory grammar accepts a new prefix: claude-code:<model>[@<temperature>]. Apply it via the standard inference settings (per-role, locked decision #3):

# Through the JSON-RPC update endpoint:
openhuman-core rpc openhuman.inference_update_model_settings \
  --json '{"chat_provider":"claude-code:claude-sonnet-4-5"}'

Role string	Field updated
chat_provider	foreground chat replies
reasoning_provider	long-context reasoning workloads
agentic_provider	multi-step agentic loops

A workload set to claude-code:<model> always spawns a fresh claude child per turn; concurrency is capped at MAX_CONCURRENT_TURNS = 4 per ClaudeCodeProvider instance.

Verifying the install

The status RPC is on the existing inference namespace:

openhuman-core rpc openhuman.inference_claude_code_status

Returns one of (CliStatus in src/openhuman/inference/provider/claude_code/types.rs):

• {"status":"ok","version":"2.0.4","path":"/usr/local/bin/claude"} — ready
• {"status":"not_installed"} — claude not on PATH
• {"status":"outdated","version":"1.9.0","min_required":"2.0.0","path":"…"} — bump CLI
• {"status":"unusable","path":"…","reason":"…"} — binary present but the version probe failed

The same status is rendered in the settings panel via ClaudeCodeStatusCard (app/src/components/settings/panels/ai/ClaudeCodeStatusCard.tsx).

Per-turn behavior

Each chat turn:

1. Resolve a per-thread CC session UUID from <workspace>/claude-code-sessions.json. New threads get a fresh RFC-4122 v4 UUID; the CLI requires v4 specifically for --resume.
2. Write mcp-config.json to a tempdir pointing at openhuman-core mcp (stdio MCP server, no extra credentials).
3. Spawn the CLI with:
   • -p --input-format stream-json --output-format stream-json --verbose --include-partial-messages
   • --mcp-config <tmp> --strict-mcp-config so only the configured MCP servers are visible
   • --disallowedTools Bash,Read,Write,Edit,Glob,Grep,WebFetch,WebSearch,TodoWrite,Task,BashOutput,KillShell — CC's own builtins stay off so OpenHuman tools (mcp__openhuman__*) are authoritative
   • --session-id <uuid> on first turn, --resume <uuid> thereafter
   • --model <model> (the suffix after claude-code:)
   • --append-system-prompt <…> if the conversation carries a system message
4. Pipe stdin: full conversation history on a new session, just the last user turn on --resume (the CLI already holds its own prior-turn context server-side).
5. Stream stdout through the JSONL parser → event mapper → ProviderDeltas on the request's stream sink.

On exit non-zero the driver bubbles stderr (capped at 16 KiB) up as the error message.

Auth resolution order

1. ANTHROPIC_API_KEY env var (highest precedence — set on the spawned child).
2. Per-thread / per-agent key from ChatRequest config (future, not yet wired).
3. ~/.claude/.credentials.json — the CLI's own OAuth tokens from claude login (Pro / Max subscription). We never read or round-trip the access token; auth detection probes this file for non-secret metadata only.
4. None — the CLI will fail with an auth error.

The openhuman.inference_claude_code_auth_status RPC probes sources 1 and 3 without spawning the CLI and surfaces the result in the Settings → AI panel.

Tool surface exposed to the CLI

The CLI sees these tools as mcp__openhuman__<name> (delivered by the existing stdio MCP server in src/openhuman/mcp_server/):

• core.list_tools, core.tool_instructions
• memory.search, memory.recall
• tree.read_chunk, tree.browse, tree.top_entities, tree.list_sources
• agent.list_subagents, agent.run_subagent (write — flagged destructiveHint per MCP spec)
• searxng_search

The MCP server enforces SecurityPolicy::ToolOperation checks; all tools except agent.run_subagent are read-only.

Limitations (v1)

• Vision input is not forwarded — set the vision_provider to a different provider when you need images.
• agentic runs share the same Semaphore(4); under load a CC turn waits in queue rather than failing fast.
• Cost accounting from the CLI's result.total_cost_usd is captured in the mapper but not yet wired into OpenHuman's billing layer (src/openhuman/cost/).