AgentAPI++ (k1dav fork)

HTTP API for programmatically controlling CLI-based AI coding agents (Claude Code, Codex, Goose, Aider, Gemini CLI, GitHub Copilot, Amp, Cursor CLI, Auggie, Opencode, Amazon Q) over an in-memory PTY.

HTTP Request → AgentAPI++ → Terminal Emulator → claude / codex / ...

Fork lineage: coder/agentapi → KooshaPari/agentapi-plusplus → this repo. Binaries are built by CI when a GitHub release is published and consumed by k1dav-c/coder-templates' ai-agents module.

Fork highlights

• Structured timeline — thinking / tool calls / tool results captured from the agent's own transcript files, exposed via GET /timeline and SSE timeline_event (see below)
• DELETE /messages resets the agent session — clears history + timeline and sends the agent's new-session command (claude: /clear, codex: /new)
• Runtime MCP config — GET /mcp / PUT /mcp manage the agent's MCP servers on the fly; ?restart=true restarts the agent process in place so changes apply immediately
• GET /links — copy-safe URL extraction: the terminal hard-wraps long URLs across lines; this returns them whole (transcript-sourced when available, heuristically rejoined otherwise)
• Chat UI process view — inline collapsible tool-call cards and a filterable timeline side panel
• API-key auth on mutating routes (--api-key / AGENTAPI_API_KEY)
• Extra read endpoints: /info, /health, /version, /ready, /messages/count

Quick Start

OS=$(uname -s | tr "[:upper:]""[:lower:]")
ARCH=$(uname -m | sed "s/x86_64/amd64/;s/aarch64/arm64/")
curl -fsSL "https://github.com/k1dav/agentapi-plusplus/releases/latest/download/agentapi-${OS}-${ARCH}" -o agentapi
chmod +x agentapi
# Start with Claude Code (specify --type explicitly, otherwise message formatting may break)
./agentapi server --type claude -- claude

The server runs on port 3284. The chat UI is at http://localhost:3284/chat, the OpenAPI schema at http://localhost:3284/openapi.json (also checked in as openapi.json).

Build from source: go build -o agentapi main.go (chat UI assets are embedded separately — see Development).

Endpoints

Method / Path	Description
GET /messages	Conversation history
POST /message	Send a message (user or raw keystrokes); auth-gated
DELETE /messages	Clear history and timeline, and reset the agent's session (?new_session=false to skip); auth-gated
GET /timeline	Structured process events; ?kind= and ?since_id= filters
GET /links	Unique URLs found in the conversation — exact even when the terminal line-wrapped them
GET /mcp	Currently configured MCP servers and the config file path
PUT /mcp	Replace the MCP server set; ?restart=true restarts the agent to apply immediately; auth-gated
GET /status	Agent status: stable or running
GET /events	SSE stream: message_update, status_change, timeline_event, agent_error
POST /upload	Upload files; auth-gated
GET /info	Version, agent type, feature flags (features.timeline)
GET /health, /version, /ready, /messages/count	Auxiliary read endpoints

Allowed hosts

By default only requests with a localhost host header are accepted. Override with --allowed-hosts / AGENTAPI_ALLOWED_HOSTS (hostnames only, no ports; * allows all).

Structured timeline (thinking & tool calls)

The PTY screen only shows the agent's final answer text. For the process — thinking, tool calls, tool results — AgentAPI++ tails the transcript files the agent itself writes to disk (read-only sidecar; the TUI is untouched):

• Claude Code: ~/.claude/projects/<sanitized-cwd>/<session>.jsonl
• Codex: ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl

Events are normalized to a single shape and assigned monotonic ids:

{
"id": 2,
"kind": "tool_call", // thinking | text | tool_call | tool_result | system"role": "assistant",
"time": "2026-07-08T09:31:59Z",
"session_id": "d1341cee-...",
"tool_name": "Bash",
"tool_input": {"command": "ls"},
"tool_use_id": "toolu_01..."// joins a tool_call with its tool_result
}

curl 'localhost:3284/timeline'# full process
curl 'localhost:3284/timeline?kind=tool_call'# only tool calls
curl 'localhost:3284/timeline?since_id=42'# incremental polling
curl -N localhost:3284/events # live: timeline_event frames

Behavior notes:

• Supported for claude and codex on the PTY transport; on by default, disable with --timeline=false, override the search directory with --timeline-dir
• Session switches (claude /clear, codex new thread) are followed automatically and marked with a system "session switched" event; files from previous runs are never re-ingested (mtime filter)
• The last 10,000 events are kept in memory; late SSE subscribers get the most recent 500 replayed (full history via GET /timeline); ids keep increasing across DELETE /messages, so since_id polling never misses events
• Thinking events appear only when the backend writes plaintext reasoning to the transcript. Some deployments encrypt/redact it at the source (Claude signature-only thinking blocks, Codex encrypted_content); tool calls and results are always plaintext and unaffected

Runtime MCP configuration

Agents load MCP config at process startup. PUT /mcp writes the agent's config file — claude: project .mcp.json in the working directory, codex: the [mcp_servers.*] tables in ~/.codex/config.toml (other content, including comments, is preserved verbatim) — and with ?restart=true restarts the agent process in place so the change takes effect immediately. AgentAPI keeps serving across the restart; the agent's conversation context is reset (a new session starts).

curl 'localhost:3284/mcp'# current servers + config path# Full replace: servers not listed are removed. Restart to apply now.
curl -X PUT 'localhost:3284/mcp?restart=true' -H 'Content-Type: application/json' -d '{ "servers": { "memory": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"]}, "remote": {"type": "http", "url": "https://mcp.example.com", "headers": {"X-Key": "..."}} }}'
curl -X PUT localhost:3284/mcp -H 'Content-Type: application/json' \
-d '{"servers":{}}'# clear all MCP servers (applies next session)

Server config objects are passed through to the agent verbatim — use whatever fields the agent supports. Supported for claude and codex on the PTY transport (features.mcp in GET /info); MCP tool invocations show up in the timeline like any other tool call.

Supported agents

claude, codex, goose, aider, gemini, copilot, amp, cursor, auggie, amazonq, opencode, custom — pass via --type. Message formatting and readiness detection are per-agent (lib/msgfmt); the structured timeline currently covers claude and codex.

API examples

# Send a message
curl -X POST http://localhost:3284/message \
-H "Content-Type: application/json" \
-d '{"type":"user","content":"list the files in this directory"}'# Watch the process live
curl -N http://localhost:3284/events
# Inspect what tools the agent used
curl -s 'http://localhost:3284/timeline?kind=tool_call'| jq '.events[] | {tool_name, tool_input}'# Start over (clears history + timeline, agent gets /clear or /new)
curl -X DELETE http://localhost:3284/messages

Chat UI

Next.js app served at /chat:

• Tool invocations render as collapsible cards in the conversation flow (input summary, running/done state, full input JSON and result on expand)
• A timeline side panel (top-right toggle) lists every process event with kind filters (Tools / Thinking / Text / System)
• Dark mode, mobile drawer layout

Architecture

Component	Description
cmd/	CLI commands (server, attach)
lib/httpapi/	HTTP server, routes, SSE event emitter
lib/termexec/	PTY process execution
lib/screentracker/	Screen snapshot → conversation messages
lib/msgfmt/	Agent-specific message formatting
lib/transcript/	Transcript tailing → structured timeline (discovery, tailer, per-agent parsers, watcher)
chat/	Next.js web UI

Configuration

export AGENTAPI_PORT=3284
export AGENTAPI_ALLOWED_HOSTS="localhost,127.0.0.1"export AGENTAPI_ALLOWED_ORIGINS="http://localhost:3284"export AGENTAPI_API_KEY="..."# enables bearer auth on mutating routesexport AGENTAPI_TIMELINE=true # structured timeline capture (default true)

Every --flag has an AGENTAPI_<FLAG> env equivalent (dashes → underscores). See agentapi server --help for the full list.

Development

go build -o agentapi main.go # server only; chat UI 404s with a hint
go test ./...
# Full build with embedded chat UI (what release CI does)cd chat && bun install --frozen-lockfile && \
NEXT_PUBLIC_BASE_PATH="/magic-base-path-placeholder" bun run build &&cd .. && \
rm -rf lib/httpapi/chat && mkdir -p lib/httpapi/chat && touch lib/httpapi/chat/marker && \
cp -r chat/out/. lib/httpapi/chat/ && go build -o agentapi main.go

Releases: publish a GitHub release; .github/workflows/release.yml builds the chat UI, embeds it, and uploads binaries for linux/darwin × amd64/arm64 as release assets.

If you change the HTTP API, regenerate the schema snapshot: go run main.go server --print-openapi dummy > openapi.json.

License

MIT — see LICENSE.