Play tic-tac-toe and never lose

Befunge, 181 168 bytes

>>4&5pp20555>>03>16p\::5g8%6p5v
 ^p5.:g605$_ #!<^_|#:-1g61+%8g<
543217539511|:_^#->#g0<>8+00p3+5%09638527419876
v<304p$_v#:->#$$:2`#3_:^
>#\3#13#<111v124236478689189378

Positions on the board are numbered 1 to 9. By default you get the first move, but if you want to allow the computer to go first, you can just enter 0 for your first move. When you've made a move, the computer will respond with a number indicating their move.

There are no checks to make sure you don't enter a valid move, and there also no checks to see if anyone has won or lost. Once their are no more moves to be made, the program just goes into an infinite loop.

It's a bit difficult to test this online, since there are no online interpreters with interactive input. However, if you know what moves you're going to make in advance (which assumes you know how the computer is going to respond), you can kind of test on TIO with those moves preprogrammed.

User plays first: Try it online!
Computer plays first: Try it online!

To makes it easier to see what's going on, I've also got a version that outputs the board between moves.

User plays first: Try it online!
Computer plays first: Try it online!

Note that you'll have to wait for TIO to timeout before you can see the results.

Explanation

The board is stored in the Befunge memory area as a flat array of 9 values, indexed from 1 to 9. This lets us use the zero offset as a special case "no move" when we want to let the computer play first. Player moves are stored as 4, and computer moves as 5. To start with all positions are initialised to 32 (the Befunge memory default), so whenever we access the board we mod with 8, so we'll get back either 0, 4 or 5.

Given that arrangement, if we sum the values of any three positions on the board, we know that the computer is one move away from winning if the total is 10, the player is one move away from winning if the total is 8, and the positions are shared between computer and player (but still one position is free) if the total is 9.

Our entire strategy is based around this concept. We have a routine that takes a list of triples indicating sets of three positions on the board, we calculate the sum of those positions, and if the sum equals a certain total, the computer moves to whichever of the positions in the set is free.

The main list of triples that we test are the winning combinations (1/2/3, 1/5/9, 1/4/7, etc.). We first look for a total of 10 (the computer is about to win), and then a total of 8 (the player is about to win and we need to block that move). Less obviously, we also check for a total of 9 (if the player and computer each have one of the positions, it's a good strategy for the computer to take the third).

Prior to that last scenario, the other strategic move we make is to check all the corner sets (1/2/4, 2/3/6, etc.) as well as two opposing corner combinations (1/8/9 and 3/7/8). If any of these combinations sum to 8, i.e. the player has taken two of the positions, it's a good strategy for the computer to take the remaining free position.

Finally, there are a two special case moves. First, we always try and take the center position before any other move. This is achieved with the same routine as all our other moves, just passing in a single triple, 5/5/5, and a target sum of 0. Additionally, if all other test have failed to find a move, we try to take one of the top corners as a last resort. Again this is simply achieved by testing the triples 1/1/1 and 3/3/3, with a target sum of 0.

I don't think this is necessarily a perfect strategy - there may be games that the computer draws which could potentially have been won - but it's good enough to never lose a match. I've run a test script that tried to play every possible move against the computer, and for every valid sequence of moves, the computer either won or drew the game.

Python 2: 399 401 349 333 317 370 bytes

2x Bug Fix: credit to l4m2

-52 chars: credit to undergroundmonorail

-16 chars: credit to Jonathan Frech

-26 chars: credit to user202729

def f(b):
 t=4,9,2,3,5,7,8,1,6;n=lambda k:[t[i]for i,j in enumerate(b)if j==k];p,o,a,I=n(2),n(1),n(0),t.index
 for i in p:
    for j in p:
     for k in a:
        if i+j+k==15and-j+i:return I(k)
 for i in o:
    for j in o:
     for k in a:
        if i+j+k==15and-j+i:return I(k)
 for i in 9,3,7,1:
    if i in a and 5 in p:return I(i)
 for i in 5,4,2,8,6:
    if i in a:return I(i)
 return I(a[0])

Try it Online!

On the first day of a linear algebra course I took last semester, my astute graduate student instructor proposed that if you represent the tic-tac-toe board as the the matrix:

4 | 9 | 2
--+---+--
3 | 5 | 7
--+---+--
8 | 1 | 6

then getting three in a row is equivalent to picking three numbers in the range [1,9] that add up to 15. This answer exploits this idea. The function takes a list containing nine numbers representing the board. 0 indicates an empty space, 1 is occupied by the opponent, and 2 represents a previous play made by the program. The first 3 lines figure out what numbers the program has picked (p), the opposition has picked (o), and are still available (a). It then looks through the available numbers and sees if any of them, combined with two numbers it has already picked add to fifteen. If it does, it will pick that square and win. If there are no immediate winning moves, it will check to see if the opponent can win using the same method. If they can, it will take their winning square. If there is neither a winning nor blocking move available, it will move in a corner. This prevents a fools mate:

- - - 
- X -
- - -

- O -             # Bad Move
- X -
- - -

- O X
- X -
- - -

- O X
- X -
O - -

- O X
- X -
O - X

If none of these situations occur, it will choose a square arbitrarily. The function outputs a number [0,8] representing the 0 indexed square chosen by the algorithm.

Edit: The algorithm now prioritizes the center over the diagonal, which will prevent another fools mate possibility pointed out by l4m2 and related strategies.

Edit: To clarify, the function takes in a board in the form of an array and outputs a move as an integer on [0,8]. Because this I/O strategy is so clunky , here's a wrapper script that makes it more interactive. It takes a single command line argument, which should be 1 if the player goes first, and 0 if the program goes first.

import sys

def f(b):
 t=4,9,2,3,5,7,8,1,6;n=lambda k:[t[i]for i,j in enumerate(b)if j==k];p,o,a,I=n(2),n(1),n(0),t.index
 for i in p:
    for j in p:
     for k in a:
        if i+j+k==15and-j+i:return I(k)
 for i in o:
    for j in o:
     for k in a:
        if i+j+k==15and-j+i:return I(k)
 for i in 9,3,7,1:
    if i in a and 5 in p:return I(i)
     for i in 5,4,2,8,6:
        if i in a:return I(i)
 return I(a[0])

board = [0,0,0,0,0,0,0,0,0]
rep = {0:"-",1:"X",2:"O"}

turn = int(sys.argv[1])
while True:
    for i in range(3):
        print rep[board[i*3]]+" "+rep[board[i*3+1]]+" "+rep[board[i*3+2]]
        print
    if turn:
        move = int(raw_input("Enter Move [0-8]: "))
    else:
        move = f(board)
    board[move] = turn+1
    turn = (turn+1)%2