← Back to home

Configuration

AgenticKode is configured through environment variables, the web UI, and flow prompts.

Environment Variables

Core

VariableDefaultDescription
DATABASE_URL(required)PostgreSQL connection string
SECRET_KEY(required)Secret for session signing and AES encryption of stored secrets
CORS_ORIGINS*Comma-separated allowed CORS origins
LOG_LEVELINFOLogging level (DEBUG, INFO, WARNING, ERROR)

AI Providers

VariableDescription
ANTHROPIC_API_KEYAPI key for the Claude CLI agent
OLLAMA_BASE_URLOllama server URL for local models (e.g., http://ollama:11434)

Git Providers

VariableDescription
GITHUB_TOKENGitHub PAT with repo scope
GITLAB_TOKENGitLab PAT with api scope
GITLAB_BASE_URLGitLab instance URL (default: https://gitlab.com)
GITEA_TOKENGitea access token
GITEA_BASE_URLGitea instance URL
BITBUCKET_USERNAMEBitbucket username
BITBUCKET_APP_PASSWORDBitbucket app password

Notifications

VariableDescription
SLACK_WEBHOOK_URLSlack incoming webhook
DISCORD_WEBHOOK_URLDiscord webhook
TELEGRAM_BOT_TOKENTelegram bot token
TELEGRAM_CHAT_IDTelegram chat ID

Agent Settings

Manage agents from the Agents page in the UI. The UI surfaces three CLI agents: claude, codex, and opencode. Each agent has:

  • Environment variables — Provider-specific env vars injected during execution
  • CLI flags — Additional command-line flags passed to the agent
  • Availability — Enable / disable without removing configuration

Claude CLI

The worker auto-installs the Claude CLI binary on workspace servers.

  • ANTHROPIC_API_KEY — Set per-agent or globally
  • CLI flags — e.g., --model claude-sonnet-4-20250514, --max-tokens 4096

Codex

OpenAI's Codex CLI. Authentication via the agent's own login flow on each workspace server (the Agents panel on a server card exposes a one-click login terminal).

OpenCode

Self-hostable open-source coding agent. Same per-server install + login flow as Codex.

Ollama (optional)

Requires a running Ollama server for local models. Manage models from the GPU Dashboard.

  • OLLAMA_BASE_URL — Server URL

Flow Prompts

A run is a single flow prompt — a named prompt plus the data AgenticKode fetches for it. The pipeline runs workspace_setup → init, fetches the flow's data, calls the agent once, then finalization. The agent's response is the run outcome. See Autonomous Mode for the full flow.

Two flow prompts ship by default:

  • implement — Writes code for a task and opens a PR. Fetches repo + issue context. The default for new runs.
  • pr_review — Reviews a PR diff and posts a comment. Fetches the PR diff; runs in generate mode (no checkout).

A flow prompt has:

  • Prompt — the instruction sent to the agent, with {{run.*}} placeholders.
  • Flow typeimplement or pr_review (drives the fixed data fetch + agent mode).
  • Agent — a specific agent (claude, codex, …) or the project/global default.
  • Triggers — optional label / issue / PR / schedule rules that route incoming events to this flow prompt (TriggerMatcher).

Earlier releases used composable bash + agent step workflow templates; those were removed in 0.6.0 (ADR-009).

Project Instructions and Secrets

Per-project configuration is managed from the project's Instructions tab.

Instructions

  • Global — Injected into all AI prompts for this project
  • Phase-specific — Override for individual phases

Use instructions to provide project context, coding standards, test requirements, and review criteria.

Secrets

Encrypted secrets stored per-project and auto-injected during execution:

  • API keys specific to the project
  • Database credentials for testing
  • Service tokens

Secrets are encrypted with AES using the SECRET_KEY and never exposed in logs or UI.

Notification Channels

Configure from Settings > Notifications:

  • Slack — Post to channels via incoming webhook
  • Discord — Post to channels via webhook
  • Telegram — Send messages via bot
  • Webhook — Generic HTTP POST callback to any URL

Notifications fire on run completion, failure, and approval requests.

Per-project and per-step workspace overrides (v0.5.1)

Each ProjectConfig has three optional fields that override the platform defaults:

  • local_path — Absolute path to an existing local git checkout. When set, workflows skip clone/fetch and operate on this folder directly. The working tree must be clean (uncommitted changes block the run); worktree isolation is auto-forced so concurrent runs don't collide.
  • worker_user_override — OS user the agent runs as for this project (overrides the workspace server's worker_user). Any step can override this further with params.run_as.

Plus a platform-wide setting in Settings → Workspace:

  • workspace.default_root — Default workspace_root applied to new workspace servers when none is supplied at create time.

Backup and Export

From Settings > Backup:

  • Export — Download full configuration as JSON (projects, agents, templates)
  • Import — Restore configuration from a backup file
  • Encryption — Optionally encrypt backups with AES

Backups include all settings but exclude run logs and task history.

Next Steps