Opencode Api

OpenCode REST API

Use this skill when you need reliable, scriptable control of OpenCode over HTTP.

Default OpenCode server base URL: http://localhost:4096.

Canonical docs: https://opencode.ai/docs/server/

Quick Start

Start a local, protected server:

export OPENCODE_SERVER_USERNAME=opencode
export OPENCODE_SERVER_PASSWORD="change-me"
opencode serve --hostname 127.0.0.1 --port 4096

Basic API helper:

api() {
  curl -sS --fail \
    -u "$OPENCODE_SERVER_USERNAME:$OPENCODE_SERVER_PASSWORD" \
    -H 'Content-Type: application/json' "$@"
}

Health check:

api "http://127.0.0.1:4096/global/health"

Core Flow

1. Check server health with GET /global/health.
2. List sessions with GET /session.
3. Create a session with POST /session if needed.
4. Send work with POST /session/{id}/message or POST /session/{id}/prompt_async.
5. Track progress via GET /session/status and SSE GET /event.
6. Resolve permission requests through POST /session/{id}/permissions/{permissionID}.

Minimal Request Examples

Create a session:

api -X POST "http://127.0.0.1:4096/session" \
  -d '{"title":"billing-platform-service#761 follow-up"}'

Send a normal prompt (waits for response):

api -X POST "http://127.0.0.1:4096/session/SESSION_ID/message" \
  -d '{
    "parts": [
      {"type":"text","text":"Address unresolved PR feedback and summarize changes."}
    ]
  }'

Send async prompt (fire-and-track):

api -X POST "http://127.0.0.1:4096/session/SESSION_ID/prompt_async" \
  -d '{
    "parts": [
      {"type":"text","text":"Apply requested review updates and report status."}
    ]
  }'

Reply to permission request:

api -X POST "http://127.0.0.1:4096/session/SESSION_ID/permissions/PERMISSION_ID" \
  -d '{"response":"allow","remember":true}'

Stream events (SSE):

curl -N -u "$OPENCODE_SERVER_USERNAME:$OPENCODE_SERVER_PASSWORD" \
  "http://127.0.0.1:4096/event"

API Surface You Will Use Most

• GET /doc - OpenAPI spec source of truth.
• GET /session - enumerate sessions.
• GET /session/status - status for all sessions in one call.
• GET /session/{id}/message - read conversation state.
• GET /session/{id}/todo - inspect tracked tasks.
• POST /session/{id}/message - synchronous prompt.
• POST /session/{id}/prompt_async - asynchronous prompt.
• POST /session/{id}/abort - cancel stuck work.
• GET /session/{id}/diff - inspect produced code changes.
• GET /find, GET /find/file, GET /file/content - pre-read context before prompting.

Performance and Reliability Optimizations

1. Keep one long-lived server process per workspace to avoid repeated startup and warm-up.
2. Prefer SSE (/event) over tight polling loops for status changes.
3. Use /prompt_async for fan-out workflows; use /session/status for cheap centralized progress checks.
4. Pull context with /find and /file/content first; then send narrower prompts to reduce token usage.
5. Use stable session titles like # to make future lookups deterministic.
6. Handle permission.updated events immediately to avoid idle stalls.
7. Use query.directory when driving multiple projects from one controller.
8. Reuse a keep-alive HTTP client instead of spawning short-lived curl in high-volume automation.

Security Guardrails

• Default bind should stay 127.0.0.1.
• Always set basic auth if the server is reachable by other processes or hosts.
• Keep CORS explicit and minimal (--cors only for trusted origins).
• If exposed remotely, put TLS and access control in front of OpenCode.

Error Handling

• 400 usually means payload shape mismatch; verify against /doc.
• 404 usually means wrong session/message/permission ID.
• If a session appears stuck, call /session/{id}/abort, then post a refined follow-up prompt.

Email-Driven Follow-Up Pattern

For PR-review automation:

1. Ingest actionable review emails from Gmail triage.
2. Extract repo, pr_number, and requested change.
3. Find or create session titled # review-followup.
4. Post follow-up via /prompt_async.
5. Watch /event for completion, permission, and error events.
6. Pull /session/{id}/diff and /session/{id}/todo for summary output.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dzianisv
• Source: dzianisv/skills
• License: MIT

Install and usage instructions live in the source repository linked above.