agencast

AgenCast runtime

AgenCast is a workflow runtime for AI agents: create a reusable procedure, validate and test its flow, run it with new inputs and inspect the results. Agents operate it through MCP or the CLI; the optional GUI helps people review and edit the same workflow files and inspect run records. Live workflow model calls use OpenRouter.

Version 0.20.0, Python 3.12 + uv. Formats per docs/spec/ (v1), design in docs/DESIGN.md. The package and the command are both called agencast (until 0.2.5 maw); the command name is in pyproject.toml ([project.scripts]).

The package also contains the skills, documentation and examples: guide.

Usage

Run the commands from the root of the repository clone.

uv run --project framework agencast --project examples/showcase validate ig-post
uv run --project framework agencast --project examples/showcase run ig-post -i topic="new coffee" --dry-run
uv run --project framework agencast --project examples/showcase run ig-post -i topic="new coffee" --fake examples/showcase/fake/ig-post.yaml
uv run --project framework agencast --project examples/showcase run ig-post -i topic="new coffee"
uv run --project framework agencast --project examples/showcase run demo-task -i books="Austen: Pride and Prejudice (1813)" --fake examples/showcase/fake/demo-task.yaml
uv run --project framework agencast --project examples/showcase runs list
uv run --project framework agencast --project examples/showcase runs show <run_id>
uv run --project framework agencast --project examples/showcase serve --host 127.0.0.1 --port 8080 [--workers 2] [--cors http://localhost:5173]
uv run --project framework agencast --project examples/showcase mcp [--allow read|run|edit] [--input-dir DIR]... [--fake [SCRIPT]] [--http [--host H] [--port P] [--allow-host H]...]
uv run --project framework agencast --project examples/showcase migrate examples/showcase/workflows/scenarios/ig-post.yaml
uv run --project framework agencast new project ~/my-project [--example showcase|tutorial]
uv run --project framework agencast new agent reviewer | new scenario check [--project <path>]
uv run --project framework agencast skills list | path | install [--to all] [--prefix DIR] [--copy] [--force]
uv run --project framework agencast docs [show spec/scenario.md]
uv run --project framework agencast projects list | add <path> [--name N] | rm <name> | trust <name> [--yes]

• A scenario can be given by name (ig-post) or by the path to its .yaml. The project root = the first folder with workflows/ going up from the current folder, or --project <path> on any command — everything works from any folder.

• Configuration workflows/config.yaml (template config.example.yaml) and the MCP server registry workflows/mcp.yaml (template mcp.example.yaml; only the owner changes both — mcp.yaml only on disk: the API never reads or writes the file, the GUI shows its servers read-only). Keys only from the environment or from .env in the project root.

• A task step starts the MCP servers from mcp.yaml once per run (stdio via npx needs Node; on a serverless host, preinstall the package). The servers' stderr is in the run record in mcp/<server>.stderr.log.

• --fake replaces model calls with no model cost and no OpenRouter key; an optional YAML contains scripted answers (described in src/agencast/fake.py). The MCP servers from mcp.yaml in a task step are real even with --fake; the example filesystem via npx needs Node.js and downloads a package on first run. The callback is really sent and needs a signing secret. Only scenarios without task (including called scenarios) and without --callback-url are guaranteed to run offline, e.g. ig-post.

• Inputs of type file and files are images (PNG, JPEG, WebP, GIF, AVIF; at most 10 MB each, up to 16): -i photo=a.jpg, -i 'refs=["a.png","b.webp"]'. They are copied into runs/<id>/inputs/ and carry width, height and format; images: on ask/task shows them to the model, on image they are reference images (only an alias with api: images). Over HTTP an image is uploaded first and passed as {"upload_id": …} (docs/spec/scenario.md, Type file; docs/spec/api.md, Uploads).

• --callback-url https://… sends the result signed with HMAC after the run (callback.secret_env).

• Run records are in runs/ (in .gitignore), files from output in outputs/ (storage.type: local).

• Every run has a report.html (a single file, CSS inside, no external resources, prompts and responses in <details>); a copy goes to storage and its URL is in the callback as report_url (storage.type: local → file://).

• new project <path> creates workflows/ (config, agent writer, scenario demo — they pass validate --offline and --fake), .env.example and .gitignore; new agent|scenario <name> adds a minimal file to the project. It never overwrites anything (docs/spec/projects.md).

• The project registry ~/.config/agencast/projects.yaml (AGENCAST_CONFIG_DIR) is filled by new project, projects add and a run that passed validation (not --dry-run; since 0.15.1 validate does not change the registry); the GUI can write in registry mode — its entries are trusted: false: such a project uses no MCP server until agencast projects trust <name>, which shows the project root and its servers and asks first (--yes outside a terminal; docs/spec/projects.md). projects_root sets the default folder for new projects (default ~/agencast-projects).

• migrate: there is nothing to convert in v1; an unknown version = config error.

• Exit code: 0 success, 1 the run ended with an error, 2 config error (validate, inputs, environment).

Webhook server (agencast serve)

Contract: docs/spec/webhook.md. Stdlib ThreadingHTTPServer, no web framework.

POST /runs            Authorization: Bearer $WEBHOOK_TOKEN
{"scenario": "ig-post", "inputs": {"topic": "…"}, "callback_url": "https://…", "request_key": "caller-4711"}
→ 202 {"run_id": "…", "queue_position": 1}   the run is queued, the result arrives at callback_url
→ 200 {"run_id": "<original>", "queue_position": null}   request_key was already used
→ 401 / 422 {"error": "…", "details": [...]}   nothing is created, no callback arrives
GET /runs/<run_id>    status: queued (+ queue_position) / running / callback body + callback_failed
GET /projects, /projects/<p>[/scenarios/<s>|/runs[/<id>[/files/<path>]]|/spend?day=]   read API (docs/spec/api.md)
POST /projects/<p>/runs   like POST /runs in project <p>; callback_url optional, "dry_run": true → plan only
POST /projects/<p>/uploads   raw bytes of one image → 201 {"upload_id"} for a file/files input (single project: POST /uploads)
POST /projects/<p>/validate   validation without writing ({path, text} or an empty body), errors as objects
POST /projects/new       creates a project from the template and writes it to the registry
POST /projects           registers an existing project with workflows/config.yaml
DELETE /projects/<p>     removes the project from the registry only, files stay
GET /, /assets/…      GUI (framework/src/agencast/ui/, no token)

• GUI (since 0.6.0): serve serves the built GUI from framework/src/agencast/ui/ (output of npm run build in the repository's ui/ folder; git ignores it, it goes into the wheel via artifacts). GET / and /assets/… need no token, the token protects only /projects… and /runs…; a path without an extension returns index.html (hash routing). Without a built GUI, GET / returns 404 with instructions. When developing the GUI from vite dev (another origin), run serve --port 8787 --cors http://localhost:5173 — without the switch there are no CORS headers.

• In a project or with --project: one project (token webhook.token_env); outside a project, registry mode: all projects from the registry, the server token AGENCAST_TOKEN, secrets from the server's environment and .env in cwd, /runs only via /projects/<p>/runs.

• --host and --port override AGENCAST_HOST and AGENCAST_PORT; the defaults are 127.0.0.1 and 8080. AGENCAST_PORT must be an integer 1–65535. These variables are read from the process environment while parsing arguments; .env in cwd is loaded only afterwards, so you cannot set the bind address or port from it.

• For registry mode, ~/.config/agencast/serve.env can contain AGENCAST_TOKEN, AGENCAST_HOST and AGENCAST_PORT:

  AGENCAST_TOKEN=<secret-token>
  AGENCAST_HOST=127.0.0.1
  AGENCAST_PORT=8080
  

  For access from other devices, set AGENCAST_HOST to the private address of the Tailscale interface. A user systemd service:

  [Unit]
  Description=AgenCast server
  [Service]
  ExecStart=%h/.local/bin/agencast serve
  EnvironmentFile=%h/.config/agencast/serve.env
  Restart=on-failure
  [Install]
  WantedBy=default.target
  

• Startup needs the variables webhook.token_env, callback.secret_env and (without --fake) the OpenRouter key; otherwise it ends with a config error.

• Runs go one after another (a single worker thread); --workers N starts N runs at once, and the completion order is then not guaranteed. The queue and request_key are files in <runs>/_queue/ — after a server restart, waiting requests are processed; a run interrupted midway is not repeated, a callback internal is sent (verify manually). On SIGTERM or Ctrl-C serve first stops the MCP servers of the runs in flight; requests still waiting stay in the queue.

• The optional limits.max_parallel_runs and limits.daily_budget_usd in config.yaml (since 0.3.1) apply to all runs over one runs/ — serve, the manual CLI and cron share one cap (docs/spec/config.md).

• Callback: https:// (exception http://127.0.0.1 for tests), signature X-Signature: sha256=<HMAC>, 3 attempts, then callback_failed.

• The server is HTTP without TLS and the GUI has no token protection. Never expose it publicly; bind only to localhost or a private network, for example Tailscale. Hosted deployment is Phase 3c.

MCP server (agencast mcp, since 0.19.0)

agencast mcp serves the registered projects (one with --project) to an MCP client over stdio, or over streamable HTTP with --http and a bearer token from AGENCAST_MCP_TOKEN: tools to list projects and scenarios, run them (fake_run free, run_scenario live), follow and read runs, and with --allow edit write scenarios, agents and skills. Runs outlive the client and the server. Image inputs come from directories the owner allows with --input-dir. Clients that render MCP Apps also get a run card. Everything is specified in docs/spec/mcp-server.md; client setup is in the main README.

Tests

cd framework && uv run pytest

Conformance suite (DESIGN §5.6, §5.9): expressions (58 cases from spike (c) adjusted per the spec + the spec rules), loader, validate, engine (error classes, retry, cascade, parallel, switch, budget, timeout, callback, masking), the task step with the fake MCP server tests/fake_mcp_server.py (permissions, schema normalization, loop, skills, dedupe_key, leftover processes) and golden scenarios — every file in examples/*/workflows/ and every example in docs/spec/. A new scenario in examples/*/workflows/scenarios/ is tested automatically; its scripted answers belong in ../examples/<project>/fake/<name>.yaml.

Structure (DESIGN D3 layers)

Not yet: a hosted wrapper and R2 storage (Phase 3c). Unclear points in the spec: docs/spec/ISSUES.md.

Historical logs are in the archive of live runs.