| name | tares |
| description | Installs Tares, the open-source platform for always-on AI agents, starts the daemon, adds a first source over the HTTP API, connects the agent it runs inside over MCP, and proves the connection with one read. Use when the user asks an agent to install or set up Tares, add a source, connect Claude Code, Codex or Cursor to Tares, or fix a local Tares install. |
Tares
Install and connect Tares (docs: https://docs.glassflow.ai/tares). Tares puts every event
from a user's systems on one timeline per thing, wakes an agent when a trigger fires, and keeps
the finding for the next reader. This skill takes a machine from nothing to: daemon running, one
source receiving, this agent connected over MCP, one successful read.
Operating rules
- Act autonomously once the user asks for Tares to be installed. Ask only for what you cannot
infer: which source to add first, and its one or two parameters (see step 4).
- Say what is about to run and that it may take a while.
uv tool install can take a minute;
post a line of progress while it runs so the user does not kill it.
- Print each command and its result. If a command fails, change something before retrying.
- Never paste secrets into the console, the chat, or a file you show. Read them from environment
variables (
GITHUB_TOKEN, ANTHROPIC_API_KEY). A variable exported only in your shell is not
seen by a tares up started elsewhere; start the daemon from the shell that has it, or tell
the user to set it where the daemon runs.
- Do not create what already exists:
GET /api/sources first, and skip a source whose name is
taken.
- Everything below goes through the HTTP API at
http://127.0.0.1:8787, not the console. The
console is for the user afterwards.
1. Prerequisites
uv --version || pipx --version
lsof -iTCP:8787 -sTCP:LISTEN -n -P
lsof -iTCP:8788 -sTCP:LISTEN -n -P
docker info >/dev/null 2>&1 && echo docker-ok
2. Install
tares --version 2>/dev/null && uv tool upgrade tares || uv tool install tares
tares --version
If tares is not on PATH after installing, uv tool update-shell (or add ~/.local/bin to PATH)
and open a new shell.
3. Start the daemon
Pick a data directory (default ~/.tares; use another one if the user wants an isolated
install). Run detached so the session survives this conversation:
nohup tares up --data-dir ~/.tares > ~/.tares-up.log 2>&1 &
for i in $(seq 1 30); do curl -sf http://127.0.0.1:8787/health && break; sleep 1; done
/health returns {"status":"ok","auth_required":false,"sources":[...]}. Print the console URL:
http://127.0.0.1:8787. To stop later: pkill -f "tares up" (or kill the PID from
lsof -iTCP:8787). To restart: the same nohup line.
If port 8787 was already taken by a running Tares, skip this step and use it.
4. Add the first source
Ask the user which one, then POST /api/sources with the exact body below. GET /api/sources
first; if the name exists, move on to step 5. Every body has the same shape:
{"name", "connector", "poll", "config"}. Full config keys per connector:
https://docs.glassflow.ai/tares/connectors. Do not invent keys.
Prometheus (ask for the URL, one PromQL, and the label that names the service):
curl -sf -X POST http://127.0.0.1:8787/api/sources -H 'Content-Type: application/json' -d '{
"name": "metrics", "connector": "prometheus", "poll": "10s",
"config": {
"url": "http://localhost:9090",
"queries": [{"promql": "rate(http_requests_total[1m])", "by_name": true}],
"labels": [
{"name": "service", "field": "metric.service", "primary": true},
{"name": "value", "field": "value", "type": "number"}
]
}}'
OTLP (nothing to ask; print the endpoint the collector should point at):
curl -sf -X POST http://127.0.0.1:8787/api/sources -H 'Content-Type: application/json' -d '{
"name": "otel", "connector": "otlp",
"config": {"labels": [{"name": "service", "field": "service", "primary": true}]}}'
Then tell the user: OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:8787 (OTLP/HTTP, paths
/v1/logs, /v1/traces, /v1/metrics).
Docker logs (ask for the container name; docker ps lists them):
curl -sf -X POST http://127.0.0.1:8787/api/sources -H 'Content-Type: application/json' -d '{
"name": "logs", "connector": "docker_logs", "poll": "5s",
"config": {
"container": "<container name>",
"labels": [{"name": "service", "const": "<service name>", "primary": true}]
}}'
GitHub (ask for owner/repo; the token comes from GITHUB_TOKEN, needed for private repos):
curl -sf -X POST http://127.0.0.1:8787/api/sources -H 'Content-Type: application/json' -d "{
\"name\": \"commits\", \"connector\": \"github\", \"poll\": \"30s\",
\"config\": {
\"repo\": \"<owner/repo>\",
\"token\": \"${GITHUB_TOKEN:-}\",
\"labels\": [{\"name\": \"repo\", \"field\": \"repo\", \"primary\": true},
{\"name\": \"author\", \"field\": \"author\"}]
}}"
Drop the token line for a public repo without a token. Never echo the body with the token in it.
Webhook (nothing to ask; create it and print the ingest URL):
curl -sf -X POST http://127.0.0.1:8787/api/sources -H 'Content-Type: application/json' -d '{
"name": "events", "connector": "webhook",
"config": {
"event_type_field": "action",
"text_template": "{action} by {sender}",
"labels": [{"name": "pipeline", "field": "pipeline", "primary": true}]
}}'
The response carries ingest_key. Print POST http://127.0.0.1:8787/ingest/<ingest_key> with a
JSON body, and send one test event yourself:
curl -sf -X POST http://127.0.0.1:8787/ingest/<ingest_key> -H 'Content-Type: application/json' \
-d '{"action": "setup", "sender": "tares-skill", "pipeline": "setup"}'
5. Verify the source is receiving
for i in $(seq 1 30); do
n=$(curl -sf http://127.0.0.1:8787/api/sources/<name> | python3 -c 'import sys,json; h=json.load(sys.stdin).get("health") or {}; print(h.get("events_total") or 0)')
[ "$n" -gt 0 ] && break; sleep 2
done
curl -sf http://127.0.0.1:8787/api/sources/<name> | python3 -c 'import sys,json; h=json.load(sys.stdin).get("health") or {}; print("status", h.get("status"), "events", h.get("events_total"), "last_error", h.get("last_error"))'
Report what it saw: the event count, or the last_error if the count stayed 0 after 60 seconds
(then fix the config and retry; see troubleshooting.md).
6. Connect this agent over MCP
Start the MCP endpoint, detached:
nohup tares mcp --transport streamable-http --port 8788 --taresd http://127.0.0.1:8787 > ~/.tares-mcp.log 2>&1 &
sleep 2; lsof -iTCP:8788 -sTCP:LISTEN -n -P
Then register it with the client you are running inside:
- Claude Code:
claude mcp add --transport http tares http://localhost:8788/mcp
- Codex CLI:
codex mcp add tares --url http://localhost:8788/mcp (keep --url; a bare URL
registers a stdio server)
- Cursor: write
~/.cursor/mcp.json (or .cursor/mcp.json in the project):
{"mcpServers": {"tares": {"url": "http://localhost:8788/mcp"}}} and tell the user to approve
it under Settings > MCP
- Claude Desktop: the same JSON in
claude_desktop_config.json, then restart the app
- Anything else: transport streamable HTTP, URL
http://localhost:8788/mcp, no auth on a
local instance. Details: https://docs.glassflow.ai/tares/agents
7. Prove it
Call the read tool once for the entity the new source produced, for example
read(selector={"service": "<service name>"}, window="15m") (for GitHub the key is repo, for
the webhook pipeline). Show the user the result. If the client needs a restart before new MCP
servers are visible, say so and show the equivalent HTTP call instead:
curl -sf -X POST http://127.0.0.1:8787/read -H 'Content-Type: application/json' \
-d '{"selector": {"service": "<service name>"}, "window": "15m"}' | head -c 800
8. Report
One short block: what was installed (version), the data dir, the console URL, how to stop and
start, which source was added and how many events it has, which client is connected. End with
exactly one next step:
Troubleshooting
See troubleshooting.md.
